diff --git a/gen_img_diffusers_control_net.py b/gen_img_diffusers_control_net.py new file mode 100644 index 00000000..53eb50e7 --- /dev/null +++ b/gen_img_diffusers_control_net.py @@ -0,0 +1,2633 @@ +""" +VGG( + (features): Sequential( + (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (1): ReLU(inplace=True) + (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (3): ReLU(inplace=True) + (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) + (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (6): ReLU(inplace=True) + (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (8): ReLU(inplace=True) + (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) + (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (11): ReLU(inplace=True) + (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (13): ReLU(inplace=True) + (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (15): ReLU(inplace=True) + (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) + (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (18): ReLU(inplace=True) + (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (20): ReLU(inplace=True) + (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (22): ReLU(inplace=True) + (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) + (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (25): ReLU(inplace=True) + (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (27): ReLU(inplace=True) + (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + (29): ReLU(inplace=True) + (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) + ) + (avgpool): AdaptiveAvgPool2d(output_size=(7, 7)) + (classifier): Sequential( + (0): Linear(in_features=25088, out_features=4096, bias=True) + (1): ReLU(inplace=True) + (2): Dropout(p=0.5, inplace=False) + (3): Linear(in_features=4096, out_features=4096, bias=True) + (4): ReLU(inplace=True) + (5): Dropout(p=0.5, inplace=False) + (6): Linear(in_features=4096, out_features=1000, bias=True) + ) +) +""" + +import json +from typing import List, Optional, Union +import glob +import importlib +import inspect +import time +import zipfile +from diffusers.utils import deprecate +from diffusers.configuration_utils import FrozenDict +import argparse +import math +import os +import random +import re +from typing import Any, Callable, List, Optional, Union + +import diffusers +import numpy as np +import torch +import torchvision +from diffusers import (AutoencoderKL, DDPMScheduler, + EulerAncestralDiscreteScheduler, DPMSolverMultistepScheduler, DPMSolverSinglestepScheduler, + LMSDiscreteScheduler, PNDMScheduler, DDIMScheduler, EulerDiscreteScheduler, HeunDiscreteScheduler, + KDPM2DiscreteScheduler, KDPM2AncestralDiscreteScheduler, + UNet2DConditionModel, StableDiffusionPipeline) +from einops import rearrange +from torch import einsum +from tqdm import tqdm +from torchvision import transforms +from transformers import CLIPTextModel, CLIPTokenizer, CLIPModel, CLIPTextConfig +import PIL +from PIL import Image +from PIL.PngImagePlugin import PngInfo + +import library.model_util as model_util + +# Tokenizer: checkpointから読み込むのではなくあらかじめ提供されているものを使う +TOKENIZER_PATH = "openai/clip-vit-large-patch14" +V2_STABLE_DIFFUSION_PATH = "stabilityai/stable-diffusion-2" # ここからtokenizerだけ使う + +DEFAULT_TOKEN_LENGTH = 75 + +# scheduler: +SCHEDULER_LINEAR_START = 0.00085 +SCHEDULER_LINEAR_END = 0.0120 +SCHEDULER_TIMESTEPS = 1000 +SCHEDLER_SCHEDULE = 'scaled_linear' + +# その他の設定 +LATENT_CHANNELS = 4 +DOWNSAMPLING_FACTOR = 8 + +# CLIP_ID_L14_336 = "openai/clip-vit-large-patch14-336" + +# CLIP guided SD関連 +CLIP_MODEL_PATH = "laion/CLIP-ViT-B-32-laion2B-s34B-b79K" +FEATURE_EXTRACTOR_SIZE = (224, 224) +FEATURE_EXTRACTOR_IMAGE_MEAN = [0.48145466, 0.4578275, 0.40821073] +FEATURE_EXTRACTOR_IMAGE_STD = [0.26862954, 0.26130258, 0.27577711] + +VGG16_IMAGE_MEAN = [0.485, 0.456, 0.406] +VGG16_IMAGE_STD = [0.229, 0.224, 0.225] +VGG16_INPUT_RESIZE_DIV = 4 + +# CLIP特徴量の取得時にcutoutを使うか：使う場合にはソースを書き換えてください +NUM_CUTOUTS = 4 +USE_CUTOUTS = False + +# region モジュール入れ替え部 +""" +高速化のためのモジュール入れ替え +""" + +# FlashAttentionを使うCrossAttention +# based on https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/memory_efficient_attention_pytorch/flash_attention.py +# LICENSE MIT https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/LICENSE + +# constants + +EPSILON = 1e-6 + +# helper functions + + +def exists(val): + return val is not None + + +def default(val, d): + return val if exists(val) else d + +# flash attention forwards and backwards + +# https://arxiv.org/abs/2205.14135 + + +class FlashAttentionFunction(torch.autograd.Function): + @ staticmethod + @ torch.no_grad() + def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size): + """ Algorithm 2 in the paper """ + + device = q.device + dtype = q.dtype + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + o = torch.zeros_like(q) + all_row_sums = torch.zeros((*q.shape[:-1], 1), dtype=dtype, device=device) + all_row_maxes = torch.full((*q.shape[:-1], 1), max_neg_value, dtype=dtype, device=device) + + scale = (q.shape[-1] ** -0.5) + + if not exists(mask): + mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size) + else: + mask = rearrange(mask, 'b n -> b 1 1 n') + mask = mask.split(q_bucket_size, dim=-1) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + mask, + all_row_sums.split(q_bucket_size, dim=-2), + all_row_maxes.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, row_mask, row_sums, row_maxes) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = einsum('... i d, ... j d -> ... i j', qc, kc) * scale + + if exists(row_mask): + attn_weights.masked_fill_(~row_mask, max_neg_value) + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, + device=device).triu(q_start_index - k_start_index + 1) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + block_row_maxes = attn_weights.amax(dim=-1, keepdims=True) + attn_weights -= block_row_maxes + exp_weights = torch.exp(attn_weights) + + if exists(row_mask): + exp_weights.masked_fill_(~row_mask, 0.) + + block_row_sums = exp_weights.sum(dim=-1, keepdims=True).clamp(min=EPSILON) + + new_row_maxes = torch.maximum(block_row_maxes, row_maxes) + + exp_values = einsum('... i j, ... j d -> ... i d', exp_weights, vc) + + exp_row_max_diff = torch.exp(row_maxes - new_row_maxes) + exp_block_row_max_diff = torch.exp(block_row_maxes - new_row_maxes) + + new_row_sums = exp_row_max_diff * row_sums + exp_block_row_max_diff * block_row_sums + + oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_((exp_block_row_max_diff / new_row_sums) * exp_values) + + row_maxes.copy_(new_row_maxes) + row_sums.copy_(new_row_sums) + + ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size) + ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes) + + return o + + @ staticmethod + @ torch.no_grad() + def backward(ctx, do): + """ Algorithm 4 in the paper """ + + causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args + q, k, v, o, l, m = ctx.saved_tensors + + device = q.device + + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + dq = torch.zeros_like(q) + dk = torch.zeros_like(k) + dv = torch.zeros_like(v) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + do.split(q_bucket_size, dim=-2), + mask, + l.split(q_bucket_size, dim=-2), + m.split(q_bucket_size, dim=-2), + dq.split(q_bucket_size, dim=-2) + ) + + for ind, (qc, oc, doc, row_mask, lc, mc, dqc) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + dk.split(k_bucket_size, dim=-2), + dv.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = einsum('... i d, ... j d -> ... i j', qc, kc) * scale + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, + device=device).triu(q_start_index - k_start_index + 1) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + exp_attn_weights = torch.exp(attn_weights - mc) + + if exists(row_mask): + exp_attn_weights.masked_fill_(~row_mask, 0.) + + p = exp_attn_weights / lc + + dv_chunk = einsum('... i j, ... i d -> ... j d', p, doc) + dp = einsum('... i d, ... j d -> ... i j', doc, vc) + + D = (doc * oc).sum(dim=-1, keepdims=True) + ds = p * scale * (dp - D) + + dq_chunk = einsum('... i j, ... j d -> ... i d', ds, kc) + dk_chunk = einsum('... i j, ... i d -> ... j d', ds, qc) + + dqc.add_(dq_chunk) + dkc.add_(dk_chunk) + dvc.add_(dv_chunk) + + return dq, dk, dv, None, None, None, None + + +def replace_unet_modules(unet: diffusers.models.unet_2d_condition.UNet2DConditionModel, mem_eff_attn, xformers): + if mem_eff_attn: + replace_unet_cross_attn_to_memory_efficient() + elif xformers: + replace_unet_cross_attn_to_xformers() + + +def replace_unet_cross_attn_to_memory_efficient(): + print("Replace CrossAttention.forward to use NAI style Hypernetwork and FlashAttention") + flash_func = FlashAttentionFunction + + def forward_flash_attn(self, x, context=None, mask=None): + q_bucket_size = 512 + k_bucket_size = 1024 + + h = self.heads + q = self.to_q(x) + + context = context if context is not None else x + context = context.to(x.dtype) + + if hasattr(self, 'hypernetwork') and self.hypernetwork is not None: + context_k, context_v = self.hypernetwork.forward(x, context) + context_k = context_k.to(x.dtype) + context_v = context_v.to(x.dtype) + else: + context_k = context + context_v = context + + k = self.to_k(context_k) + v = self.to_v(context_v) + del context, x + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v)) + + out = flash_func.apply(q, k, v, mask, False, q_bucket_size, k_bucket_size) + + out = rearrange(out, 'b h n d -> b n (h d)') + + # diffusers 0.7.0~ + out = self.to_out[0](out) + out = self.to_out[1](out) + return out + + diffusers.models.attention.CrossAttention.forward = forward_flash_attn + + +def replace_unet_cross_attn_to_xformers(): + print("Replace CrossAttention.forward to use NAI style Hypernetwork and xformers") + try: + import xformers.ops + except ImportError: + raise ImportError("No xformers / xformersがインストールされていないようです") + + def forward_xformers(self, x, context=None, mask=None): + h = self.heads + q_in = self.to_q(x) + + context = default(context, x) + context = context.to(x.dtype) + + if hasattr(self, 'hypernetwork') and self.hypernetwork is not None: + context_k, context_v = self.hypernetwork.forward(x, context) + context_k = context_k.to(x.dtype) + context_v = context_v.to(x.dtype) + else: + context_k = context + context_v = context + + k_in = self.to_k(context_k) + v_in = self.to_v(context_v) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b n h d', h=h), (q_in, k_in, v_in)) + del q_in, k_in, v_in + + q = q.contiguous() + k = k.contiguous() + v = v.contiguous() + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None) # 最適なのを選んでくれる + + out = rearrange(out, 'b n h d -> b n (h d)', h=h) + + # diffusers 0.7.0~ + out = self.to_out[0](out) + out = self.to_out[1](out) + return out + + diffusers.models.attention.CrossAttention.forward = forward_xformers +# endregion + +# region 画像生成の本体：lpw_stable_diffusion.py （ASL）からコピーして修正 +# https://github.com/huggingface/diffusers/blob/main/examples/community/lpw_stable_diffusion.py +# Pipelineだけ独立して使えないのと機能追加するのとでコピーして修正 + + +class PipelineLike(): + r""" + Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing + weighting in prompt. + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + def __init__( + self, + device, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + clip_skip: int, + clip_model: CLIPModel, + clip_guidance_scale: float, + clip_image_guidance_scale: float, + vgg16_model: torchvision.models.VGG, + vgg16_guidance_scale: float, + vgg16_layer_no: int, + # safety_checker: StableDiffusionSafetyChecker, + # feature_extractor: CLIPFeatureExtractor, + ): + super().__init__() + self.device = device + self.clip_skip = clip_skip + + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + self.vae = vae + self.text_encoder = text_encoder + self.tokenizer = tokenizer + self.unet = unet + self.scheduler = scheduler + self.safety_checker = None + + # Textual Inversion + self.token_replacements = {} + + # CLIP guidance + self.clip_guidance_scale = clip_guidance_scale + self.clip_image_guidance_scale = clip_image_guidance_scale + self.clip_model = clip_model + self.normalize = transforms.Normalize(mean=FEATURE_EXTRACTOR_IMAGE_MEAN, std=FEATURE_EXTRACTOR_IMAGE_STD) + self.make_cutouts = MakeCutouts(FEATURE_EXTRACTOR_SIZE) + + # VGG16 guidance + self.vgg16_guidance_scale = vgg16_guidance_scale + if self.vgg16_guidance_scale > 0.0: + return_layers = {f'{vgg16_layer_no}': 'feat'} + self.vgg16_feat_model = torchvision.models._utils.IntermediateLayerGetter(vgg16_model.features, return_layers=return_layers) + self.vgg16_normalize = transforms.Normalize(mean=VGG16_IMAGE_MEAN, std=VGG16_IMAGE_STD) + + def set_lora_network(self, lora_network): + self.lora_network = lora_network + + # Textual Inversion + def add_token_replacement(self, target_token_id, rep_token_ids): + self.token_replacements[target_token_id] = rep_token_ids + + def replace_token(self, tokens): + new_tokens = [] + for token in tokens: + if token in self.token_replacements: + new_tokens.extend(self.token_replacements[token]) + else: + new_tokens.append(token) + return new_tokens + + # region xformersとか使う部分：独自に書き換えるので関係なし + def enable_xformers_memory_efficient_attention(self): + r""" + Enable memory efficient attention as implemented in xformers. + When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference + time. Speed up at training time is not guaranteed. + Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention + is used. + """ + self.unet.set_use_memory_efficient_attention_xformers(True) + + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention as implemented in xformers. + """ + self.unet.set_use_memory_efficient_attention_xformers(False) + + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. + When this option is enabled, the attention module will split the input tensor in slices, to compute attention + in several steps. This is useful to save some memory in exchange for a small speed decrease. + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, + `attention_head_dim` must be a multiple of `slice_size`. + """ + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = self.unet.config.attention_head_dim // 2 + self.unet.