added bootstrap avg / ensemble preds

main
Felix Jan Michael Mucha 2025-02-16 03:56:50 +01:00
parent 603eab83b4
commit 4469f55889
6 changed files with 299 additions and 124 deletions

108
BERT.py
View File

@ -3,10 +3,12 @@ import random
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.data import DataLoader, Subset
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
from transformers import BertForSequenceClassification, AutoTokenizer
import numpy as np
from datetime import datetime
import json
import Datasets
import dataset_helper
@ -53,20 +55,16 @@ if __name__ == '__main__':
# Config
"max_len": 128,
# Training
"epochs": 10,
"epochs": 1,
"patience": 7,
"batch_size": 32,
"learning_rate": 0.001,
"learning_rate": 1e-6,
"weight_decay": 5e-4 ,
# Model
"filter_sizes": [2, 3, 4, 5],
"num_filters": 150,
"dropout": 0.6
}
# Configs
MODEL_NAME = 'BERT.pt'
HIST_NAME = 'BERT_history'
GLOVE_PATH = 'data/glove.6B.100d.txt'
DATA_PATH = 'data/hack.csv'
FREEZE_BERT = False
@ -74,6 +72,11 @@ if __name__ == '__main__':
TEST_SIZE = 0.1
VAL_SIZE = 0.1
N_MODELS = 2
models = []
timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
# Daten laden und vorbereiten
embedding_matrix, word_index, vocab_size, d_model = dataset_helper.get_embedding_matrix(
gloVe_path=GLOVE_PATH, emb_len=EMBEDDING_DIM)
@ -96,42 +99,77 @@ if __name__ == '__main__':
val_loader = DataLoader(val_dataset, batch_size=params["batch_size"], shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=params["batch_size"], shuffle=False)
# Modell initialisieren
model = CustomBert(dropout=params["dropout"])
subset_size = len(train_dataset) // N_MODELS
device = ml_helper.get_device(verbose=True, include_mps=False)
model = model.to(device)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=params["learning_rate"], weight_decay=params["weight_decay"])
early_stopping = EarlyStopping.EarlyStoppingCallback(patience=params["patience"], verbose=True, model_name=MODEL_NAME)
for i in range(N_MODELS):
model_name = f'BERT.pt'
hist_name = f'BERT_history'
hist = ml_history.History()
if N_MODELS > 1:
model_name = f'BERT_{i}_ensemble.pt'
hist_name = f'BERT_{i}_ensemble_history'
# Training und Validierung
for epoch in range(params["epochs"]):
ml_train.train_epoch(model, train_loader, criterion, optimizer, device, hist, epoch, params["epochs"], bert_freeze=FREEZE_BERT, is_bert=True)
subset_indices = dataset_helper.ensemble_data_idx(train_dataset.labels, N_MODELS, i, methods='bootstrap')
train_dataset_sub = Subset(train_dataset, subset_indices)
train_loader = DataLoader(train_dataset_sub, batch_size=params["batch_size"], shuffle=True)
val_rmse = ml_train.validate_epoch(model, val_loader, epoch, criterion, device, hist, is_bert=True)
early_stopping(val_rmse, model)
if early_stopping.early_stop:
print("Early stopping triggered.")
break
model = CustomBert(dropout=params["dropout"])
model = model.to(device)
# Load best model
model.load_state_dict(torch.load('models/checkpoints/' + MODEL_NAME))
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=params["learning_rate"], weight_decay=params["weight_decay"])
early_stopping = EarlyStopping.EarlyStoppingCallback(patience=params["patience"], verbose=True, model_name=model_name)
# Test Evaluation
test_labels, test_preds = ml_train.test_loop(model, test_loader, device, is_bert=True)
hist = ml_history.History()
hist.add_test_results(test_labels, test_preds)
# Training und Validierung
for epoch in range(params["epochs"]):
ml_train.train_epoch(model, train_loader, criterion, optimizer, device, hist, epoch, params["epochs"], bert_freeze=FREEZE_BERT, is_bert=True)
# save training history
hist.save_history(HIST_NAME)
val_rmse = ml_train.validate_epoch(model, val_loader, epoch, criterion, device, hist, is_bert=True)
# RMSE, MAE und R²-Score für das Test-Set
test_mae = mean_absolute_error(test_labels, test_preds)
test_rmse = np.sqrt(mean_squared_error(test_labels, test_preds))
test_r2 = r2_score(test_labels, test_preds)
print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
early_stopping(val_rmse, model)
if early_stopping.early_stop:
print("Early stopping triggered.")
break
# Load best model
model.load_state_dict(torch.load('models/checkpoints/' + model_name, weights_only=False))
models.append(model)
# Test Evaluation
test_labels, test_preds = ml_train.test_loop(model, test_loader, device, is_bert=True)
hist.add_test_results(test_labels, test_preds)
# save training history
hist.save_history(hist_name, timestamp)
# RMSE, MAE und R²-Score für das Test-Set
test_mae = mean_absolute_error(test_labels, test_preds)
test_rmse = np.sqrt(mean_squared_error(test_labels, test_preds))
test_r2 = r2_score(test_labels, test_preds)
print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
if N_MODELS >1:
# Ensemble Prediction
ensemble_test_preds = ml_train.ensemble_predict(models, test_loader, device, is_bert=True)
ensemble_avg_preds = np.mean(ensemble_test_preds, axis=0)
# Save ensemble predictions as json
ensemble_preds_path = f'histories/ensemble_preds_BERT_{timestamp}.json'
with open(ensemble_preds_path, 'w') as f:
json.dump(ensemble_avg_preds.tolist(), f)
# Test Evaluation
test_labels = test_dataset.labels
test_mse = mean_squared_error(test_labels, ensemble_avg_preds)
test_mae = mean_absolute_error(test_labels, ensemble_avg_preds)
test_r2 = r2_score(test_labels, ensemble_avg_preds)
print(f"Ensemble Test RMSE: {test_mse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")

