VictorMylle/Thesis
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Added LSTM model

Browse Source
This commit is contained in:
Victor Mylle
2023-11-28 22:27:15 +00:00
parent ffa19592f9
commit eba10c8f83
6 changed files with 117 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
src/data/dataset.py
View File

@@ -12,10 +12,12 @@ class NrvDataset(Dataset):
full_day_skip: bool = False,
sequence_length=96,
predict_sequence_length=96,
lstm: bool = False,
):
self.data_config = data_config
self.dataframe = dataframe
self.full_day_skip = full_day_skip
self.lstm = lstm
# reset dataframe index
self.dataframe.reset_index(drop=True, inplace=True)
@@ -107,19 +109,26 @@ class NrvDataset(Dataset):
history_features = history_df[self.history_features].values
# combine the history features to one tensor (first one feature, then the next one, etc.)
history_features = torch.tensor(history_features).reshape(-1)
history_features = torch.tensor(history_features)
# get forecast features
forecast_features = forecast_df[self.forecast_features].values
forecast_features = torch.tensor(forecast_features).view(-1)
forecast_features = torch.tensor(forecast_features)
# add last time feature of the history
time_feature = history_df["time_feature"].iloc[-1]
## all features
all_features = torch.cat(
[nrv_features, history_features, forecast_features, torch.tensor([time_feature])], dim=0
)
if not self.lstm:
all_features = torch.cat(
[nrv_features, history_features.reshape(-1), forecast_features.reshape(-1), torch.tensor([time_feature])], dim=0
)
else:
time_features = torch.tensor(history_df["time_feature"].values).reshape(-1, 1)
# combine (96, ) and (96, 2) to (96, 3)
all_features = torch.cat(
[nrv_features.unsqueeze(1), time_features], dim=1
)
# Target sequence, flattened if necessary
nrv_target = forecast_df["nrv"].values
@@ -133,7 +142,7 @@ class NrvDataset(Dataset):
# all features and target to float
all_features = all_features.float()
# to tensors
# to tens&éazzaéaz"ezéors
nrv_target = torch.tensor(nrv_target).float()
return all_features, nrv_target, idx

7
src/data/preprocessing.py
View File

@@ -36,9 +36,10 @@ class DataConfig:
class DataProcessor:
def __init__(self, data_config: DataConfig, path:str="./"):
def __init__(self, data_config: DataConfig, lstm: bool = False, path:str="./"):
self.batch_size = 2048
self.path = path
self.lstm = lstm
self.train_range = (
-np.inf,
@@ -204,6 +205,7 @@ class DataProcessor:
data_config=self.data_config,
full_day_skip=self.full_day_skip,
predict_sequence_length=predict_sequence_length,
lstm=self.lstm,
)
return self.get_dataloader(train_dataset, shuffle=shuffle)
@@ -234,6 +236,7 @@ class DataProcessor:
data_config=self.data_config,
full_day_skip=self.full_day_skip,
predict_sequence_length=predict_sequence_length,
lstm=self.lstm,
)
return self.get_dataloader(test_dataset, shuffle=False)
@@ -274,7 +277,7 @@ class DataProcessor:
predict_sequence_length=self.output_size
)
input, _, _ = next(iter(data_loader))
return input.shape[-1]
return input.shape
def get_time_feature_size(self):
time_feature_size = 1

45
src/models/lstm_model.py Normal file
View File

@@ -0,0 +1,45 @@
import torch
class LSTMModel(torch.nn.Module):
def __init__(self, inputSize, output_size, num_layers: int, hidden_size: int, dropout: float = 0.2):
super(LSTMModel, self).__init__()
self.inputSize = inputSize
self.output_size = output_size
self.num_layers = num_layers
self.hidden_size = hidden_size
self.dropout = dropout
self.lstm = torch.nn.LSTM(input_size=inputSize[-1], hidden_size=hidden_size, num_layers=num_layers, dropout=dropout, batch_first=True)
self.linear = torch.nn.Linear(hidden_size, output_size)
def forward(self, x):
# Forward pass through the LSTM layers
_, (hidden_state, _) = self.lstm(x)
# Use the hidden state from the last time step for the output
output = self.linear(hidden_state[-1])
return output
class GRUModel(torch.nn.Module):
def __init__(self, inputSize, output_size, num_layers: int, hidden_size: int, dropout: float = 0.2):
super(GRUModel, self).__init__()
self.inputSize = inputSize
self.output_size = output_size
self.num_layers = num_layers
self.hidden_size = hidden_size
self.dropout = dropout
self.gru = torch.nn.GRU(input_size=inputSize[-1], hidden_size=hidden_size, num_layers=num_layers, dropout=dropout, batch_first=True)
self.linear = torch.nn.Linear(hidden_size, output_size)
def forward(self, x):
# Forward pass through the GRU layers
_, hidden_state = self.gru(x)
# Use the hidden state from the last time step for the output
output = self.linear(hidden_state[-1])
return output