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go + back to computing attention in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def enable_sequential_cpu_offload(self): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + """ + # accelerateが必要になるのでとりあえず省略 + raise NotImplementedError("cpu_offload is omitted.") + # if is_accelerate_available(): + # from accelerate import cpu_offload + # else: + # raise ImportError("Please install accelerate via `pip install accelerate`") + + # device = self.device + + # for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]: + # if cpu_offloaded_model is not None: + # cpu_offload(cpu_offloaded_model, device) +# endregion + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + init_image: Union[torch.FloatTensor, PIL.Image.Image, List[PIL.Image.Image]] = None, + mask_image: Union[torch.FloatTensor, PIL.Image.Image, List[PIL.Image.Image]] = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_scale: float = None, + strength: float = 0.8, + # num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + # return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: Optional[int] = 1, + img2img_noise=None, + clip_prompts=None, + clip_guide_images=None, + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + init_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `init_image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1. + `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + num_images_per_prompt = 1 # fixed + + if isinstance(prompt, str): + batch_size = 1 + prompt = [prompt] + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # get prompt text embeddings + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + if not do_classifier_free_guidance and negative_scale is not None: + print(f"negative_scale is ignored if guidance scalle <= 1.0") + negative_scale = None + + # get unconditional embeddings for classifier free guidance + if negative_prompt is None: + negative_prompt = [""] * batch_size + elif isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + if batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + text_embeddings, uncond_embeddings, prompt_tokens = get_weighted_text_embeddings( + pipe=self, + prompt=prompt, + uncond_prompt=negative_prompt if do_classifier_free_guidance else None, + max_embeddings_multiples=max_embeddings_multiples, + clip_skip=self.clip_skip, + **kwargs, + ) + + if negative_scale is not None: + _, real_uncond_embeddings, _ = get_weighted_text_embeddings( + pipe=self, + prompt=prompt, # こちらのトークン長に合わせてuncondを作るので75トークン超で必須 + uncond_prompt=[""]*batch_size, + max_embeddings_multiples=max_embeddings_multiples, + clip_skip=self.clip_skip, + **kwargs, + ) + + if do_classifier_free_guidance: + if negative_scale is None: + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + else: + text_embeddings = torch.cat([uncond_embeddings, text_embeddings, real_uncond_embeddings]) + + # CLIP guidanceで使用するembeddingsを取得する + if self.clip_guidance_scale > 0: + clip_text_input = prompt_tokens + if clip_text_input.shape[1] > self.tokenizer.model_max_length: + # TODO 75文字を超えたら警告を出す？ + print("trim text input", clip_text_input.shape) + clip_text_input = torch.cat([clip_text_input[:, :self.tokenizer.model_max_length-1], + clip_text_input[:, -1].unsqueeze(1)], dim=1) + print("trimmed", clip_text_input.shape) + + for i, clip_prompt in enumerate(clip_prompts): + if clip_prompt is not None: # clip_promptがあれば上書きする + clip_text_input[i] = self.tokenizer(clip_prompt, padding="max_length", max_length=self.tokenizer.model_max_length, + truncation=True, return_tensors="pt",).input_ids.to(self.device) + + text_embeddings_clip = self.clip_model.get_text_features(clip_text_input) + text_embeddings_clip = text_embeddings_clip / text_embeddings_clip.norm(p=2, dim=-1, keepdim=True) # prompt複数件でもOK + + if self.clip_image_guidance_scale > 0 or self.vgg16_guidance_scale > 0 and clip_guide_images is not None: + if isinstance(clip_guide_images, PIL.Image.Image): + clip_guide_images = [clip_guide_images] + + if self.clip_image_guidance_scale > 0: + clip_guide_images = [preprocess_guide_image(im) for im in clip_guide_images] + clip_guide_images = torch.cat(clip_guide_images, dim=0) + + clip_guide_images = self.normalize(clip_guide_images).to(self.device).to(text_embeddings.dtype) + image_embeddings_clip = self.clip_model.get_image_features(clip_guide_images) + image_embeddings_clip = image_embeddings_clip / image_embeddings_clip.norm(p=2, dim=-1, keepdim=True) + if len(image_embeddings_clip) == 1: + image_embeddings_clip = image_embeddings_clip.repeat((batch_size, 1, 1, 1)) + else: + size = (width // VGG16_INPUT_RESIZE_DIV, height // VGG16_INPUT_RESIZE_DIV) # とりあえず1/4に（小さいか?） + clip_guide_images = [preprocess_vgg16_guide_image(im, size) for im in clip_guide_images] + clip_guide_images = torch.cat(clip_guide_images, dim=0) + + clip_guide_images = self.vgg16_normalize(clip_guide_images).to(self.device).to(text_embeddings.dtype) + image_embeddings_vgg16 = self.vgg16_feat_model(clip_guide_images)['feat'] + if len(image_embeddings_vgg16) == 1: + image_embeddings_vgg16 = image_embeddings_vgg16.repeat((batch_size, 1, 1, 1)) + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps, self.device) + + latents_dtype = text_embeddings.dtype + init_latents_orig = None + mask = None + noise = None + + # if init_image is None: + # get the initial random noise unless the user supplied it + + # Unlike in other pipelines, latents need to be generated in the target device + # for 1-to-1 results reproducibility with the CompVis implementation. + # However this currently doesn't work in `mps`. + latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8,) + + if latents is None: + if self.device.type == "mps": + # randn does not exist on mps + latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype,).to(self.device) + else: + latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype,) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + latents = latents.to(self.device) + + timesteps = self.scheduler.timesteps.to(self.device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + + if init_image is not None: + # init imageをhintとして使う + # image to tensor + if isinstance(init_image, PIL.Image.Image): + init_image = [init_image] + if isinstance(init_image[0], PIL.Image.Image): + init_image = [preprocess_image(im) for im in init_image] + init_image = torch.cat(init_image) + + # mask image to tensor + if mask_image is not None: + if isinstance(mask_image, PIL.Image.Image): + mask_image = [mask_image] + if isinstance(mask_image[0], PIL.Image.Image): + mask_image = torch.cat([preprocess_mask(im) for im in mask_image]) # H*W, 0 for repaint + + # encode the init image into latents and scale the latents + init_image = init_image.to(device=self.device, dtype=latents_dtype) + init_latent_dist = self.vae.encode(init_image).latent_dist + init_latents = init_latent_dist.sample(generator=generator) + init_latents = 0.18215 * init_latents + if len(init_latents) == 1: + init_latents = init_latents.repeat((batch_size, 1, 1, 1)) + init_latents_orig = init_latents + + # # preprocess mask + # if mask_image is not None: + # mask = mask_image.to(device=self.device, dtype=latents_dtype) + # if len(mask) == 1: + # mask = mask.repeat((batch_size, 1, 1, 1)) + + # # check sizes + # if not mask.shape == init_latents.shape: + # raise ValueError("The mask and init_image should be the same size!") + + # init imageをhintとして使う + hint_latents = init_latents + # org_dtype = init_image.dtype + # hint = torch.nn.functional.interpolate(init_image.to(torch.float32), scale_factor=(1/8, 1/8), mode="bilinear") + # hint = hint[:, 0].unsqueeze(1) # RGB -> BW + # hint = hint.to(org_dtype) + """ + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + timesteps = self.scheduler.timesteps[-init_timestep] + timesteps = torch.tensor([timesteps] * batch_size * num_images_per_prompt, device=self.device) + + # add noise to latents using the timesteps + latents = self.scheduler.add_noise(init_latents, img2img_noise, timesteps) + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].to(self.device) + """ + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + hint_latents = torch.cat([hint_latents, hint_latents]) + + num_latent_input = (3 if negative_scale is not None else 2) if do_classifier_free_guidance else 1 + for i, t in enumerate(tqdm(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = latents.repeat((num_latent_input, 1, 1, 1)) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + self.lora_network.set_as_control_path(True) + # self.unet(latent_model_input * hint, t, encoder_hidden_states=text_embeddings).sample + self.unet(hint_latents, t, encoder_hidden_states=text_embeddings) + self.lora_network.set_as_control_path(False) + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + if negative_scale is None: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(num_latent_input) # uncond by negative prompt + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + else: + noise_pred_negative, noise_pred_text, noise_pred_uncond = noise_pred.chunk(num_latent_input) # uncond is real uncond + noise_pred = noise_pred_uncond + guidance_scale * \ + (noise_pred_text - noise_pred_uncond) - negative_scale * (noise_pred_negative - noise_pred_uncond) + + # perform clip guidance + if self.clip_guidance_scale > 0 or self.clip_image_guidance_scale > 0 or self.vgg16_guidance_scale > 0: + text_embeddings_for_guidance = (text_embeddings.chunk(num_latent_input)[ + 1] if do_classifier_free_guidance else text_embeddings) + + if self.clip_guidance_scale > 0: + noise_pred, latents = self.cond_fn(latents, t, i, text_embeddings_for_guidance, noise_pred, + text_embeddings_clip, self.clip_guidance_scale, NUM_CUTOUTS, USE_CUTOUTS,) + if self.clip_image_guidance_scale > 0 and clip_guide_images is not None: + noise_pred, latents = self.cond_fn(latents, t, i, text_embeddings_for_guidance, noise_pred, + image_embeddings_clip, self.clip_image_guidance_scale, NUM_CUTOUTS, USE_CUTOUTS,) + if self.vgg16_guidance_scale > 0 and clip_guide_images is not None: + noise_pred, latents = self.cond_fn_vgg16(latents, t, i, text_embeddings_for_guidance, noise_pred, + image_embeddings_vgg16, self.vgg16_guidance_scale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + if mask is not None: + # masking + init_latents_proper = self.scheduler.add_noise(init_latents_orig, img2img_noise, torch.tensor([t])) + latents = (init_latents_proper * mask) + (latents * (1 - mask)) + + # call the callback, if provided + if i % callback_steps == 0: + if callback is not None: + callback(i, t, latents) + if is_cancelled_callback is not None and is_cancelled_callback(): + return None + + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + + image = (image / 2 + 0.5).clamp(0, 1) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to( + self.device + ) + image, has_nsfw_concept = self.safety_checker( + images=image, + clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype), + ) + else: + has_nsfw_concept = None + + if output_type == "pil": + # image = self.numpy_to_pil(image) + image = (image * 255).round().astype("uint8") + image = [Image.fromarray(im) for im in image] + + # if not return_dict: + return (image, has_nsfw_concept) + + # return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + def text2img( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function for text-to-image generation. + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + **kwargs, + ) + + def img2img( + self, + init_image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[torch.Generator] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function for image-to-image generation. + Args: + init_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1. + `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + init_image=init_image, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + strength=strength, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + **kwargs, + ) + + def inpaint( + self, + init_image: Union[torch.FloatTensor, PIL.Image.Image], + mask_image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[torch.Generator] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: Optional[int] = 1, + **kwargs, + ): + r""" + Function for inpaint. + Args: + init_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `init_image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength` + is 1, the denoising process will be run on the masked area for the full number of iterations specified + in `num_inference_steps`. `init_image` will be used as a reference for the masked area, adding more + noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur. + num_inference_steps (`int`, *optional*, defaults to 50): + The reference number of denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. This parameter will be modulated by `strength`, as explained above. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + init_image=init_image, + mask_image=mask_image, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + strength=strength, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + callback_steps=callback_steps, + **kwargs, + ) + + # CLIP guidance StableDiffusion + # copy from https://github.com/huggingface/diffusers/blob/main/examples/community/clip_guided_stable_diffusion.py + + # バッチを分解して1件ずつ処理する + def cond_fn(self, latents, timestep, index, text_embeddings, noise_pred_original, guide_embeddings_clip, clip_guidance_scale, + num_cutouts, use_cutouts=True, ): + if len(latents) == 1: + return self.cond_fn1(latents, timestep, index, text_embeddings, noise_pred_original, guide_embeddings_clip, clip_guidance_scale, + num_cutouts, use_cutouts) + + noise_pred = [] + cond_latents = [] + for i in range(len(latents)): + lat1 = latents[i].unsqueeze(0) + tem1 = text_embeddings[i].unsqueeze(0) + npo1 = noise_pred_original[i].unsqueeze(0) + gem1 = guide_embeddings_clip[i].unsqueeze(0) + npr1, cla1 = self.cond_fn1(lat1, timestep, index, tem1, npo1, gem1, clip_guidance_scale, num_cutouts, use_cutouts) + noise_pred.append(npr1) + cond_latents.append(cla1) + + noise_pred = torch.cat(noise_pred) + cond_latents = torch.cat(cond_latents) + return noise_pred, cond_latents + + @torch.enable_grad() + def cond_fn1(self, latents, timestep, index, text_embeddings, noise_pred_original, guide_embeddings_clip, clip_guidance_scale, + num_cutouts, use_cutouts=True, ): + latents = latents.detach().requires_grad_() + + if isinstance(self.scheduler, LMSDiscreteScheduler): + sigma = self.scheduler.sigmas[index] + # the model input needs to be scaled to match the continuous ODE formulation in K-LMS + latent_model_input = latents / ((sigma**2 + 1) ** 0.5) + else: + latent_model_input = latents + + # predict the noise residual + noise_pred = self.unet(latent_model_input, timestep, encoder_hidden_states=text_embeddings).sample + + if isinstance(self.scheduler, (PNDMScheduler, DDIMScheduler)): + alpha_prod_t = self.scheduler.alphas_cumprod[timestep] + beta_prod_t = 1 - alpha_prod_t + # compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf + pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5) + + fac = torch.sqrt(beta_prod_t) + sample = pred_original_sample * (fac) + latents * (1 - fac) + elif isinstance(self.scheduler, LMSDiscreteScheduler): + sigma = self.scheduler.sigmas[index] + sample = latents - sigma * noise_pred + else: + raise ValueError(f"scheduler type {type(self.scheduler)} not supported") + + sample = 1 / 0.18215 * sample + image = self.vae.decode(sample).sample + image = (image / 2 + 0.5).clamp(0, 1) + + if use_cutouts: + image = self.make_cutouts(image, num_cutouts) + else: + image = transforms.Resize(FEATURE_EXTRACTOR_SIZE)(image) + image = self.normalize(image).to(latents.dtype) + + image_embeddings_clip = self.clip_model.get_image_features(image) + image_embeddings_clip = image_embeddings_clip / image_embeddings_clip.norm(p=2, dim=-1, keepdim=True) + + if use_cutouts: + dists = spherical_dist_loss(image_embeddings_clip, guide_embeddings_clip) + dists = dists.view([num_cutouts, sample.shape[0], -1]) + loss = dists.sum(2).mean(0).sum() * clip_guidance_scale + else: + # バッチサイズが複数だと正しく動くかわからない + loss = spherical_dist_loss(image_embeddings_clip, guide_embeddings_clip).mean() * clip_guidance_scale + + grads = -torch.autograd.grad(loss, latents)[0] + + if isinstance(self.scheduler, LMSDiscreteScheduler): + latents = latents.detach() + grads * (sigma**2) + noise_pred = noise_pred_original + else: + noise_pred = noise_pred_original - torch.sqrt(beta_prod_t) * grads + return noise_pred, latents + + # バッチを分解して一件ずつ処理する + def cond_fn_vgg16(self, latents, timestep, index, text_embeddings, noise_pred_original, guide_embeddings, guidance_scale): + if len(latents) == 1: + return self.cond_fn_vgg16_b1(latents, timestep, index, text_embeddings, noise_pred_original, guide_embeddings, guidance_scale) + + noise_pred = [] + cond_latents = [] + for i in range(len(latents)): + lat1 = latents[i].unsqueeze(0) + tem1 = text_embeddings[i].unsqueeze(0) + npo1 = noise_pred_original[i].unsqueeze(0) + gem1 = guide_embeddings[i].unsqueeze(0) + npr1, cla1 = self.cond_fn_vgg16_b1(lat1, timestep, index, tem1, npo1, gem1, guidance_scale) + noise_pred.append(npr1) + cond_latents.append(cla1) + + noise_pred = torch.cat(noise_pred) + cond_latents = torch.cat(cond_latents) + return noise_pred, cond_latents + + # 1件だけ処理する + @torch.enable_grad() + def cond_fn_vgg16_b1(self, latents, timestep, index, text_embeddings, noise_pred_original, guide_embeddings, guidance_scale): + latents = latents.detach().requires_grad_() + + if isinstance(self.scheduler, LMSDiscreteScheduler): + sigma = self.scheduler.sigmas[index] + # the model input needs to be scaled to match the continuous ODE formulation in K-LMS + latent_model_input = latents / ((sigma**2 + 1) ** 0.5) + else: + latent_model_input = latents + + # predict the noise residual + noise_pred = self.unet(latent_model_input, timestep, encoder_hidden_states=text_embeddings).sample + + if isinstance(self.scheduler, (PNDMScheduler, DDIMScheduler)): + alpha_prod_t = self.scheduler.alphas_cumprod[timestep] + beta_prod_t = 1 - alpha_prod_t + # compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf + pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5) + + fac = torch.sqrt(beta_prod_t) + sample = pred_original_sample * (fac) + latents * (1 - fac) + elif isinstance(self.scheduler, LMSDiscreteScheduler): + sigma = self.scheduler.sigmas[index] + sample = latents - sigma * noise_pred + else: + raise ValueError(f"scheduler type {type(self.scheduler)} not supported") + + sample = 1 / 0.18215 * sample + image = self.vae.decode(sample).sample + image = (image / 2 + 0.5).clamp(0, 1) + image = transforms.Resize((image.shape[-2] // VGG16_INPUT_RESIZE_DIV, image.shape[-1] // VGG16_INPUT_RESIZE_DIV))(image) + image = self.vgg16_normalize(image).to(latents.dtype) + + image_embeddings = self.vgg16_feat_model(image)['feat'] + + # バッチサイズが複数だと正しく動くかわからない + loss = ((image_embeddings - guide_embeddings) ** 2).mean() * guidance_scale # MSE style transferでコンテンツの損失はMSEなので + + grads = -torch.autograd.grad(loss, latents)[0] + if isinstance(self.scheduler, LMSDiscreteScheduler): + latents = latents.detach() + grads * (sigma**2) + noise_pred = noise_pred_original + else: + noise_pred = noise_pred_original - torch.sqrt(beta_prod_t) * grads + return noise_pred, latents + + +class MakeCutouts(torch.nn.Module): + def __init__(self, cut_size, cut_power=1.0): + super().__init__() + + self.cut_size = cut_size + self.cut_power = cut_power + + def forward(self, pixel_values, num_cutouts): + sideY, sideX = pixel_values.shape[2:4] + max_size = min(sideX, sideY) + min_size = min(sideX, sideY, self.cut_size) + cutouts = [] + for _ in range(num_cutouts): + size = int(torch.rand([]) ** self.cut_power * (max_size - min_size) + min_size) + offsetx = torch.randint(0, sideX - size + 1, ()) + offsety = torch.randint(0, sideY - size + 1, ()) + cutout = pixel_values[:, :, offsety: offsety + size, offsetx: offsetx + size] + cutouts.append(torch.nn.functional.adaptive_avg_pool2d(cutout, self.cut_size)) + return torch.cat(cutouts) + + +def spherical_dist_loss(x, y): + x = torch.nn.functional.normalize(x, dim=-1) + y = torch.nn.functional.normalize(y, dim=-1) + return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2) + + +re_attention = re.compile( + r""" +\\\(| +\\\)| +\\\[| +\\]| +\\\\| +\\| +\(| +\[| +:([+-]?[.\d]+)\)| +\)| +]| +[^\\()\[\]:]+| +: +""", + re.X, +) + + +def parse_prompt_attention(text): + """ + Parses a string with attention tokens and returns a list of pairs: text and its associated weight. + Accepted tokens are: + (abc) - increases attention to abc by a multiplier of 1.1 + (abc:3.12) - increases attention to abc by a multiplier of 3.12 + [abc] - decreases attention to abc by a multiplier of 1.1 + \( - literal character '(' + \[ - literal character '[' + \) - literal character ')' + \] - literal character ']' + \\ - literal character '\' + anything else - just text + >>> parse_prompt_attention('normal text') + [['normal text', 1.0]] + >>> parse_prompt_attention('an (important) word') + [['an ', 1.0], ['important', 1.1], [' word', 1.0]] + >>> parse_prompt_attention('(unbalanced') + [['unbalanced', 1.1]] + >>> parse_prompt_attention('\(literal\]') + [['(literal]', 1.0]] + >>> parse_prompt_attention('(unnecessary)(parens)') + [['unnecessaryparens', 1.1]] + >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).') + [['a ', 1.0], + ['house', 1.5730000000000004], + [' ', 1.1], + ['on', 1.0], + [' a ', 1.1], + ['hill', 0.55], + [', sun, ', 1.1], + ['sky', 1.4641000000000006], + ['.', 1.1]] + """ + + res = [] + round_brackets = [] + square_brackets = [] + + round_bracket_multiplier = 1.1 + square_bracket_multiplier = 1 / 1.1 + + def multiply_range(start_position, multiplier): + for p in range(start_position, len(res)): + res[p][1] *= multiplier + + for m in re_attention.finditer(text): + text = m.group(0) + weight = m.group(1) + + if text.startswith("\\"): + res.append([text[1:], 1.0]) + elif text == "(": + round_brackets.append(len(res)) + elif text == "[": + square_brackets.append(len(res)) + elif weight is not None and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), float(weight)) + elif text == ")" and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), round_bracket_multiplier) + elif text == "]" and len(square_brackets) > 0: + multiply_range(square_brackets.pop(), square_bracket_multiplier) + else: + res.append([text, 1.0]) + + for pos in round_brackets: + multiply_range(pos, round_bracket_multiplier) + + for pos in square_brackets: + multiply_range(pos, square_bracket_multiplier) + + if len(res) == 0: + res = [["", 1.0]] + + # merge runs of identical weights + i = 0 + while i + 1 < len(res): + if res[i][1] == res[i + 1][1]: + res[i][0] += res[i + 1][0] + res.pop(i + 1) + else: + i += 1 + + return res + + +def get_prompts_with_weights(pipe: PipelineLike, prompt: List[str], max_length: int): + r""" + Tokenize a list of prompts and return its tokens with weights of each token. + No padding, starting or ending token is included. + """ + tokens = [] + weights = [] + truncated = False + for text in prompt: + texts_and_weights = parse_prompt_attention(text) + text_token = [] + text_weight = [] + for word, weight in texts_and_weights: + # tokenize and discard the starting and the ending token + token = pipe.tokenizer(word).input_ids[1:-1] + + token = pipe.replace_token(token) + + text_token += token + # copy the weight by length of token + text_weight += [weight] * len(token) + # stop if the text is too long (longer than truncation limit) + if len(text_token) > max_length: + truncated = True + break + # truncate + if len(text_token) > max_length: + truncated = True + text_token = text_token[:max_length] + text_weight = text_weight[:max_length] + tokens.append(text_token) + weights.append(text_weight) + if truncated: + print("warning: Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples") + return tokens, weights + + +def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, pad, no_boseos_middle=True, chunk_length=77): + r""" + Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length. + """ + max_embeddings_multiples = (max_length - 2) // (chunk_length - 2) + weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length + for i in range(len(tokens)): + tokens[i] = [bos] + tokens[i] + [eos] + [pad] * (max_length - 2 - len(tokens[i])) + if no_boseos_middle: + weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i])) + else: + w = [] + if len(weights[i]) == 0: + w = [1.0] * weights_length + else: + for j in range(max_embeddings_multiples): + w.append(1.0) # weight for starting token in this chunk + w += weights[i][j * (chunk_length - 2): min(len(weights[i]), (j + 1) * (chunk_length - 2))] + w.append(1.0) # weight for ending token in this chunk + w += [1.0] * (weights_length - len(w)) + weights[i] = w[:] + + return tokens, weights + + +def get_unweighted_text_embeddings( + pipe: PipelineLike, + text_input: torch.Tensor, + chunk_length: int, + clip_skip: int, + eos: int, + pad: int, + no_boseos_middle: Optional[bool] = True, +): + """ + When the length of tokens is a multiple of the capacity of the text encoder, + it should be split into chunks and sent to the text encoder individually. + """ + max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2) + if max_embeddings_multiples > 1: + text_embeddings = [] + for i in range(max_embeddings_multiples): + # extract the i-th chunk + text_input_chunk = text_input[:, i * (chunk_length - 2): (i + 1) * (chunk_length - 2) + 2].clone() + + # cover the head and the tail by the starting and the ending tokens + text_input_chunk[:, 0] = text_input[0, 0] + if pad == eos: # v1 + text_input_chunk[:, -1] = text_input[0, -1] + else: # v2 + if text_input_chunk[:, -1] != eos and text_input_chunk[:, -1] != pad: # 最後に普通の文字がある + text_input_chunk[:, -1] = eos + if text_input_chunk[:, 1] == pad: # BOSだけであとはPAD + text_input_chunk[:, 1] = eos + + if clip_skip is None or clip_skip == 1: + text_embedding = pipe.text_encoder(text_input_chunk)[0] + else: + enc_out = pipe.text_encoder(text_input_chunk, output_hidden_states=True, return_dict=True) + text_embedding = enc_out['hidden_states'][-clip_skip] + text_embedding = pipe.text_encoder.text_model.final_layer_norm(text_embedding) + + if no_boseos_middle: + if i == 0: + # discard the ending token + text_embedding = text_embedding[:, :-1] + elif i == max_embeddings_multiples - 1: + # discard the starting token + text_embedding = text_embedding[:, 1:] + else: + # discard both starting and ending tokens + text_embedding = text_embedding[:, 1:-1] + + text_embeddings.append(text_embedding) + text_embeddings = torch.concat(text_embeddings, axis=1) + else: + if clip_skip is None or clip_skip == 1: + text_embeddings = pipe.text_encoder(text_input)[0] + else: + enc_out = pipe.text_encoder(text_input, output_hidden_states=True, return_dict=True) + text_embeddings = enc_out['hidden_states'][-clip_skip] + text_embeddings = pipe.text_encoder.text_model.final_layer_norm(text_embeddings) + return text_embeddings + + +def get_weighted_text_embeddings( + pipe: PipelineLike, + prompt: Union[str, List[str]], + uncond_prompt: Optional[Union[str, List[str]]] = None, + max_embeddings_multiples: Optional[int] = 1, + no_boseos_middle: Optional[bool] = False, + skip_parsing: Optional[bool] = False, + skip_weighting: Optional[bool] = False, + clip_skip=None, + **kwargs, +): + r""" + Prompts can be assigned with local weights using brackets. For example, + prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful', + and the embedding tokens corresponding to the words get multiplied by a constant, 1.1. + Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean. + Args: + pipe (`DiffusionPipeline`): + Pipe to provide access to the tokenizer and the text encoder. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + uncond_prompt (`str` or `List[str]`): + The unconditional prompt or prompts for guide the image generation. If unconditional prompt + is provided, the embeddings of prompt and uncond_prompt are concatenated. + max_embeddings_multiples (`int`, *optional*, defaults to `1`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + no_boseos_middle (`bool`, *optional*, defaults to `False`): + If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and + ending token in each of the chunk in the middle. + skip_parsing (`bool`, *optional*, defaults to `False`): + Skip the parsing of brackets. + skip_weighting (`bool`, *optional*, defaults to `False`): + Skip the weighting. When the parsing is skipped, it is forced True. + """ + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + if isinstance(prompt, str): + prompt = [prompt] + + if not skip_parsing: + prompt_tokens, prompt_weights = get_prompts_with_weights(pipe, prompt, max_length - 2) + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens, uncond_weights = get_prompts_with_weights(pipe, uncond_prompt, max_length - 2) + else: + prompt_tokens = [ + token[1:-1] for token in pipe.tokenizer(prompt, max_length=max_length, truncation=True).input_ids + ] + prompt_weights = [[1.0] * len(token) for token in prompt_tokens] + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens = [ + token[1:-1] + for token in pipe.tokenizer(uncond_prompt, max_length=max_length, truncation=True).input_ids + ] + uncond_weights = [[1.0] * len(token) for token in uncond_tokens] + + # round up the longest length of tokens to a multiple of (model_max_length - 2) + max_length = max([len(token) for token in prompt_tokens]) + if uncond_prompt is not None: + max_length = max(max_length, max([len(token) for token in uncond_tokens])) + + max_embeddings_multiples = min( + max_embeddings_multiples, + (max_length - 1) // (pipe.tokenizer.model_max_length - 2) + 1, + ) + max_embeddings_multiples = max(1, max_embeddings_multiples) + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + + # pad the length of tokens and weights + bos = pipe.tokenizer.bos_token_id + eos = pipe.tokenizer.eos_token_id + pad = pipe.tokenizer.pad_token_id + prompt_tokens, prompt_weights = pad_tokens_and_weights( + prompt_tokens, + prompt_weights, + max_length, + bos, + eos, + pad, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=pipe.device) + if uncond_prompt is not None: + uncond_tokens, uncond_weights = pad_tokens_and_weights( + uncond_tokens, + uncond_weights, + max_length, + bos, + eos, + pad, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + uncond_tokens = torch.tensor(uncond_tokens, dtype=torch.long, device=pipe.device) + + # get the embeddings + text_embeddings = get_unweighted_text_embeddings( + pipe, + prompt_tokens, + pipe.tokenizer.model_max_length, + clip_skip, + eos, pad, + no_boseos_middle=no_boseos_middle, + ) + prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=pipe.device) + if uncond_prompt is not None: + uncond_embeddings = get_unweighted_text_embeddings( + pipe, + uncond_tokens, + pipe.tokenizer.model_max_length, + clip_skip, + eos, pad, + no_boseos_middle=no_boseos_middle, + ) + uncond_weights = torch.tensor(uncond_weights, dtype=uncond_embeddings.dtype, device=pipe.device) + + # assign weights to the prompts and normalize in the sense of mean + # TODO: should we normalize by chunk or in a whole (current implementation)? + # →全体でいいんじゃないかな + if (not skip_parsing) and (not skip_weighting): + previous_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= prompt_weights.unsqueeze(-1) + current_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + if uncond_prompt is not None: + previous_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= uncond_weights.unsqueeze(-1) + current_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + + if uncond_prompt is not None: + return text_embeddings, uncond_embeddings, prompt_tokens + return text_embeddings, None, prompt_tokens + + +def preprocess_guide_image(image): + image = image.resize(FEATURE_EXTRACTOR_SIZE, resample=Image.NEAREST) # cond_fnと合わせる + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) # nchw + image = torch.from_numpy(image) + return image # 0 to 1 + + +# VGG16の入力は任意サイズでよいので入力画像を適宜リサイズする +def preprocess_vgg16_guide_image(image, size): + image = image.resize(size, resample=Image.NEAREST) # cond_fnと合わせる + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) # nchw + image = torch.from_numpy(image) + return image # 0 to 1 + + +def preprocess_image(image): + w, h = image.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask): + mask = mask.convert("L") + w, h = mask.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + mask = mask.resize((w // 8, h // 8), resample=PIL.Image.LANCZOS) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = mask[None].transpose(0, 1, 2, 3) # what does this step do? + mask = 1 - mask # repaint white, keep black + mask = torch.from_numpy(mask) + return mask + + +# endregion + + +# def load_clip_l14_336(dtype): +# print(f"loading CLIP: {CLIP_ID_L14_336}") +# text_encoder = CLIPTextModel.from_pretrained(CLIP_ID_L14_336, torch_dtype=dtype) +# return text_encoder + + +def main(args): + if args.fp16: + dtype = torch.float16 + elif args.bf16: + dtype = torch.bfloat16 + else: + dtype = torch.float32 + + highres_fix = args.highres_fix_scale is not None + assert not highres_fix or args.image_path is None, f"highres_fix doesn't work with img2img / highres_fixはimg2imgと同時に使えません" + + if args.v_parameterization and not args.v2: + print("v_parameterization should be with v2 / v1でv_parameterizationを使用することは想定されていません") + if args.v2 and args.clip_skip is not None: + print("v2 with clip_skip will be unexpected / v2でclip_skipを使用することは想定されていません") + + # モデルを読み込む + if not os.path.isfile(args.ckpt): # ファイルがないならパターンで探し、一つだけ該当すればそれを使う + files = glob.glob(args.ckpt) + if len(files) == 1: + args.ckpt = files[0] + + use_stable_diffusion_format = os.path.isfile(args.ckpt) + if use_stable_diffusion_format: + print("load StableDiffusion checkpoint") + text_encoder, vae, unet = model_util.load_models_from_stable_diffusion_checkpoint(args.v2, args.ckpt) + else: + print("load Diffusers pretrained models") + loading_pipe = StableDiffusionPipeline.from_pretrained(args.ckpt, safety_checker=None, torch_dtype=dtype) + text_encoder = loading_pipe.text_encoder + vae = loading_pipe.vae + unet = loading_pipe.unet + tokenizer = loading_pipe.tokenizer + del loading_pipe + + # VAEを読み込む + if args.vae is not None: + vae = model_util.load_vae(args.vae, dtype) + print("additional VAE loaded") + + # # 置換するCLIPを読み込む + # if args.replace_clip_l14_336: + # text_encoder = load_clip_l14_336(dtype) + # print(f"large clip {CLIP_ID_L14_336} is loaded") + + if args.clip_guidance_scale > 0.0 or args.clip_image_guidance_scale: + print("prepare clip model") + clip_model = CLIPModel.from_pretrained(CLIP_MODEL_PATH, torch_dtype=dtype) + else: + clip_model = None + + if args.vgg16_guidance_scale > 0.0: + print("prepare resnet model") + vgg16_model = torchvision.models.vgg16(torchvision.models.VGG16_Weights.IMAGENET1K_V1) + else: + vgg16_model = None + + # xformers、Hypernetwork対応 + if not args.diffusers_xformers: + replace_unet_modules(unet, not args.xformers, args.xformers) + + # tokenizerを読み込む + print("loading tokenizer") + if use_stable_diffusion_format: + if args.v2: + tokenizer = CLIPTokenizer.from_pretrained(V2_STABLE_DIFFUSION_PATH, subfolder="tokenizer") + else: + tokenizer = CLIPTokenizer.from_pretrained(TOKENIZER_PATH) # , model_max_length=max_token_length + 2) + + # schedulerを用意する + sched_init_args = {} + scheduler_num_noises_per_step = 1 + if args.sampler == "ddim": + scheduler_cls = DDIMScheduler + scheduler_module = diffusers.schedulers.scheduling_ddim + elif args.sampler == "ddpm": # ddpmはおかしくなるのでoptionから外してある + scheduler_cls = DDPMScheduler + scheduler_module = diffusers.schedulers.scheduling_ddpm + elif args.sampler == "pndm": + scheduler_cls = PNDMScheduler + scheduler_module = diffusers.schedulers.scheduling_pndm + elif args.sampler == 'lms' or args.sampler == 'k_lms': + scheduler_cls = LMSDiscreteScheduler + scheduler_module = diffusers.schedulers.scheduling_lms_discrete + elif args.sampler == 'euler' or args.sampler == 'k_euler': + scheduler_cls = EulerDiscreteScheduler + scheduler_module = diffusers.schedulers.scheduling_euler_discrete + elif args.sampler == 'euler_a' or args.sampler == 'k_euler_a': + scheduler_cls = EulerAncestralDiscreteScheduler + scheduler_module = diffusers.schedulers.scheduling_euler_ancestral_discrete + elif args.sampler == "dpmsolver" or args.sampler == "dpmsolver++": + scheduler_cls = DPMSolverMultistepScheduler + sched_init_args['algorithm_type'] = args.sampler + scheduler_module = diffusers.schedulers.scheduling_dpmsolver_multistep + elif args.sampler == "dpmsingle": + scheduler_cls = DPMSolverSinglestepScheduler + scheduler_module = diffusers.schedulers.scheduling_dpmsolver_singlestep + elif args.sampler == "heun": + scheduler_cls = HeunDiscreteScheduler + scheduler_module = diffusers.schedulers.scheduling_heun_discrete + elif args.sampler == 'dpm_2' or args.sampler == 'k_dpm_2': + scheduler_cls = KDPM2DiscreteScheduler + scheduler_module = diffusers.schedulers.scheduling_k_dpm_2_discrete + elif args.sampler == 'dpm_2_a' or args.sampler == 'k_dpm_2_a': + scheduler_cls = KDPM2AncestralDiscreteScheduler + scheduler_module = diffusers.schedulers.scheduling_k_dpm_2_ancestral_discrete + scheduler_num_noises_per_step = 2 + + if args.v_parameterization: + sched_init_args['prediction_type'] = 'v_prediction' + + # samplerの乱数をあらかじめ指定するための処理 + + # replace randn + class NoiseManager: + def __init__(self): + self.sampler_noises = None + self.sampler_noise_index = 0 + + def reset_sampler_noises(self, noises): + self.sampler_noise_index = 0 + self.sampler_noises = noises + + def randn(self, shape, device=None, dtype=None, layout=None, generator=None): + # print("replacing", shape, len(self.sampler_noises), self.sampler_noise_index) + if self.sampler_noises is not None and self.sampler_noise_index < len(self.sampler_noises): + noise = self.sampler_noises[self.sampler_noise_index] + if shape != noise.shape: + noise = None + else: + noise = None + + if noise == None: + print(f"unexpected noise request: {self.sampler_noise_index}, {shape}") + noise = torch.randn(shape, dtype=dtype, device=device, generator=generator) + + self.sampler_noise_index += 1 + return noise + + class TorchRandReplacer: + def __init__(self, noise_manager): + self.noise_manager = noise_manager + + def __getattr__(self, item): + if item == 'randn': + return self.noise_manager.randn + if hasattr(torch, item): + return getattr(torch, item) + raise AttributeError("'{}' object has no attribute '{}'".format(type(self).__name__, item)) + + noise_manager = NoiseManager() + if scheduler_module is not None: + scheduler_module.torch = TorchRandReplacer(noise_manager) + + scheduler = scheduler_cls(num_train_timesteps=SCHEDULER_TIMESTEPS, + beta_start=SCHEDULER_LINEAR_START, beta_end=SCHEDULER_LINEAR_END, + beta_schedule=SCHEDLER_SCHEDULE, **sched_init_args) + + # clip_sample=Trueにする + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is False: + print("set clip_sample to True") + scheduler.config.clip_sample = True + + # deviceを決定する + device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # "mps"を考量してない + + # custom pipelineをコピったやつを生成する + vae.to(dtype).to(device) + text_encoder.to(dtype).to(device) + unet.to(dtype).to(device) + if clip_model is not None: + clip_model.to(dtype).to(device) + if vgg16_model is not None: + vgg16_model.to(dtype).to(device) + + # networkを組み込む + if args.network_module: + networks = [] + for i, network_module in enumerate(args.network_module): + print("import network module:", network_module) + imported_module = importlib.import_module(network_module) + + network_mul = 1.0 if args.network_mul is None or len(args.network_mul) <= i else args.network_mul[i] + + net_kwargs = {} + if args.network_args and i < len(args.network_args): + network_args = args.network_args[i] + # TODO escape special chars + network_args = network_args.split(";") + for net_arg in network_args: + key, value = net_arg.split("=") + net_kwargs[key] = value + + if args.network_weights and i < len(args.network_weights): + network_weight = args.network_weights[i] + print("load network weights from:", network_weight) + + if model_util.is_safetensors(network_weight): + from safetensors.torch import safe_open + with safe_open(network_weight, framework="pt") as f: + metadata = f.metadata() + if metadata is not None: + print(f"metadata for: {network_weight}: {metadata}") + + network = imported_module.create_network_from_weights(network_mul, network_weight, vae, text_encoder, unet, **net_kwargs) + else: + raise ValueError("No weight. Weight is required.") + if network is None: + return + + network.apply_to(text_encoder, unet) + + if args.opt_channels_last: + network.to(memory_format=torch.channels_last) + network.to(dtype).to(device) + + networks.append(network) + else: + networks = [] + + if args.opt_channels_last: + print(f"set optimizing: channels last") + text_encoder.to(memory_format=torch.channels_last) + vae.to(memory_format=torch.channels_last) + unet.to(memory_format=torch.channels_last) + if clip_model is not None: + clip_model.to(memory_format=torch.channels_last) + if networks: + for network in networks: + network.to(memory_format=torch.channels_last) + if vgg16_model is not None: + vgg16_model.to(memory_format=torch.channels_last) + + pipe = PipelineLike(device, vae, text_encoder, tokenizer, unet, scheduler, args.clip_skip, + clip_model, args.clip_guidance_scale, args.clip_image_guidance_scale, + vgg16_model, args.vgg16_guidance_scale, args.vgg16_guidance_layer) + print("pipeline is ready.") + if len(networks) > 0: + pipe.set_lora_network(networks[0]) + + if args.diffusers_xformers: + pipe.enable_xformers_memory_efficient_attention() + + # Textual Inversionを処理する + if args.textual_inversion_embeddings: + token_ids_embeds = [] + for embeds_file in args.textual_inversion_embeddings: + if model_util.is_safetensors(embeds_file): + from safetensors.torch import load_file + data = load_file(embeds_file) + else: + data = torch.load(embeds_file, map_location="cpu") + + embeds = next(iter(data.values())) + if type(embeds) != torch.Tensor: + raise ValueError(f"weight file does not contains Tensor / 重みファイルのデータがTensorではありません: {embeds_file}") + + num_vectors_per_token = embeds.size()[0] + token_string = os.path.splitext(os.path.basename(embeds_file))[0] + token_strings = [token_string] + [f"{token_string}{i+1}" for i in range(num_vectors_per_token - 1)] + + # add new word to tokenizer, count is num_vectors_per_token + num_added_tokens = tokenizer.add_tokens(token_strings) + assert num_added_tokens == num_vectors_per_token, f"tokenizer has same word to token string (filename). please rename the file / 指定した名前（ファイル名）のトークンが既に存在します。