119
CNN.py
View File

@ -3,9 +3,11 @@ import random
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.data import DataLoader, Subset
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
import numpy as np
from datetime import datetime
import json
import Datasets
import dataset_helper
@ -57,7 +59,7 @@ if __name__ == '__main__':
# Config
"max_len": 280,
# Training
"epochs": 25,
"epochs": 5,
"patience": 7,
"batch_size": 32,
"learning_rate": 0.001,
@ -69,14 +71,17 @@ if __name__ == '__main__':
}
# Configs
MODEL_NAME = 'CNN.pt'
HIST_NAME = 'CNN_history'
GLOVE_PATH = 'data/glove.6B.100d.txt'
DATA_PATH = 'data/hack.csv'
EMBEDDING_DIM = 100
TEST_SIZE = 0.1
VAL_SIZE = 0.1
N_MODELS = 1
models = []
timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
# Daten laden und vorbereiten
embedding_matrix, word_index, vocab_size, d_model = dataset_helper.get_embedding_matrix(
gloVe_path=GLOVE_PATH, emb_len=EMBEDDING_DIM)
@ -95,53 +100,83 @@ if __name__ == '__main__':
val_loader = DataLoader(val_dataset, batch_size=params["batch_size"], shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=params["batch_size"], shuffle=False)
# Modell initialisieren
model = EnhancedCNNRegressor(
vocab_size=vocab_size,
embedding_dim=EMBEDDING_DIM,
filter_sizes=params["filter_sizes"],
num_filters=params["num_filters"],
embedding_matrix=embedding_matrix,
dropout=params["dropout"]
)
subset_size = len(train_dataset) // N_MODELS
device = ml_helper.get_device(verbose=True, include_mps=False)
model = model.to(device)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=params["learning_rate"], weight_decay=params["weight_decay"])
early_stopping = EarlyStopping.EarlyStoppingCallback(patience=params["patience"], verbose=True, model_name=MODEL_NAME)
for i in range(N_MODELS):
model_name = f'CNN.pt'
hist_name = f'CNN_history'
hist = ml_history.History()
if N_MODELS > 1:
model_name = f'CNN_{i}_ensemble.pt'
hist_name = f'CNN_{i}_ensemble_history'
subset_indices = dataset_helper.ensemble_data_idx(train_dataset.labels, N_MODELS, i, methods='bootstrap')
train_dataset_sub = Subset(train_dataset, subset_indices)
train_loader = DataLoader(train_dataset_sub, batch_size=params["batch_size"], shuffle=True)
# Training und Validierung
for epoch in range(params["epochs"]):
ml_train.train_epoch(model, train_loader, criterion, optimizer, device, hist, epoch, params["epochs"])
model = EnhancedCNNRegressor(
vocab_size=vocab_size,
embedding_dim=EMBEDDING_DIM,
filter_sizes=params["filter_sizes"],
num_filters=params["num_filters"],
embedding_matrix=embedding_matrix,
dropout=params["dropout"]
)
model = model.to(device)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=params["learning_rate"], weight_decay=params["weight_decay"])
early_stopping = EarlyStopping.EarlyStoppingCallback(patience=params["patience"], verbose=True, model_name=model_name)
val_rmse = ml_train.validate_epoch(model, val_loader, epoch, criterion, device, hist)
hist = ml_history.History()
early_stopping(val_rmse, model)
if early_stopping.early_stop:
print("Early stopping triggered.")
break
# Training und Validierung
for epoch in range(params["epochs"]):
ml_train.train_epoch(model, train_loader, criterion, optimizer, device, hist, epoch, params["epochs"])
# save training history
hist.save_history(HIST_NAME)
val_rmse = ml_train.validate_epoch(model, val_loader, epoch, criterion, device, hist)
# Load best model
model.load_state_dict(torch.load('models/checkpoints/' + MODEL_NAME))
early_stopping(val_rmse, model)
if early_stopping.early_stop:
print("Early stopping triggered.")
break
# Test Evaluation
test_labels, test_preds = ml_train.test_loop(model, test_loader, device)
# Load best model
model.load_state_dict(torch.load('models/checkpoints/' + model_name, weights_only=False))
models.append(model)
hist.add_test_results(test_labels, test_preds)
# Test Evaluation
test_labels, test_preds = ml_train.test_loop(model, test_loader, device)
# save training history
hist.save_history(HIST_NAME)
hist.add_test_results(test_labels, test_preds)
# save training history
hist.save_history(hist_name, timestamp)
# RMSE, MAE und R²-Score für das Test-Set
test_mae = mean_absolute_error(test_labels, test_preds)
test_rmse = np.sqrt(mean_squared_error(test_labels, test_preds))
test_r2 = r2_score(test_labels, test_preds)
print(f"Model: {model_name} Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
if N_MODELS >1:
# Ensemble Prediction
ensemble_test_preds = ml_train.ensemble_predict(models, test_loader, device)
ensemble_avg_preds = np.mean(ensemble_test_preds, axis=0)
# Save ensemble predictions as json
ensemble_preds_path = f'histories/ensemble_preds_CNN_{timestamp}.json'
with open(ensemble_preds_path, 'w') as f:
json.dump(ensemble_avg_preds.tolist(), f)
# Test Evaluation
test_labels = test_dataset.labels.to_numpy()
test_mse = mean_squared_error(test_labels, ensemble_avg_preds)
test_mae = mean_absolute_error(test_labels, ensemble_avg_preds)
test_r2 = r2_score(test_labels, ensemble_avg_preds)
print(f"Ensemble Test RMSE: {test_mse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
# RMSE, MAE und R²-Score für das Test-Set
test_mae = mean_absolute_error(test_labels, test_preds)
test_rmse = np.sqrt(mean_squared_error(test_labels, test_preds))
test_r2 = r2_score(test_labels, test_preds)
print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")