19
src/models/time_embedding_layer.py
View File

@@ -10,19 +10,20 @@ class TimeEmbedding(nn.Module):
def forward(self, x):
# Extract the last 'time_features' from the input
time_feature = x[:, -1]
time_feature = x[..., -1] # Use ellipsis to access the last dimension
# convert to int
time_feature = time_feature.int()
# Embed these time features
# print max value of time_feature
if time_feature.max() > self.time_features:
# print the row from x that includes the max value in the last column
print(x[time_feature == time_feature.max()])
print("time feature max value is greater than time features")
embedded_time = self.embedding(time_feature)
# Concatenate the embedded features with the original input (minus the last 'time feature')
return torch.cat((x[:, :-1], embedded_time), dim=1)
return torch.cat((x[..., :-1], embedded_time), dim=-1) # Use -1 to specify the last dimension
def output_dim(self, input_dim):
return input_dim + self.embedding.embedding_dim - 1
# Create a list from the input dimension
input_dim_list = list(input_dim)
# Modify the last dimension
input_dim_list[-1] = input_dim_list[-1] - 1 + self.embedding.embedding_dim
# Convert the list back to a torch.Size object
output_dim = torch.Size(input_dim_list)
return output_dim

21
src/training_scripts/autoregressive_quantiles.py
View File

@@ -1,3 +1,4 @@
from src.models.lstm_model import LSTMModel, GRUModel
from src.data import DataProcessor, DataConfig
from src.trainers.quantile_trainer import AutoRegressiveQuantileTrainer, NonAutoRegressiveQuantileRegression
from src.trainers.probabilistic_baseline import ProbabilisticBaselineTrainer
@@ -28,19 +29,21 @@ data_config.LOAD_FORECAST = True
data_config.QUARTER = True
data_config.DAY_OF_WEEK = True
data_config = task.connect(data_config, name="data_features")
# data_config = task.connect(data_config, name="data_features")
data_processor = DataProcessor(data_config, path="")
data_processor.set_batch_size(1024)
data_processor = DataProcessor(data_config, path="", lstm=True)
data_processor.set_batch_size(512)
data_processor.set_full_day_skip(False)
#### Hyperparameters ####
data_processor.set_output_size(1)
inputDim = data_processor.get_input_size()
learningRate = 0.0001
learningRate = 0.001
epochs = 100
print("Input dim: ", inputDim)
# add parameters to clearml
quantiles = task.get_parameter("general/quantiles", cast=True)
if quantiles is None:
@@ -49,8 +52,9 @@ if quantiles is None:
# model = LinearRegression(inputDim, len(quantiles))
time_embedding = TimeEmbedding(data_processor.get_time_feature_size(), 4)
non_linear_regression_model = NonLinearRegression(time_embedding.output_dim(inputDim), len(quantiles), hiddenSize=1024, numLayers=5)
model = nn.Sequential(time_embedding, non_linear_regression_model)
# non_linear_regression_model = NonLinearRegression(time_embedding.output_dim(inputDim), len(quantiles), hiddenSize=1024, numLayers=5)
lstm_model = GRUModel(time_embedding.output_dim(inputDim), len(quantiles), hidden_size=512, num_layers=2)
model = nn.Sequential(time_embedding, lstm_model)
optimizer = torch.optim.Adam(model.parameters(), lr=learningRate)
#### Trainer ####
@@ -62,9 +66,10 @@ trainer = AutoRegressiveQuantileTrainer(
"cuda",
debug=True,
)
trainer.add_metrics_to_track(
[PinballLoss(quantiles), MSELoss(), L1Loss(), CRPSLoss()]
)
trainer.early_stopping(patience=10)
trainer.plot_every(5)
trainer.train(task=task, epochs=epochs, remotely=False)
trainer.plot_every(100)
trainer.train(task=task, epochs=epochs, remotely=True)