ファイルをリネームしてください: {embeds_file}" + + token_ids = tokenizer.convert_tokens_to_ids(token_strings) + print(f"Textual Inversion embeddings `{token_string}` loaded. Tokens are added: {token_ids}") + assert min(token_ids) == token_ids[0] and token_ids[-1] == token_ids[0] + len(token_ids) - 1, f"token ids is not ordered" + assert len(tokenizer) - 1 == token_ids[-1], f"token ids is not end of tokenize: {len(tokenizer)}" + + if num_vectors_per_token > 1: + pipe.add_token_replacement(token_ids[0], token_ids) + + token_ids_embeds.append((token_ids, embeds)) + + text_encoder.resize_token_embeddings(len(tokenizer)) + token_embeds = text_encoder.get_input_embeddings().weight.data + for token_ids, embeds in token_ids_embeds: + for token_id, embed in zip(token_ids, embeds): + token_embeds[token_id] = embed + + # promptを取得する + if args.from_file is not None: + print(f"reading prompts from {args.from_file}") + with open(args.from_file, "r", encoding="utf-8") as f: + prompt_list = f.read().splitlines() + prompt_list = [d for d in prompt_list if len(d.strip()) > 0] + elif args.prompt is not None: + prompt_list = [args.prompt] + else: + prompt_list = [] + + if args.interactive: + args.n_iter = 1 + + # img2imgの前処理、画像の読み込みなど + def load_images(path): + if os.path.isfile(path): + paths = [path] + else: + paths = glob.glob(os.path.join(path, "*.png")) + glob.glob(os.path.join(path, "*.jpg")) + \ + glob.glob(os.path.join(path, "*.jpeg")) + glob.glob(os.path.join(path, "*.webp")) + paths.sort() + + images = [] + for p in paths: + image = Image.open(p) + if image.mode != "RGB": + print(f"convert image to RGB from {image.mode}: {p}") + image = image.convert("RGB") + images.append(image) + + return images + + def resize_images(imgs, size): + resized = [] + for img in imgs: + r_img = img.resize(size, Image.Resampling.LANCZOS) + if hasattr(img, 'filename'): # filename属性がない場合があるらしい + r_img.filename = img.filename + resized.append(r_img) + return resized + + if args.image_path is not None: + print(f"load image for img2img: {args.image_path}") + init_images = load_images(args.image_path) + assert len(init_images) > 0, f"No image / 画像がありません: {args.image_path}" + print(f"loaded {len(init_images)} images for img2img") + else: + init_images = None + + if args.mask_path is not None: + print(f"load mask for inpainting: {args.mask_path}") + mask_images = load_images(args.mask_path) + assert len(mask_images) > 0, f"No mask image / マスク画像がありません: {args.image_path}" + print(f"loaded {len(mask_images)} mask images for inpainting") + else: + mask_images = None + + # promptがないとき、画像のPngInfoから取得する + if init_images is not None and len(prompt_list) == 0 and not args.interactive: + print("get prompts from images' meta data") + for img in init_images: + if 'prompt' in img.text: + prompt = img.text['prompt'] + if 'negative-prompt' in img.text: + prompt += " --n " + img.text['negative-prompt'] + prompt_list.append(prompt) + + # プロンプトと画像を一致させるため指定回数だけ繰り返す（画像を増幅する） + l = [] + for im in init_images: + l.extend([im] * args.images_per_prompt) + init_images = l + + if mask_images is not None: + l = [] + for im in mask_images: + l.extend([im] * args.images_per_prompt) + mask_images = l + + # 画像サイズにオプション指定があるときはリサイズする + if init_images is not None and args.W is not None and args.H is not None: + print(f"resize img2img source images to {args.W}*{args.H}") + init_images = resize_images(init_images, (args.W, args.H)) + if mask_images is not None: + print(f"resize img2img mask images to {args.W}*{args.H}") + mask_images = resize_images(mask_images, (args.W, args.H)) + + prev_image = None # for VGG16 guided + if args.guide_image_path is not None: + print(f"load image for CLIP/VGG16 guidance: {args.guide_image_path}") + guide_images = load_images(args.guide_image_path) + print(f"loaded {len(guide_images)} guide images for CLIP/VGG16 guidance") + if len(guide_images) == 0: + print(f"No guide image, use previous generated image. / ガイド画像がありません。直前に生成した画像を使います: {args.image_path}") + guide_images = None + else: + guide_images = None + + # seed指定時はseedを決めておく + if args.seed is not None: + random.seed(args.seed) + predefined_seeds = [random.randint(0, 0x7fffffff) for _ in range(args.n_iter * len(prompt_list) * args.images_per_prompt)] + if len(predefined_seeds) == 1: + predefined_seeds[0] = args.seed + else: + predefined_seeds = None + + # デフォルト画像サイズを設定する：img2imgではこれらの値は無視される（またはW*Hにリサイズ済み） + if args.W is None: + args.W = 512 + if args.H is None: + args.H = 512 + + # 画像生成のループ + os.makedirs(args.outdir, exist_ok=True) + max_embeddings_multiples = 1 if args.max_embeddings_multiples is None else args.max_embeddings_multiples + + for gen_iter in range(args.n_iter): + print(f"iteration {gen_iter+1}/{args.n_iter}") + iter_seed = random.randint(0, 0x7fffffff) + + # バッチ処理の関数 + def process_batch(batch, highres_fix, highres_1st=False): + batch_size = len(batch) + + # highres_fixの処理 + if highres_fix and not highres_1st: + # 1st stageのバッチを作成して呼び出す + print("process 1st stage1") + batch_1st = [] + for params1, (width, height, steps, scale, negative_scale, strength) in batch: + width_1st = int(width * args.highres_fix_scale + .5) + height_1st = int(height * args.highres_fix_scale + .5) + width_1st = width_1st - width_1st % 32 + height_1st = height_1st - height_1st % 32 + batch_1st.append((params1, (width_1st, height_1st, args.highres_fix_steps, scale, negative_scale, strength))) + images_1st = process_batch(batch_1st, True, True) + + # 2nd stageのバッチを作成して以下処理する + print("process 2nd stage1") + batch_2nd = [] + for i, (b1, image) in enumerate(zip(batch, images_1st)): + image = image.resize((width, height), resample=PIL.Image.LANCZOS) + (step, prompt, negative_prompt, seed, _, _, clip_prompt, guide_image), params2 = b1 + batch_2nd.append(((step, prompt, negative_prompt, seed+1, image, None, clip_prompt, guide_image), params2)) + batch = batch_2nd + + (step_first, _, _, _, init_image, mask_image, _, guide_image), (width, + height, steps, scale, negative_scale, strength) = batch[0] + noise_shape = (LATENT_CHANNELS, height // DOWNSAMPLING_FACTOR, width // DOWNSAMPLING_FACTOR) + + prompts = [] + negative_prompts = [] + start_code = torch.zeros((batch_size, *noise_shape), device=device, dtype=dtype) + noises = [torch.zeros((batch_size, *noise_shape), device=device, dtype=dtype) + for _ in range(steps * scheduler_num_noises_per_step)] + seeds = [] + clip_prompts = [] + + if init_image is not None: # img2img? + i2i_noises = torch.zeros((batch_size, *noise_shape), device=device, dtype=dtype) + init_images = [] + + if mask_image is not None: + mask_images = [] + else: + mask_images = None + else: + i2i_noises = None + init_images = None + mask_images = None + + if guide_image is not None: # CLIP image guided? + guide_images = [] + else: + guide_images = None + + # バッチ内の位置に関わらず同じ乱数を使うためにここで乱数を生成しておく。あわせてimage/maskがbatch内で同一かチェックする + all_images_are_same = True + all_masks_are_same = True + all_guide_images_are_same = True + for i, ((_, prompt, negative_prompt, seed, init_image, mask_image, clip_prompt, guide_image), _) in enumerate(batch): + prompts.append(prompt) + negative_prompts.append(negative_prompt) + seeds.append(seed) + clip_prompts.append(clip_prompt) + + if init_image is not None: + init_images.append(init_image) + if i > 0 and all_images_are_same: + all_images_are_same = init_images[-2] is init_image + + if mask_image is not None: + mask_images.append(mask_image) + if i > 0 and all_masks_are_same: + all_masks_are_same = mask_images[-2] is mask_image + + if guide_image is not None: + guide_images.append(guide_image) + if i > 0 and all_guide_images_are_same: + all_guide_images_are_same = guide_images[-2] is guide_image + + # make start code + torch.manual_seed(seed) + start_code[i] = torch.randn(noise_shape, device=device, dtype=dtype) + + # make each noises + for j in range(steps * scheduler_num_noises_per_step): + noises[j][i] = torch.randn(noise_shape, device=device, dtype=dtype) + + if i2i_noises is not None: # img2img noise + i2i_noises[i] = torch.randn(noise_shape, device=device, dtype=dtype) + + noise_manager.reset_sampler_noises(noises) + + # すべての画像が同じなら1枚だけpipeに渡すことでpipe側で処理を高速化する + if init_images is not None and all_images_are_same: + init_images = init_images[0] + if mask_images is not None and all_masks_are_same: + mask_images = mask_images[0] + if guide_images is not None and all_guide_images_are_same: + guide_images = guide_images[0] + + # generate + images = pipe(prompts, negative_prompts, init_images, mask_images, height, width, steps, scale, negative_scale, strength, latents=start_code, + output_type='pil', max_embeddings_multiples=max_embeddings_multiples, img2img_noise=i2i_noises, clip_prompts=clip_prompts, clip_guide_images=guide_images)[0] + if highres_1st and not args.highres_fix_save_1st: + return images + + # save image + highres_prefix = ("0" if highres_1st else "1") if highres_fix else "" + ts_str = time.strftime('%Y%m%d%H%M%S', time.localtime()) + for i, (image, prompt, negative_prompts, seed, clip_prompt) in enumerate(zip(images, prompts, negative_prompts, seeds, clip_prompts)): + metadata = PngInfo() + metadata.add_text("prompt", prompt) + metadata.add_text("seed", str(seed)) + metadata.add_text("sampler", args.sampler) + metadata.add_text("steps", str(steps)) + metadata.add_text("scale", str(scale)) + if negative_prompt is not None: + metadata.add_text("negative-prompt", negative_prompt) + if negative_scale is not None: + metadata.add_text("negative-scale", str(negative_scale)) + if clip_prompt is not None: + metadata.add_text("clip-prompt", clip_prompt) + + if args.use_original_file_name and init_images is not None: + if type(init_images) is list: + fln = os.path.splitext(os.path.basename(init_images[i % len(init_images)].filename))[0] + ".png" + else: + fln = os.path.splitext(os.path.basename(init_images.filename))[0] + ".png" + elif args.sequential_file_name: + fln = f"im_{highres_prefix}{step_first + i + 1:06d}.png" + else: + fln = f"im_{ts_str}_{highres_prefix}{i:03d}_{seed}.png" + + image.save(os.path.join(args.outdir, fln), pnginfo=metadata) + + if not args.no_preview and not highres_1st and args.interactive: + try: + import cv2 + for prompt, image in zip(prompts, images): + cv2.imshow(prompt[:128], np.array(image)[:, :, ::-1]) # プロンプトが長いと死ぬ + cv2.waitKey() + cv2.destroyAllWindows() + except ImportError: + print("opencv-python is not installed, cannot preview / opencv-pythonがインストールされていないためプレビューできません") + + return images + + # 画像生成のプロンプトが一周するまでのループ + prompt_index = 0 + global_step = 0 + batch_data = [] + while args.interactive or prompt_index < len(prompt_list): + if len(prompt_list) == 0: + # interactive + valid = False + while not valid: + print("\nType prompt:") + try: + prompt = input() + except EOFError: + break + + valid = len(prompt.strip().split(' --')[0].strip()) > 0 + if not valid: # EOF, end app + break + else: + prompt = prompt_list[prompt_index] + + # parse prompt + width = args.W + height = args.H + scale = args.scale + negative_scale = args.negative_scale + steps = args.steps + seeds = None + strength = 0.8 if args.strength is None else args.strength + negative_prompt = "" + clip_prompt = None + + prompt_args = prompt.strip().split(' --') + prompt = prompt_args[0] + print(f"prompt {prompt_index+1}/{len(prompt_list)}: {prompt}") + + for parg in prompt_args[1:]: + try: + m = re.match(r'w (\d+)', parg, re.IGNORECASE) + if m: + width = int(m.group(1)) + print(f"width: {width}") + continue + + m = re.match(r'h (\d+)', parg, re.IGNORECASE) + if m: + height = int(m.group(1)) + print(f"height: {height}") + continue + + m = re.match(r's (\d+)', parg, re.IGNORECASE) + if m: # steps + steps = max(1, min(1000, int(m.group(1)))) + print(f"steps: {steps}") + continue + + m = re.match(r'd ([\d,]+)', parg, re.IGNORECASE) + if m: # seed + seeds = [int(d) for d in m.group(1).split(',')] + print(f"seeds: {seeds}") + continue + + m = re.match(r'l ([\d\.]+)', parg, re.IGNORECASE) + if m: # scale + scale = float(m.group(1)) + print(f"scale: {scale}") + continue + + m = re.match(r'nl ([\d\.]+|none|None)', parg, re.IGNORECASE) + if m: # negative scale + if m.group(1).lower() == 'none': + negative_scale = None + else: + negative_scale = float(m.group(1)) + print(f"negative scale: {negative_scale}") + continue + + m = re.match(r't ([\d\.]+)', parg, re.IGNORECASE) + if m: # strength + strength = float(m.group(1)) + print(f"strength: {strength}") + continue + + m = re.match(r'n (.+)', parg, re.IGNORECASE) + if m: # negative prompt + negative_prompt = m.group(1) + print(f"negative prompt: {negative_prompt}") + continue + + m = re.match(r'c (.+)', parg, re.IGNORECASE) + if m: # clip prompt + clip_prompt = m.group(1) + print(f"clip prompt: {clip_prompt}") + continue + except ValueError as ex: + print(f"Exception in parsing / 解析エラー: {parg}") + print(ex) + + if seeds is not None: + # 数が足りないなら繰り返す + if len(seeds) < args.images_per_prompt: + seeds = seeds * int(math.ceil(args.images_per_prompt / len(seeds))) + seeds = seeds[:args.images_per_prompt] + else: + if predefined_seeds is not None: + seeds = predefined_seeds[-args.images_per_prompt:] + predefined_seeds = predefined_seeds[:-args.images_per_prompt] + elif args.iter_same_seed: + seeds = [iter_seed] * args.images_per_prompt + else: + seeds = [random.randint(0, 0x7fffffff) for _ in range(args.images_per_prompt)] + if args.interactive: + print(f"seed: {seeds}") + + init_image = mask_image = guide_image = None + for seed in seeds: # images_per_promptの数だけ + # 同一イメージを使うとき、本当はlatentに変換しておくと無駄がないが面倒なのでとりあえず毎回処理する + if init_images is not None: + init_image = init_images[global_step % len(init_images)] + + # 32単位に丸めたやつにresizeされるので踏襲する + width, height = init_image.size + width = width - width % 32 + height = height - height % 32 + if width != init_image.size[0] or height != init_image.size[1]: + print(f"img2img image size is not divisible by 32 so aspect ratio is changed / img2imgの画像サイズが32で割り切れないためリサイズされます。画像が歪みます") + + if mask_images is not None: + mask_image = mask_images[global_step % len(mask_images)] + + if guide_images is not None: + guide_image = guide_images[global_step % len(guide_images)] + elif args.clip_image_guidance_scale > 0 or args.vgg16_guidance_scale > 0: + if prev_image is None: + print("Generate 1st image without guide image.") + else: + print("Use previous image as guide image.") + guide_image = prev_image + + # TODO named tupleか何かにする + b1 = ((global_step, prompt, negative_prompt, seed, init_image, mask_image, clip_prompt, guide_image), + (width, height, steps, scale, negative_scale, strength)) + if len(batch_data) > 0 and batch_data[-1][1] != b1[1]: # バッチ分割必要？ + process_batch(batch_data, highres_fix) + batch_data.clear() + + batch_data.append(b1) + if len(batch_data) == args.batch_size: + prev_image = process_batch(batch_data, highres_fix)[0] + batch_data.clear() + + global_step += 1 + + prompt_index += 1 + + if len(batch_data) > 0: + process_batch(batch_data, highres_fix) + batch_data.clear() + + print("done!") + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + + parser.add_argument("--v2", action='store_true', help='load Stable Diffusion v2.0 model / Stable Diffusion 2.0のモデルを読み込む') + parser.add_argument("--v_parameterization", action='store_true', + help='enable v-parameterization training / v-parameterization学習を有効にする') + parser.add_argument("--prompt", type=str, default=None, help="prompt / プロンプト") + parser.add_argument("--from_file", type=str, default=None, + help="if specified, load prompts from this file / 指定時はプロンプトをファイルから読み込む") + parser.add_argument("--interactive", action='store_true', help='interactive mode (generates one image) / 対話モード（生成される画像は1枚になります）') + parser.add_argument("--no_preview", action='store_true', help='do not show generated image in interactive mode / 対話モードで画像を表示しない') + parser.add_argument("--image_path", type=str, default=None, help="image to inpaint or to generate from / img2imgまたはinpaintを行う元画像") + parser.add_argument("--mask_path", type=str, default=None, help="mask in inpainting / inpaint時のマスク") + parser.add_argument("--strength", type=float, default=None, help="img2img strength / img2img時のstrength") + parser.add_argument("--images_per_prompt", type=int, default=1, help="number of images per prompt / プロンプトあたりの出力枚数") + parser.add_argument("--outdir", type=str, default="outputs", help="dir to write results to / 生成画像の出力先") + parser.add_argument("--sequential_file_name", action='store_true', help="sequential output file name / 生成画像のファイル名を連番にする") + parser.add_argument("--use_original_file_name", action='store_true', + help="prepend original file name in img2img / img2imgで元画像のファイル名を生成画像のファイル名の先頭に付ける") + # parser.add_argument("--ddim_eta", type=float, default=0.0, help="ddim eta (eta=0.0 corresponds to deterministic sampling", ) + parser.add_argument("--n_iter", type=int, default=1, help="sample this often / 繰り返し回数") + parser.add_argument("--H", type=int, default=None, help="image height, in pixel space / 生成画像高さ") + parser.add_argument("--W", type=int, default=None, help="image width, in pixel space / 生成画像幅") + parser.add_argument("--batch_size", type=int, default=1, help="batch size / バッチサイズ") + parser.add_argument("--steps", type=int, default=50, help="number of ddim sampling steps / サンプリングステップ数") + parser.add_argument('--sampler', type=str, default='ddim', + choices=['ddim', 'pndm', 'lms', 'euler', 'euler_a', 'heun', 'dpm_2', 'dpm_2_a', 'dpmsolver', + 'dpmsolver++', 'dpmsingle', + 'k_lms', 'k_euler', 'k_euler_a', 'k_dpm_2', 'k_dpm_2_a'], + help=f'sampler (scheduler) type / サンプラー（スケジューラ）の種類') + parser.add_argument("--scale", type=float, default=7.5, + help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty)) / guidance scale") + parser.add_argument("--ckpt", type=str, default=None, help="path to checkpoint of model / モデルのcheckpointファイルまたはディレクトリ") + parser.add_argument("--vae", type=str, default=None, + help="path to checkpoint of vae to replace / VAEを入れ替える場合、VAEのcheckpointファイルまたはディレクトリ") + # parser.add_argument("--replace_clip_l14_336", action='store_true', + # help="Replace CLIP (Text Encoder) to l/14@336 / CLIP(Text Encoder)をl/14@336に入れ替える") + parser.add_argument("--seed", type=int, default=None, + help="seed, or seed of seeds in multiple generation / 1枚生成時のseed、または複数枚生成時の乱数seedを決めるためのseed") + parser.add_argument("--iter_same_seed", action='store_true', + help='use same seed for all prompts in iteration if no seed specified / 乱数seedの指定がないとき繰り返し内はすべて同じseedを使う（プロンプト間の差異の比較用）') + parser.add_argument("--fp16", action='store_true', help='use fp16 / fp16を指定し省メモリ化する') + parser.add_argument("--bf16", action='store_true', help='use bfloat16 / bfloat16を指定し省メモリ化する') + parser.add_argument("--xformers", action='store_true', help='use xformers / xformersを使用し高速化する') + parser.add_argument("--diffusers_xformers", action='store_true', + help='use xformers by diffusers (Hypernetworks doesn\'t work) / Diffusersでxformersを使用する（Hypernetwork利用不可）') + parser.add_argument("--opt_channels_last", action='store_true', + help='set channels last option to model / モデルにchannels lastを指定し最適化する') + parser.add_argument("--network_module", type=str, default=None, nargs='*', + help='Hypernetwork module to use / Hypernetworkを使う時そのモジュール名') + parser.add_argument("--network_weights", type=str, default=None, nargs='*', + help='Hypernetwork weights to load / Hypernetworkの重み') + parser.add_argument("--network_mul", type=float, default=None, nargs='*', help='Hypernetwork multiplier / Hypernetworkの効果の倍率') + parser.add_argument("--network_args", type=str, default=None, nargs='*', + help='additional argmuments for network (key=value) / ネットワークへの追加の引数') + parser.add_argument("--textual_inversion_embeddings", type=str, default=None, nargs='*', + help='Embeddings files of Textual Inversion / Textual Inversionのembeddings') + parser.add_argument("--clip_skip", type=int, default=None, help='layer number from bottom to use in CLIP / CLIPの後ろからn層目の出力を使う') + parser.add_argument("--max_embeddings_multiples", type=int, default=None, + help='max embeding multiples, max token length is 75 * multiples / トークン長をデフォルトの何倍とするか 75*この値 がトークン長となる') + parser.add_argument("--clip_guidance_scale", type=float, default=0.0, + help='enable CLIP guided SD, scale for guidance (DDIM, PNDM, LMS samplers only) / CLIP guided SDを有効にしてこのscaleを適用する（サンプラーはDDIM、PNDM、LMSのみ）') + parser.add_argument("--clip_image_guidance_scale", type=float, default=0.0, + help='enable CLIP guided SD by image, scale for guidance / 画像によるCLIP guided SDを有効にしてこのscaleを適用する') + parser.add_argument("--vgg16_guidance_scale", type=float, default=0.0, + help='enable VGG16 guided SD by image, scale for guidance / 画像によるVGG16 guided SDを有効にしてこのscaleを適用する') + parser.add_argument("--vgg16_guidance_layer", type=int, default=20, + help='layer of VGG16 to calculate contents guide (1~30, 20 for conv4_2) / VGG16のcontents guideに使うレイヤー番号 (1~30、20はconv4_2)') + parser.add_argument("--guide_image_path", type=str, default=None, help="image to CLIP guidance / CLIP guided SDでガイドに使う画像") + parser.add_argument("--highres_fix_scale", type=float, default=None, + help="enable highres fix, reso scale for 1st stage / highres fixを有効にして最初の解像度をこのscaleにする") + parser.add_argument("--highres_fix_steps", type=int, default=28, + help="1st stage steps for highres fix / highres fixの最初のステージのステップ数") + parser.add_argument("--highres_fix_save_1st", action='store_true', + help="save 1st stage images for highres fix / highres fixの最初のステージの画像を保存する") + parser.add_argument("--negative_scale", type=float, default=None, + help="set another guidance scale for negative prompt / ネガティブプロンプトのscaleを指定する") + + args = parser.parse_args() + main(args) diff --git a/networks/control_net_lora.py b/networks/control_net_lora.py new file mode 100644 index 00000000..58771922 --- /dev/null +++ b/networks/control_net_lora.py @@ -0,0 +1,252 @@ +# LoRA network module +# reference: +# https://github.com/microsoft/LoRA/blob/main/loralib/layers.py +# https://github.com/cloneofsimo/lora/blob/master/lora_diffusion/lora.py + +import math +import os +from typing import List +import torch + +from library import train_util + + +class LoRAModule(torch.nn.Module): + """ + replaces forward method of the original Linear, instead of replacing the original Linear module. + """ + + def __init__(self, lora_name, org_module: torch.nn.Module, multiplier=1.0, lora_dim=4, alpha=1): + """ if alpha == 0 or None, alpha is rank (no scaling). """ + super().__init__() + self.lora_name = lora_name + self.lora_dim = lora_dim + + if org_module.__class__.__name__ == 'Conv2d': + in_dim = org_module.in_channels + out_dim = org_module.out_channels + self.lora_down = torch.nn.Conv2d(in_dim, lora_dim, (1, 1), bias=False) + self.lora_up = torch.nn.Conv2d(lora_dim, out_dim, (1, 1), bias=False) + else: + in_dim = org_module.in_features + out_dim = org_module.out_features + self.lora_down = torch.nn.Linear(in_dim, lora_dim, bias=False) + self.lora_up = torch.nn.Linear(lora_dim, out_dim, bias=False) + + if type(alpha) == torch.Tensor: + alpha = alpha.detach().float().numpy() # without casting, bf16 causes error + alpha = lora_dim if alpha is None or alpha == 0 else alpha + self.scale = alpha / self.lora_dim + self.register_buffer('alpha', torch.tensor(alpha)) # 定数として扱える + + # same as microsoft's + torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5)) + torch.nn.init.zeros_(self.lora_up.weight) + + self.multiplier = multiplier + self.org_module = org_module # remove in applying + + def apply_to(self): + self.org_forward = self.org_module.forward + self.org_module.forward = self.forward + del self.org_module + + def set_as_control_path(self, control_path): + self.is_control_path = control_path + + def forward(self, x): + if self.is_control_path: + lora_x = self.lora_up(self.lora_down(x)) * self.multiplier * self.scale + self.previous_lora_x = lora_x + else: + lora_x = self.previous_lora_x + del self.previous_lora_x + return self.org_forward(x) + lora_x + + +def create_network(multiplier, network_dim, network_alpha, vae, text_encoder, unet, **kwargs): + if network_dim is None: + network_dim = 4 # default + network = LoRANetwork(text_encoder, unet, multiplier=multiplier, lora_dim=network_dim, alpha=network_alpha) + return network + + +def create_network_from_weights(multiplier, file, vae, text_encoder, unet, **kwargs): + if os.path.splitext(file)[1] == '.safetensors': + from safetensors.torch import load_file, safe_open + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location='cpu') + + # get dim (rank) + network_alpha = None + network_dim = None + for key, value in weights_sd.items(): + if network_alpha is None and 'alpha' in key: + network_alpha = value + if network_dim is None and 'lora_down' in key and len(value.size()) == 2: + network_dim = value.size()[0] + + if network_alpha is None: + network_alpha = network_dim + + network = LoRANetwork(text_encoder, unet, multiplier=multiplier, lora_dim=network_dim, alpha=network_alpha) + network.weights_sd = weights_sd + return network + + +class LoRANetwork(torch.nn.Module): + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel", "Attention"] + TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPMLP"] + LORA_PREFIX_UNET = 'lora_unet' + LORA_PREFIX_TEXT_ENCODER = 'lora_te' + + def __init__(self, text_encoder, unet, multiplier=1.0, lora_dim=4, alpha=1) -> None: + super().__init__() + self.multiplier = multiplier + self.lora_dim = lora_dim + self.alpha = alpha + + # create module instances + def create_modules(prefix, root_module: torch.nn.Module, target_replace_modules) -> List[LoRAModule]: + loras = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + if child_module.__class__.__name__ == "Linear" or (child_module.__class__.__name__ == "Conv2d" and child_module.kernel_size == (1, 1)): + lora_name = prefix + '.' + name + '.' + child_name + lora_name = lora_name.replace('.', '_') + lora = LoRAModule(lora_name, child_module, self.multiplier, self.lora_dim, self.alpha) + loras.append(lora) + return loras + + self.text_encoder_loras = create_modules(LoRANetwork.LORA_PREFIX_TEXT_ENCODER, + text_encoder, LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE) + print(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.") + + self.unet_loras = create_modules(LoRANetwork.LORA_PREFIX_UNET, unet, LoRANetwork.UNET_TARGET_REPLACE_MODULE) + print(f"create LoRA for U-Net: {len(self.unet_loras)} modules.") + + self.weights_sd = None + + # assertion + names = set() + for lora in self.text_encoder_loras + self.unet_loras: + assert lora.lora_name not in names, f"duplicated lora name: {lora.lora_name}" + names.add(lora.lora_name) + + def load_weights(self, file): + if os.path.splitext(file)[1] == '.safetensors': + from safetensors.torch import load_file, safe_open + self.weights_sd = load_file(file) + else: + self.weights_sd = torch.load(file, map_location='cpu') + + def apply_to(self, text_encoder, unet, apply_text_encoder=None, apply_unet=None): + if self.weights_sd: + weights_has_text_encoder = weights_has_unet = False + for key in self.weights_sd.keys(): + if key.startswith(LoRANetwork.LORA_PREFIX_TEXT_ENCODER): + weights_has_text_encoder = True + elif key.startswith(LoRANetwork.LORA_PREFIX_UNET): + weights_has_unet = True + + if apply_text_encoder is None: + apply_text_encoder = weights_has_text_encoder + else: + assert apply_text_encoder == weights_has_text_encoder, f"text encoder weights: {weights_has_text_encoder} but text encoder flag: {apply_text_encoder} / 重みとText Encoderのフラグが矛盾しています" + + if apply_unet is None: + apply_unet = weights_has_unet + else: + assert apply_unet == weights_has_unet, f"u-net weights: {weights_has_unet} but u-net flag: {apply_unet} / 重みとU-Netのフラグが矛盾しています" + else: + assert apply_text_encoder is not None and apply_unet is not None, f"internal error: flag not set" + + assert not apply_text_encoder, "ControlNet does not support for text encoder" + + if apply_text_encoder: + print("enable LoRA for text encoder") + else: + self.text_encoder_loras = [] + + if apply_unet: + print("enable LoRA for U-Net") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + lora.apply_to() + self.add_module(lora.lora_name, lora) + + if self.weights_sd: + # if some weights are not in state dict, it is ok because initial LoRA does nothing (lora_up is initialized by zeros) + info = self.load_state_dict(self.weights_sd, False) + print(f"weights are loaded: {info}") + + def set_as_control_path(self, control_path): + for lora in self.text_encoder_loras + self.unet_loras: + lora.set_as_control_path(control_path) + + def enable_gradient_checkpointing(self): + # not supported + pass + + def prepare_optimizer_params(self, text_encoder_lr, unet_lr): + def enumerate_params(loras): + params = [] + for lora in loras: + params.extend(lora.parameters()) + return params + + self.requires_grad_(True) + all_params = [] + + if self.text_encoder_loras: + param_data = {'params': enumerate_params(self.text_encoder_loras)} + if text_encoder_lr is not None: + param_data['lr'] = text_encoder_lr + all_params.append(param_data) + + if self.unet_loras: + param_data = {'params': enumerate_params(self.unet_loras)} + if unet_lr is not None: + param_data['lr'] = unet_lr + all_params.append(param_data) + + return all_params + + def prepare_grad_etc(self, text_encoder, unet): + self.requires_grad_(True) + + def on_epoch_start(self, text_encoder, unet): + self.train() + + def get_trainable_params(self): + return self.parameters() + + def save_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + state_dict = self.state_dict() + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == '.safetensors': + from safetensors.torch import save_file + + # Precalculate model hashes to save time on indexing + if metadata is None: + metadata = {} + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) diff --git a/train_control_net.py b/train_control_net.py new file mode 100644 index 00000000..74d384bf --- /dev/null +++ b/train_control_net.py @@ -0,0 +1,756 @@ +from diffusers.optimization import SchedulerType, TYPE_TO_SCHEDULER_FUNCTION +from torch.optim import Optimizer +from typing import Optional, Tuple, Union +import importlib +import argparse +import gc +import math +import os +import random +import time +import json + +from tqdm import tqdm +import torch +from accelerate.utils import set_seed +import diffusers +from diffusers import DDPMScheduler + +import library.train_util as train_util +from library.train_util