VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-08 22:35:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into original-u-net

Browse Source
This commit is contained in:
ykume
2023-06-11 17:35:20 +09:00
parent 4b7b3bc04a 6417f5d7c1
commit 33a6234b52
3 changed files with 17 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
README.md
View File

@@ -140,6 +140,11 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
## Change History ## Change History
### 8 Jun. 2023, 2023/06/08
- Fixed a bug where clip skip did not work when training with weighted captions (`--weighted_captions` specified) and when generating sample images during training.
- 重みづけキャプションでの学習時（`--weighted_captions`指定時）および学習中のサンプル画像生成時にclip skipが機能しない不具合を修正しました。
### 6 Jun. 2023, 2023/06/06 ### 6 Jun. 2023, 2023/06/06
- Fix `train_network.py` to probably work with older versions of LyCORIS. - Fix `train_network.py` to probably work with older versions of LyCORIS.

10
library/custom_train_functions.py
View File

@@ -265,11 +265,6 @@ def get_unweighted_text_embeddings(
text_embedding = enc_out["hidden_states"][-clip_skip] text_embedding = enc_out["hidden_states"][-clip_skip]
text_embedding = text_encoder.text_model.final_layer_norm(text_embedding) text_embedding = text_encoder.text_model.final_layer_norm(text_embedding)
# cover the head and the tail by the starting and the ending tokens
text_input_chunk[:, 0] = text_input[0, 0]
text_input_chunk[:, -1] = text_input[0, -1]
text_embedding = text_encoder(text_input_chunk, attention_mask=None)[0]
if no_boseos_middle: if no_boseos_middle:
if i == 0: if i == 0:
# discard the ending token # discard the ending token
@@ -284,7 +279,12 @@ def get_unweighted_text_embeddings(
text_embeddings.append(text_embedding) text_embeddings.append(text_embedding)
text_embeddings = torch.concat(text_embeddings, axis=1) text_embeddings = torch.concat(text_embeddings, axis=1)
else: else:
if clip_skip is None or clip_skip == 1:
text_embeddings = text_encoder(text_input)[0] text_embeddings = text_encoder(text_input)[0]
else:
enc_out = text_encoder(text_input, output_hidden_states=True, return_dict=True)
text_embeddings = enc_out["hidden_states"][-clip_skip]
text_embeddings = text_encoder.text_model.final_layer_norm(text_embeddings)
return text_embeddings return text_embeddings

10
library/lpw_stable_diffusion.py
View File

@@ -245,11 +245,6 @@ def get_unweighted_text_embeddings(
text_embedding = enc_out["hidden_states"][-clip_skip] text_embedding = enc_out["hidden_states"][-clip_skip]
text_embedding = pipe.text_encoder.text_model.final_layer_norm(text_embedding) text_embedding = pipe.text_encoder.text_model.final_layer_norm(text_embedding)
# cover the head and the tail by the starting and the ending tokens
text_input_chunk[:, 0] = text_input[0, 0]
text_input_chunk[:, -1] = text_input[0, -1]
text_embedding = pipe.text_encoder(text_input_chunk, attention_mask=None)[0]
if no_boseos_middle: if no_boseos_middle:
if i == 0: if i == 0:
# discard the ending token # discard the ending token
@@ -264,7 +259,12 @@ def get_unweighted_text_embeddings(
text_embeddings.append(text_embedding) text_embeddings.append(text_embedding)
text_embeddings = torch.concat(text_embeddings, axis=1) text_embeddings = torch.concat(text_embeddings, axis=1)
else: else:
if clip_skip is None or clip_skip == 1:
text_embeddings = pipe.text_encoder(text_input)[0] 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 return text_embeddings