View File

@ -1,11 +1,14 @@
import math
import random
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.data import DataLoader, Subset
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
import numpy as np
from datetime import datetime
import json
import Datasets
import dataset_helper
@ -14,6 +17,12 @@ import ml_helper
import ml_history
import ml_train
SEED = 501
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed_all(SEED)
torch.backends.cudnn.deterministic = True
class PositionalEncoding(nn.Module):
"""
@ -102,7 +111,7 @@ if __name__ == '__main__':
# Config
"max_len": 280,
# Training
"epochs": 25,
"epochs": 1,
"patience": 7,
"batch_size": 32,
"learning_rate": 1e-4, # 1e-4
@ -113,17 +122,19 @@ if __name__ == '__main__':
'hiden_dim': 2048,
'num_layers': 6
}
# TODO set seeds
# Configs
MODEL_NAME = 'transfomrer.pt'
HIST_NAME = 'transformer_history'
GLOVE_PATH = 'data/glove.6B.100d.txt'
DATA_PATH = 'data/hack.csv'
EMBEDDING_DIM = 100
TEST_SIZE = 0.1
VAL_SIZE = 0.1
N_MODELS = 2
models = []
timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
# Daten laden und vorbereiten
embedding_matrix, word_index, vocab_size, d_model = dataset_helper.get_embedding_matrix(
gloVe_path=GLOVE_PATH, emb_len=EMBEDDING_DIM)
@ -142,55 +153,83 @@ if __name__ == '__main__':
val_loader = DataLoader(val_dataset, batch_size=params["batch_size"], shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=params["batch_size"], shuffle=False)
# Modell initialisieren
model = TransformerBinaryClassifier(
embeddings=embedding_matrix,
nhead=params['nhead'],
dim_feedforward=params['hiden_dim'],
num_layers=params['num_layers'],
positional_dropout=params["dropout"],
classifier_dropout=params["dropout"],
)
subset_size = len(train_dataset) // N_MODELS
device = ml_helper.get_device(verbose=True, include_mps=False)
model = model.to(device)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=params["learning_rate"]) #, weight_decay=params["weight_decay"])
early_stopping = EarlyStopping.EarlyStoppingCallback(patience=params["patience"], verbose=True, model_name=MODEL_NAME)
for i in range(N_MODELS):
model_name = f'Transformer.pt'
hist_name = f'Transformer_history'
if N_MODELS > 1:
model_name = f'Transformer_{i}_ensemble.pt'
hist_name = f'Transformer_{i}_ensemble_history'
hist = ml_history.History()
subset_indices = dataset_helper.ensemble_data_idx(train_dataset.labels, N_MODELS, i, methods='bootstrap')
train_dataset_sub = Subset(train_dataset, subset_indices)
train_loader = DataLoader(train_dataset_sub, batch_size=params["batch_size"], shuffle=True)
# Training und Validierung
for epoch in range(params["epochs"]):
ml_train.train_epoch(model, train_loader, criterion, optimizer, device, hist, epoch, params["epochs"])
# Modell initialisieren
model = TransformerBinaryClassifier(
embeddings=embedding_matrix,
nhead=params['nhead'],
dim_feedforward=params['hiden_dim'],