import BaseDataset, ImageInfo, glob_images +import networks.control_net_lora as control_net_rola + + +def collate_fn(examples): + return examples[0] + + +def generate_step_logs(args: argparse.Namespace, current_loss, avr_loss, lr_scheduler): + logs = {"loss/current": current_loss, "loss/average": avr_loss} + + if args.network_train_unet_only: + logs["lr/unet"] = lr_scheduler.get_last_lr()[0] + elif args.network_train_text_encoder_only: + logs["lr/textencoder"] = lr_scheduler.get_last_lr()[0] + else: + logs["lr/textencoder"] = lr_scheduler.get_last_lr()[0] + logs["lr/unet"] = lr_scheduler.get_last_lr()[-1] # may be same to textencoder + + return logs + + +# Monkeypatch newer get_scheduler() function overridng current version of diffusers.optimizer.get_scheduler +# code is taken from https://github.com/huggingface/diffusers diffusers.optimizer, commit d87cc15977b87160c30abaace3894e802ad9e1e6 +# Which is a newer release of diffusers than currently packaged with sd-scripts +# This code can be removed when newer diffusers version (v0.12.1 or greater) is tested and implemented to sd-scripts + + +def get_scheduler_fix( + name: Union[str, SchedulerType], + optimizer: Optimizer, + num_warmup_steps: Optional[int] = None, + num_training_steps: Optional[int] = None, + num_cycles: int = 1, + power: float = 1.0, +): + """ + Unified API to get any scheduler from its name. + Args: + name (`str` or `SchedulerType`): + The name of the scheduler to use. + optimizer (`torch.optim.Optimizer`): + The optimizer that will be used during training. + num_warmup_steps (`int`, *optional*): + The number of warmup steps to do. This is not required by all schedulers (hence the argument being + optional), the function will raise an error if it's unset and the scheduler type requires it. + num_training_steps (`int``, *optional*): + The number of training steps to do. This is not required by all schedulers (hence the argument being + optional), the function will raise an error if it's unset and the scheduler type requires it. + num_cycles (`int`, *optional*): + The number of hard restarts used in `COSINE_WITH_RESTARTS` scheduler. + power (`float`, *optional*, defaults to 1.0): + Power factor. See `POLYNOMIAL` scheduler + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + """ + name = SchedulerType(name) + schedule_func = TYPE_TO_SCHEDULER_FUNCTION[name] + if name == SchedulerType.CONSTANT: + return schedule_func(optimizer) + + # All other schedulers require `num_warmup_steps` + if num_warmup_steps is None: + raise ValueError(f"{name} requires `num_warmup_steps`, please provide that argument.") + + if name == SchedulerType.CONSTANT_WITH_WARMUP: + return schedule_func(optimizer, num_warmup_steps=num_warmup_steps) + + # All other schedulers require `num_training_steps` + if num_training_steps is None: + raise ValueError(f"{name} requires `num_training_steps`, please provide that argument.") + + if name == SchedulerType.COSINE_WITH_RESTARTS: + return schedule_func( + optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps, num_cycles=num_cycles + ) + + if name == SchedulerType.POLYNOMIAL: + return schedule_func( + optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps, power=power + ) + + return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps) + + +class ImagesWithHintDataset(BaseDataset): + def __init__(self, batch_size, train_data_dir, tokenizer, max_token_length, caption_extension, shuffle_caption, shuffle_keep_tokens, resolution, enable_bucket, min_bucket_reso, max_bucket_reso, bucket_reso_steps, bucket_no_upscale, flip_aug, color_aug, face_crop_aug_range, random_crop, debug_dataset) -> None: + super().__init__(tokenizer, max_token_length, shuffle_caption, shuffle_keep_tokens, + resolution, flip_aug, color_aug, face_crop_aug_range, random_crop, debug_dataset) + + assert resolution is not None, f"resolution is required / resolution（解像度）指定は必須です" + + self.batch_size = batch_size + self.size = min(self.width, self.height) # 短いほう + self.latents_cache = None + + self.enable_bucket = enable_bucket + if self.enable_bucket: + assert min(resolution) >= min_bucket_reso, f"min_bucket_reso must be equal or less than resolution / min_bucket_resoは最小解像度より大きくできません。解像度を大きくするかmin_bucket_resoを小さくしてください" + assert max(resolution) <= max_bucket_reso, f"max_bucket_reso must be equal or greater than resolution / max_bucket_resoは最大解像度より小さくできません。解像度を小さくするかmin_bucket_resoを大きくしてください" + self.min_bucket_reso = min_bucket_reso + self.max_bucket_reso = max_bucket_reso + self.bucket_reso_steps = bucket_reso_steps + self.bucket_no_upscale = bucket_no_upscale + else: + self.min_bucket_reso = None + self.max_bucket_reso = None + self.bucket_reso_steps = None # この情報は使われない + self.bucket_no_upscale = False + + # fill50k + print("loading fill50k dataset") + with open(os.path.join(train_data_dir, "prompt.json")) as f: + annos = f.readlines() + + captions = [] + src_paths = [] + trg_paths = [] + for anno in annos: + anno1 = json.loads(anno) + captions.append(anno1["prompt"]) + src_paths.append(os.path.join(train_data_dir, anno1["source"])) + trg_paths.append(os.path.join(train_data_dir, anno1["target"])) + + self.set_tag_frequency(os.path.basename(train_data_dir), captions) # タグ頻度を記録 + self.dataset_dirs_info[os.path.basename(train_data_dir)] = {"n_repeats": 1, "img_count": len(src_paths)} + + for src_path, trg_path, caption in zip(src_paths, trg_paths, captions): + info = ImageInfo(src_path, 1, caption, False, src_path) + self.register_image(info) + + num_train_images = len(src_paths) + print(f"{num_train_images} train images with repeating.") + self.num_train_images = num_train_images + self.num_reg_images = 0 + + """ + def read_caption(img_path): + # captionの候補ファイル名を作る + base_name = os.path.splitext(img_path)[0] + base_name_face_det = base_name + tokens = base_name.split("_") + if len(tokens) >= 5: + base_name_face_det = "_".join(tokens[:-4]) + cap_paths = [base_name + caption_extension, base_name_face_det + caption_extension] + + caption = None + for cap_path in cap_paths: + if os.path.isfile(cap_path): + with open(cap_path, "rt", encoding='utf-8') as f: + try: + lines = f.readlines() + except UnicodeDecodeError as e: + print(f"illegal char in file (not UTF-8) / ファイルにUTF-8以外の文字があります: {cap_path}") + raise e + assert len(lines) > 0, f"caption file is empty / キャプションファイルが空です: {cap_path}" + caption = lines[0].strip() + break + return caption + + def load_dreambooth_dir(dir): + if not os.path.isdir(dir): + # print(f"ignore file: {dir}") + return 0, [], [] + + tokens = os.path.basename(dir).split('_') + try: + n_repeats = int(tokens[0]) + except ValueError as e: + print(f"ignore directory without repeats / 繰り返し回数のないディレクトリを無視します: {dir}") + return 0, [], [] + + caption_by_folder = '_'.join(tokens[1:]) + img_paths = glob_images(dir, "*") + print(f"found directory {n_repeats}_{caption_by_folder} contains {len(img_paths)} image files") + + # 画像ファイルごとにプロンプトを読み込み、もしあればそちらを使う + captions = [] + for img_path in img_paths: + cap_for_img = read_caption(img_path) + captions.append(caption_by_folder if cap_for_img is None else cap_for_img) + + self.set_tag_frequency(os.path.basename(dir), captions) # タグ頻度を記録 + + return n_repeats, img_paths, captions + + print("prepare train images.") + train_dirs = os.listdir(train_data_dir) + num_train_images = 0 + for dir in train_dirs: + n_repeats, img_paths, captions = load_dreambooth_dir(os.path.join(train_data_dir, dir)) + num_train_images += n_repeats * len(img_paths) + + for img_path, caption in zip(img_paths, captions): + info = ImageInfo(img_path, n_repeats, caption, False, img_path) + self.register_image(info) + + self.dataset_dirs_info[os.path.basename(dir)] = {"n_repeats": n_repeats, "img_count": len(img_paths)} + + print(f"{num_train_images} train images with repeating.") + self.num_train_images = num_train_images + self.num_reg_images = 0 + """ + + def __getitem__(self, index): + # latentsのcacheをサポートしてない + if index == 0: + self.shuffle_buckets() + + bucket = self.bucket_manager.buckets[self.buckets_indices[index].bucket_index] + bucket_batch_size = self.buckets_indices[index].bucket_batch_size + image_index = self.buckets_indices[index].batch_index * bucket_batch_size + + loss_weights = [] + captions = [] + input_ids_list = [] + images = [] + hint_images = [] + + for image_key in bucket[image_index:image_index + bucket_batch_size]: + image_info = self.image_data[image_key] + loss_weights.append(1.0) + + # image/latentsを処理する + # 画像を読み込み、必要ならcropする + src_path = image_info.absolute_path + trg_path = src_path.replace("source", "target") + img, face_cx, face_cy, face_w, face_h = self.load_image_with_face_info(trg_path) + hint_img, face_cx, face_cy, face_w, face_h = self.load_image_with_face_info(src_path) + assert img.shape[0:2] == hint_img.shape[0:2] + im_h, im_w = img.shape[0:2] + + if self.enable_bucket: + img = self.trim_and_resize_if_required(img, image_info.bucket_reso, image_info.resized_size) + else: + if face_cx > 0: # 顔位置情報あり + img = self.crop_target(img, face_cx, face_cy, face_w, face_h) + elif im_h > self.height or im_w > self.width: + assert self.random_crop, f"image too large, but cropping and bucketing are disabled / 画像サイズが大きいのでface_crop_aug_rangeかrandom_crop、またはbucketを有効にしてください: {image_info.absolute_path}" + if im_h > self.height: + p = random.randint(0, im_h - self.height) + img = img[p:p + self.height] + if im_w > self.width: + p = random.randint(0, im_w - self.width) + img = img[:, p:p + self.width] + + im_h, im_w = img.shape[0:2] + assert im_h == self.height and im_w == self.width, f"image size is small / 画像サイズが小さいようです: {image_info.absolute_path}" + + # augmentation + if self.aug is not None: + # TODO color aug does not work + auged = self.aug(image=img, image2=hint_img) + img = auged['image'] + hint_img = auged['image2'] + + image = self.image_transforms(img) # -1.0~1.0のtorch.Tensorになる + hint_image = self.image_transforms(hint_img) # -1.0~1.0のtorch.Tensorになる + + images.append(image) + hint_images.append(hint_image) + + caption = self.process_caption(image_info.caption) + captions.append(caption) + if not self.token_padding_disabled: # this option might be omitted in future + input_ids_list.append(self.get_input_ids(caption)) + + example = {} + example['loss_weights'] = torch.FloatTensor(loss_weights) + + if self.token_padding_disabled: + # padding=True means pad in the batch + example['input_ids'] = self.tokenizer(captions, padding=True, truncation=True, return_tensors="pt").input_ids + else: + # batch processing seems to be good + example['input_ids'] = torch.stack(input_ids_list) + + images = torch.stack(images) + images = images.to(memory_format=torch.contiguous_format).float() + example['images'] = images + + hint_images = torch.stack(hint_images) + hint_images = hint_images.to(memory_format=torch.contiguous_format).float() + example['hint_images'] = hint_images + + example['latents'] = None + + if self.debug_dataset: + example['image_keys'] = bucket[image_index:image_index + self.batch_size] + example['captions'] = captions + return example + + +def train(args): + session_id = random.randint(0, 2**32) + training_started_at = time.time() + train_util.verify_training_args(args) + train_util.prepare_dataset_args(args, True) + + cache_latents = args.cache_latents + use_dreambooth_method = args.in_json is None + + if args.seed is not None: + set_seed(args.seed) + + tokenizer = train_util.load_tokenizer(args) + + # データセットを準備する + train_dataset = ImagesWithHintDataset(args.train_batch_size, args.train_data_dir, + tokenizer, args.max_token_length, args.caption_extension, args.shuffle_caption, args.keep_tokens, + args.resolution, args.enable_bucket, args.min_bucket_reso, args.max_bucket_reso, + args.bucket_reso_steps, args.bucket_no_upscale, + args.flip_aug, args.color_aug, args.face_crop_aug_range, + args.random_crop, args.debug_dataset) + + # 学習データのdropout率を設定する + train_dataset.set_caption_dropout(args.caption_dropout_rate, args.caption_dropout_every_n_epochs, args.caption_tag_dropout_rate) + + train_dataset.make_buckets() + + if args.debug_dataset: + train_util.debug_dataset(train_dataset) + return + if len(train_dataset) == 0: + print("No data found. Please verify arguments (train_data_dir must be the parent of folders with images) / 画像がありません。引数指定を確認してください（train_data_dirには画像があるフォルダではなく、画像があるフォルダの親フォルダを指定する必要があります）") + return + + # acceleratorを準備する + print("prepare accelerator") + accelerator, unwrap_model = train_util.prepare_accelerator(args) + + # mixed precisionに対応した型を用意しておき適宜castする + weight_dtype, save_dtype = train_util.prepare_dtype(args) + + # モデルを読み込む + text_encoder, vae, unet, _ = train_util.load_target_model(args, weight_dtype) + + # モデルに xformers とか memory efficient attention を組み込む + train_util.replace_unet_modules(unet, args.mem_eff_attn, args.xformers) + + # 学習を準備する + if cache_latents: + vae.to(accelerator.device, dtype=weight_dtype) + vae.requires_grad_(False) + vae.eval() + with torch.no_grad(): + train_dataset.cache_latents(vae) + vae.to("cpu") + if torch.cuda.is_available(): + torch.cuda.empty_cache() + gc.collect() + + # prepare network + print("import network module:", args.network_module) + network_module = importlib.import_module(args.network_module) + + net_kwargs = {} + if args.network_args is not None: + for net_arg in args.network_args: + key, value = net_arg.split('=') + net_kwargs[key] = value + + # if a new network is added in future, add if ~ then blocks for each network (;'∀') + network: control_net_rola.LoRANetwork = network_module.create_network( + 1.0, args.network_dim, args.network_alpha, vae, text_encoder, unet, **net_kwargs) + if network is None: + return + + if args.network_weights is not None: + print("load network weights from:", args.network_weights) + network.load_weights(args.network_weights) + + train_unet = not args.network_train_text_encoder_only + train_text_encoder = not args.network_train_unet_only + network.apply_to(text_encoder, unet, train_text_encoder, train_unet) + + if args.gradient_checkpointing: + unet.enable_gradient_checkpointing() + text_encoder.gradient_checkpointing_enable() + network.enable_gradient_checkpointing() # may have no effect + + # 学習に必要なクラスを準備する + print("prepare optimizer, data loader etc.") + + # 8-bit Adamを使う + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError("No bitsand bytes / bitsandbytesがインストールされていないようです") + print("use 8-bit Adam optimizer") + optimizer_class = bnb.optim.AdamW8bit + else: + optimizer_class = torch.optim.AdamW + + trainable_params = network.prepare_optimizer_params(args.text_encoder_lr, args.unet_lr) + + # betaやweight