42
src/training_scripts/hyperparameter_optimizer.py
View File

@@ -5,7 +5,6 @@ from clearml.automation.optuna import OptimizerOptuna
from clearml.automation import (
DiscreteParameterRange, HyperParameterOptimizer, RandomSearch,
UniformIntegerParameterRange)
from src.data.preprocessing import DataConfig
# trying to load Bayesian optimizer package
try:
@@ -21,17 +20,28 @@ except ImportError as ex:
'we will be using RandomSearch strategy instead')
aSearchStrategy = RandomSearch
# input task id to optimize
input_task_id = input("Please enter the task id to optimize: ")
# input task id to optimize using argparse
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--task_id", help="Task ID to optimize", type=str)
args = parser.parse_args()
input_task_id = args.task_id
# check if task id is valid
if not Task.get_task(task_id=input_task_id):
raise ValueError("Invalid task id")
task = Task.init(project_name='Hyper-Parameter Optimization',
task_name='Automatic Hyper-Parameter Optimization',
Task.add_requirements("requirements.txt")
Task.ignore_requirements("torch")
Task.ignore_requirements("torchvision")
Task.ignore_requirements("tensorboard")
task = Task.init(project_name='Thesis/NrvForecast',
task_name='Autoregressive Quantile Regression Hyper-Parameter Optimization',
task_type=Task.TaskTypes.optimizer,
reuse_last_task_id=False)
task.set_base_docker(f"docker.io/clearml/pytorch-cuda-gcc:2.0.0-cuda11.7-cudnn8-runtime --env GIT_SSL_NO_VERIFY=true --env CLEARML_AGENT_GIT_USER=VictorMylle --env CLEARML_AGENT_GIT_PASS=Voetballer1" )
task.set_packages("requirements.txt")
execution_queue = "default"
@@ -40,36 +50,42 @@ execution_queue = "default"
#### Quantiles ####
quantile_lists = [
[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], # Deciles
[0.25, 0.5, 0.75], # Quartiles
[0.05, 0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95], # 10% Increments, Excluding Extremes
[0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.99], # Combining Deciles with Extremes
[0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1], # Including 0 and 1
[0.01, 0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95, 0.99], # Mixed Small and Large Increments
[0.2, 0.4, 0.6, 0.8], # 20% Increments
[0.125, 0.375, 0.625, 0.875], # Eighths
[0.05, 0.10, 0.20, 0.30, 0.40, 0.50, 0.60, 0.70, 0.80, 0.90], # 10% Increments
[0.01, 0.02, 0.03, 0.04, 0.05, 0.1, 0.15, 0.2, 0.3, 0.5] # Mixed Fine and Coarser Increments
[0.05, 0.10, 0.20, 0.30, 0.40, 0.50, 0.60, 0.70, 0.80, 0.90, 0.95], # 10% Increments
]
quantiles_range = DiscreteParameterRange("general/quantiles", values=quantile_lists)
#### Data Config ####
quarter_range = DiscreteParameterRange("data_features/quarter", values=[True, False])
day_of_week_range = DiscreteParameterRange("data_features/day_of_week", values=[True, False])
load_forecast_range = DiscreteParameterRange("data_features/load_forecast", values=[True, False])
load_history_range = DiscreteParameterRange("data_features/load_history", values=[True, False])
### OPTIMIZER OBJECT ###
optimizer = HyperParameterOptimizer(
base_task_id=input_task_id,
objective_metric_title="PinballLoss",
objective_metric_series="test",
objective_metric_title="Summary",
objective_metric_series="test_CRPSLoss",
objective_metric_sign="min",
execution_queue=execution_queue,
max_number_of_concurrent_tasks=1,
optimizer_class=aSearchStrategy,
max_iteration_per_job=50,
# save_top_k_tasks_only=3,
pool_period_min=0.2,
total_max_jobs=15,
hyper_parameters=[
quantiles_range,
quarter_range,
day_of_week_range,
load_forecast_range,
load_history_range
]
)