num_layers=params['num_layers'],
positional_dropout=params["dropout"],
classifier_dropout=params["dropout"],
)
model = model.to(device)
val_rmse = ml_train.validate_epoch(model, val_loader, epoch, criterion, device, hist)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=params["learning_rate"]) #, weight_decay=params["weight_decay"])
early_stopping = EarlyStopping.EarlyStoppingCallback(patience=params["patience"], verbose=True, model_name=model_name)
early_stopping(val_rmse, model)
if early_stopping.early_stop:
print("Early stopping triggered.")
break
hist = ml_history.History()
# save training history
hist.save_history(HIST_NAME)
# Training und Validierung
for epoch in range(params["epochs"]):
ml_train.train_epoch(model, train_loader, criterion, optimizer, device, hist, epoch, params["epochs"])
# save training history
hist.save_history(HIST_NAME)
val_rmse = ml_train.validate_epoch(model, val_loader, epoch, criterion, device, hist)
# Load best model
model.load_state_dict(torch.load('models/checkpoints/' + MODEL_NAME))
early_stopping(val_rmse, model)
if early_stopping.early_stop:
print("Early stopping triggered.")
break
# Test Evaluation
test_labels, test_preds = ml_train.test_loop(model, test_loader, device)
# Load best model
model.load_state_dict(torch.load('models/checkpoints/' + model_name, weights_only=False))
models.append(model)
# Test Evaluation
test_labels, test_preds = ml_train.test_loop(model, test_loader, device)
hist.add_test_results(test_labels, test_preds)
hist.add_test_results(test_labels, test_preds)
# save training history
hist.save_history(HIST_NAME)
# save training history
hist.save_history(hist_name, timestamp)
# RMSE, MAE und R²-Score für das Test-Set
test_mae = mean_absolute_error(test_labels, test_preds)
test_rmse = np.sqrt(mean_squared_error(test_labels, test_preds))
test_r2 = r2_score(test_labels, test_preds)
print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
# RMSE, MAE und R²-Score für das Test-Set
test_mae = mean_absolute_error(test_labels, test_preds)
test_rmse = np.sqrt(mean_squared_error(test_labels, test_preds))
test_r2 = r2_score(test_labels, test_preds)
print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
if N_MODELS >1:
# Ensemble Prediction
ensemble_test_preds = ml_train.ensemble_predict(models, test_loader, device)
ensemble_avg_preds = np.mean(ensemble_test_preds, axis=0)
# Save ensemble predictions as json
ensemble_preds_path = f'histories/ensemble_preds_Transformer_{timestamp}.json'
with open(ensemble_preds_path, 'w') as f:
json.dump(ensemble_avg_preds.tolist(), f)
# Test Evaluation
test_labels = test_dataset.labels.to_numpy()
test_mse = mean_squared_error(test_labels, ensemble_avg_preds)
test_mae = mean_absolute_error(test_labels, ensemble_avg_preds)
test_r2 = r2_score(test_labels, ensemble_avg_preds)
print(f"Ensemble Test RMSE: {test_mse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")