decayはdiffusers DreamBoothもDreamBooth SDもデフォルト値のようなのでオプションはとりあえず省略 + optimizer = optimizer_class(trainable_params, lr=args.learning_rate) + + # dataloaderを準備する + # DataLoaderのプロセス数：0はメインプロセスになる + n_workers = min(args.max_data_loader_n_workers, os.cpu_count() - 1) # cpu_count-1 ただし最大で指定された数まで + train_dataloader = torch.utils.data.DataLoader( + train_dataset, batch_size=1, shuffle=False, collate_fn=collate_fn, num_workers=n_workers, persistent_workers=args.persistent_data_loader_workers) + + # 学習ステップ数を計算する + if args.max_train_epochs is not None: + args.max_train_steps = args.max_train_epochs * len(train_dataloader) + print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}") + + # lr schedulerを用意する + # lr_scheduler = diffusers.optimization.get_scheduler( + lr_scheduler = get_scheduler_fix( + args.lr_scheduler, optimizer, num_warmup_steps=args.lr_warmup_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + num_cycles=args.lr_scheduler_num_cycles, power=args.lr_scheduler_power) + + # 実験的機能：勾配も含めたfp16学習を行う　モデル全体をfp16にする + if args.full_fp16: + assert args.mixed_precision == "fp16", "full_fp16 requires mixed precision='fp16' / full_fp16を使う場合はmixed_precision='fp16'を指定してください。" + print("enable full fp16 training.") + network.to(weight_dtype) + + # acceleratorがなんかよろしくやってくれるらしい + if train_unet and train_text_encoder: + unet, text_encoder, network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, text_encoder, network, optimizer, train_dataloader, lr_scheduler) + elif train_unet: + unet, network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, network, optimizer, train_dataloader, lr_scheduler) + elif train_text_encoder: + text_encoder, network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + text_encoder, network, optimizer, train_dataloader, lr_scheduler) + else: + network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + network, optimizer, train_dataloader, lr_scheduler) + + unet.requires_grad_(False) + unet.to(accelerator.device, dtype=weight_dtype) + text_encoder.requires_grad_(False) + text_encoder.to(accelerator.device, dtype=weight_dtype) + if args.gradient_checkpointing: # according to TI example in Diffusers, train is required + unet.train() + text_encoder.train() + + # set top parameter requires_grad = True for gradient checkpointing works + text_encoder.text_model.embeddings.requires_grad_(True) + else: + unet.eval() + text_encoder.eval() + + network.prepare_grad_etc(text_encoder, unet) + + if not cache_latents: + vae.requires_grad_(False) + vae.eval() + vae.to(accelerator.device, dtype=weight_dtype) + + # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする + if args.full_fp16: + train_util.patch_accelerator_for_fp16_training(accelerator) + + # resumeする + if args.resume is not None: + print(f"resume training from state: {args.resume}") + accelerator.load_state(args.resume) + + # epoch数を計算する + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + if (args.save_n_epoch_ratio is not None) and (args.save_n_epoch_ratio > 0): + args.save_every_n_epochs = math.floor(num_train_epochs / args.save_n_epoch_ratio) or 1 + + # 学習する + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + print("running training / 学習開始") + print(f" num train images * repeats / 学習画像の数×繰り返し回数: {train_dataset.num_train_images}") + print(f" num reg images / 正則化画像の数: {train_dataset.num_reg_images}") + print(f" num batches per epoch / 1epochのバッチ数: {len(train_dataloader)}") + print(f" num epochs / epoch数: {num_train_epochs}") + print(f" batch size per device / バッチサイズ: {args.train_batch_size}") + print(f" total train batch size (with parallel & distributed & accumulation) / 総バッチサイズ（並列学習、勾配合計含む）: {total_batch_size}") + print(f" gradient accumulation steps / 勾配を合計するステップ数 = {args.gradient_accumulation_steps}") + print(f" total optimization steps / 学習ステップ数: {args.max_train_steps}") + + metadata = { + "ss_session_id": session_id, # random integer indicating which group of epochs the model came from + "ss_training_started_at": training_started_at, # unix timestamp + "ss_output_name": args.output_name, + "ss_learning_rate": args.learning_rate, + "ss_text_encoder_lr": args.text_encoder_lr, + "ss_unet_lr": args.unet_lr, + "ss_num_train_images": train_dataset.num_train_images, # includes repeating + "ss_num_reg_images": train_dataset.num_reg_images, + "ss_num_batches_per_epoch": len(train_dataloader), + "ss_num_epochs": num_train_epochs, + "ss_batch_size_per_device": args.train_batch_size, + "ss_total_batch_size": total_batch_size, + "ss_gradient_checkpointing": args.gradient_checkpointing, + "ss_gradient_accumulation_steps": args.gradient_accumulation_steps, + "ss_max_train_steps": args.max_train_steps, + "ss_lr_warmup_steps": args.lr_warmup_steps, + "ss_lr_scheduler": args.lr_scheduler, + "ss_network_module": "control_net_" + args.network_module, + "ss_network_dim": args.network_dim, # None means default because another network than LoRA may have another default dim + "ss_network_alpha": args.network_alpha, # some networks may not use this value + "ss_mixed_precision": args.mixed_precision, + "ss_full_fp16": bool(args.full_fp16), + "ss_v2": bool(args.v2), + "ss_resolution": args.resolution, + "ss_clip_skip": args.clip_skip, + "ss_max_token_length": args.max_token_length, + "ss_color_aug": bool(args.color_aug), + "ss_flip_aug": bool(args.flip_aug), + "ss_random_crop": bool(args.random_crop), + "ss_shuffle_caption": bool(args.shuffle_caption), + "ss_cache_latents": bool(args.cache_latents), + "ss_enable_bucket": bool(train_dataset.enable_bucket), + "ss_min_bucket_reso": train_dataset.min_bucket_reso, + "ss_max_bucket_reso": train_dataset.max_bucket_reso, + "ss_seed": args.seed, + "ss_keep_tokens": args.keep_tokens, + "ss_dataset_dirs": json.dumps(train_dataset.dataset_dirs_info), + "ss_reg_dataset_dirs": json.dumps(train_dataset.reg_dataset_dirs_info), + "ss_tag_frequency": json.dumps(train_dataset.tag_frequency), + "ss_bucket_info": json.dumps(train_dataset.bucket_info), + "ss_training_comment": args.training_comment # will not be updated after training + } + + # uncomment if another network is added + # for key, value in net_kwargs.items(): + # metadata["ss_arg_" + key] = value + + if args.pretrained_model_name_or_path is not None: + sd_model_name = args.pretrained_model_name_or_path + if os.path.exists(sd_model_name): + metadata["ss_sd_model_hash"] = train_util.model_hash(sd_model_name) + metadata["ss_new_sd_model_hash"] = train_util.calculate_sha256(sd_model_name) + sd_model_name = os.path.basename(sd_model_name) + metadata["ss_sd_model_name"] = sd_model_name + + if args.vae is not None: + vae_name = args.vae + if os.path.exists(vae_name): + metadata["ss_vae_hash"] = train_util.model_hash(vae_name) + metadata["ss_new_vae_hash"] = train_util.calculate_sha256(vae_name) + vae_name = os.path.basename(vae_name) + metadata["ss_vae_name"] = vae_name + + metadata = {k: str(v) for k, v in metadata.items()} + + progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps") + global_step = 0 + + noise_scheduler = DDPMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", + num_train_timesteps=1000, clip_sample=False) + + if accelerator.is_main_process: + accelerator.init_trackers("network_train") + + for epoch in range(num_train_epochs): + print(f"epoch {epoch+1}/{num_train_epochs}") + train_dataset.set_current_epoch(epoch + 1) + + metadata["ss_epoch"] = str(epoch+1) + + network.on_epoch_start(text_encoder, unet) + + loss_total = 0 + for step, batch in enumerate(train_dataloader): + with accelerator.accumulate(network): + with torch.no_grad(): + if "latents" in batch and batch["latents"] is not None: + latents = batch["latents"].to(accelerator.device) + else: + # latentに変換 + latents = vae.encode(batch["images"].to(dtype=weight_dtype)).latent_dist.sample() + hint_latents = vae.encode(batch["hint_images"].to(dtype=weight_dtype)).latent_dist.sample() + latents = latents * 0.18215 + hint_latents = hint_latents * 0.18215 + # hint = torch.nn.functional.interpolate(batch["hint_images"], scale_factor=(1/8, 1/8), mode="bilinear") + # hint = hint[:, 0].unsqueeze(1) # RGB -> BW + b_size = latents.shape[0] + + with torch.set_grad_enabled(train_text_encoder): + # Get the text embedding for conditioning + input_ids = batch["input_ids"].to(accelerator.device) + encoder_hidden_states = train_util.get_hidden_states(args, input_ids, tokenizer, text_encoder, weight_dtype) + + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents, device=latents.device) + + # Sample a random timestep for each image + timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (b_size,), device=latents.device) + timesteps = timesteps.long() + + # Add noise to the latents according to the noise magnitude at each timestep + # (this is the forward diffusion process) + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + # Predict the noise residual + network.set_as_control_path(True) + unet(hint_latents, timesteps, encoder_hidden_states) # めちゃくちゃ乱暴だが入力にhintを加える + # unet(noisy_latents * hint, timesteps, encoder_hidden_states) # めちゃくちゃ乱暴だが入力にhintを乗算 + network.set_as_control_path(False) + noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample + + if args.v_parameterization: + # v-parameterization training + target = noise_scheduler.get_velocity(latents, noise, timesteps) + else: + target = noise + + loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none") + loss = loss.mean([1, 2, 3]) + + loss_weights = batch["loss_weights"] # 各sampleごとのweight + loss = loss * loss_weights + + loss = loss.mean() # 平均なのでbatch_sizeで割る必要なし + + accelerator.backward(loss) + if accelerator.sync_gradients: + params_to_clip = network.get_trainable_params() + accelerator.clip_grad_norm_(params_to_clip, 1.0) # args.max_grad_norm) + + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad(set_to_none=True) + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + progress_bar.update(1) + global_step += 1 + + current_loss = loss.detach().item() + loss_total += current_loss + avr_loss = loss_total / (step+1) + logs = {"loss": avr_loss} # , "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if args.logging_dir is not None: + logs = generate_step_logs(args, current_loss, avr_loss, lr_scheduler) + accelerator.log(logs, step=global_step) + + if global_step >= args.max_train_steps: + break + + if args.logging_dir is not None: + logs = {"loss/epoch": loss_total / len(train_dataloader)} + accelerator.log(logs, step=epoch+1) + + accelerator.wait_for_everyone() + + if args.save_every_n_epochs is not None: + model_name = train_util.DEFAULT_EPOCH_NAME if args.output_name is None else args.output_name + + def save_func(): + ckpt_name = train_util.EPOCH_FILE_NAME.format(model_name, epoch + 1) + '.' + args.save_model_as + ckpt_file = os.path.join(args.output_dir, ckpt_name) + print(f"saving checkpoint: {ckpt_file}") + unwrap_model(network).save_weights(ckpt_file, save_dtype, None if args.no_metadata else metadata) + + def remove_old_func(old_epoch_no): + old_ckpt_name = train_util.EPOCH_FILE_NAME.format(model_name, old_epoch_no) + '.' + args.save_model_as + old_ckpt_file = os.path.join(args.output_dir, old_ckpt_name) + if os.path.exists(old_ckpt_file): + print(f"removing old checkpoint: {old_ckpt_file}") + os.remove(old_ckpt_file) + + saving = train_util.save_on_epoch_end(args, save_func, remove_old_func, epoch + 1, num_train_epochs) + if saving and args.save_state: + train_util.save_state_on_epoch_end(args, accelerator, model_name, epoch + 1) + + # end of epoch + + metadata["ss_epoch"] = str(num_train_epochs) + + is_main_process = accelerator.is_main_process + if is_main_process: + network = unwrap_model(network) + + accelerator.end_training() + + if args.save_state: + train_util.save_state_on_train_end(args, accelerator) + + del accelerator # この後メモリを使うのでこれは消す + + if is_main_process: + os.makedirs(args.output_dir, exist_ok=True) + + model_name = train_util.DEFAULT_LAST_OUTPUT_NAME if args.output_name is None else args.output_name + ckpt_name = model_name + '.' + args.save_model_as + ckpt_file = os.path.join(args.output_dir, ckpt_name) + + print(f"save trained model to {ckpt_file}") + network.save_weights(ckpt_file, save_dtype, None if args.no_metadata else metadata) + print("model saved.") + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + + train_util.add_sd_models_arguments(parser) + train_util.add_dataset_arguments(parser, True, True, True) + train_util.add_training_arguments(parser, True) + + parser.add_argument("--no_metadata", action='store_true', help="do not save metadata in output model / メタデータを出力先モデルに保存しない") + parser.add_argument("--save_model_as", type=str, default="safetensors", choices=[None, "ckpt", "pt", "safetensors"], + help="format to save the model (default is .safetensors) / モデル保存時の形式（デフォルトはsafetensors）") + + parser.add_argument("--unet_lr", type=float, default=None, help="learning rate for U-Net / U-Netの学習率") + parser.add_argument("--text_encoder_lr", type=float, default=None, help="learning rate for Text Encoder / Text Encoderの学習率") + parser.add_argument("--lr_scheduler_num_cycles", type=int, default=1, + help="Number of restarts for cosine scheduler with restarts / cosine with restartsスケジューラでのリスタート回数") + parser.add_argument("--lr_scheduler_power", type=float, default=1, + help="Polynomial power for polynomial scheduler / polynomialスケジューラでのpolynomial power") + + parser.add_argument("--network_weights", type=str, default=None, + help="pretrained weights for network / 学習するネットワークの初期重み") + parser.add_argument("--network_module", type=str, default=None, help='network module to train / 学習対象のネットワークのモジュール') + parser.add_argument("--network_dim", type=int, default=None, + help='network dimensions (depends on each network) / モジュールの次元数（ネットワークにより定義は異なります）') + parser.add_argument("--network_alpha", type=float, default=1, + help='alpha for LoRA weight scaling, default 1 (same as network_dim for same behavior as old version) / LoRaの重み調整のalpha値、デフォルト1（旧バージョンと同じ動作をするにはnetwork_dimと同じ値を指定）') + parser.add_argument("--network_args", type=str, default=None, nargs='*', + help='additional argmuments for network (key=value) / ネットワークへの追加の引数') + parser.add_argument("--network_train_unet_only", action="store_true", help="only training U-Net part / U-Net関連部分のみ学習する") + parser.add_argument("--network_train_text_encoder_only", action="store_true", + help="only training Text Encoder part / Text Encoder関連部分のみ学習する") + parser.add_argument("--training_comment", type=str, default=None, + help="arbitrary comment string stored in metadata / メタデータに記録する任意のコメント文字列") + + args = parser.parse_args() + train(args)