View File

@ -8,6 +8,7 @@ import torch
import regex as re
def load_glove_embeddings(glove_file_path, emb_len=100):
print('Loading GloVe embeddings...')
embeddings_index = {}
with open(glove_file_path, 'r', encoding='utf-8') as f:
for line in f:
@ -99,4 +100,39 @@ def split_data(X, y, test_size=0.1, val_size=0.1):
for key in ret_dict.keys():
print(key, len(ret_dict[key]['X']), len(ret_dict[key]['y']))
return ret_dict
return ret_dict
def ensemble_data_idx(labels, n_models, cur_models_idx, methods='bootstrap'):
if methods == 'bootstrap':
# Calculate the size of the subset
subset_size = len(labels) // n_models
# Calculate the start and end index of the subset
start_idx = cur_models_idx * subset_size
end_idx = start_idx + subset_size
# Calculate the indices of the subset
subset_indices = list(range(0, start_idx)) + list(range(end_idx, len(labels)))
return subset_indices
if methods == 'shuffle':
subset_indices = np.random.permutation(len(labels))
return subset_indices
if methods == 'random':
subset_indices = np.random.choice(len(labels), len(labels), replace=False)
return subset_indices
if methods == 'flatten_normal_dist':
# TODO: test this and plot if it works
subset_size = len(labels) // n_models
std_range = 1
mean = np.mean(labels)
std = np.std(labels)
# Randomly select samples arounnd the mean in the std
del_subset_indices = np.random.choice(np.where((labels >= mean - std_range * std) & (labels <= mean + std_range * std))[0], size=subset_size, replace=False)
subset = np.delete(labels, del_subset_indices)
# TODO i dont think this really uses the indices
subset_indices = np.where(np.isin(labels, subset))[0]
return subset_indices
else:
raise ValueError(f"Unknown method: {methods}")

View File

@ -99,11 +99,12 @@ class History:
return history_to_save
def save_history(self, hist_name):
def save_history(self, hist_name, timestamp=None):
directory = "histories"
if not os.path.exists(directory):
os.makedirs(directory) # Create the directory if it does not exist
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
if timestamp is None:
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
filepath = os.path.join(directory, f"{hist_name}_{timestamp}.json")
# Needed for saving the history to a json file:

View File

@ -85,3 +85,29 @@ def test_loop(model, test_loader, device, is_bert=False):
test_labels.extend(labels.cpu().detach().numpy())
return test_labels, test_preds
def ensemble_predict(models, test_loader, device, is_bert=False):
for model in models:
model.eval()
test_preds = []
with torch.no_grad():
for batch in test_loader:
if is_bert:
input_ids = batch['input_ids'].to(device)
attention_mask = batch['attention_mask'].to(device)
predictions = [model(input_ids, attention_mask=attention_mask).float().cpu().detach().numpy() for model in models]
else:
X_batch, y_batch = batch
X_batch, y_batch = X_batch.to(device), y_batch.to(device).float()
predictions = [model(X_batch).float().cpu().detach().numpy() for model in models]
predictions = predictions
test_preds.append(predictions)
#check if predictions are empty lists
if not test_preds[0]:
raise ValueError("No predictions were made in ensemble prediction.")
test_preds = np.concatenate(test_preds, axis=1)
return test_preds