-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_CoLT5_refined.py
230 lines (182 loc) · 7.77 KB
/
train_CoLT5_refined.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
import torch
import torch.nn as nn
from torch.utils.data import DataLoader, random_split
from datasets import load_dataset
from transformers import T5Tokenizer, DataCollatorWithPadding
from colt5_attention.transformer_block import ConditionalRoutedTransformerBlock, ConditionalRoutedDecoderBlock
from colt5_attention.colt5_model import CoLT5
from tqdm import tqdm
import os
import matplotlib.pyplot as plt # For plotting loss curves
# ============================
# 1. Data Loading and Preprocessing
# ============================
# Load the train split of TriviaQA
dataset = load_dataset('trivia_qa', 'unfiltered', split='train')
# Initialize tokenizer
tokenizer = T5Tokenizer.from_pretrained('t5-small')
def preprocess_function(examples):
# Prepare inputs
inputs = [f"trivia question: {question}" for question in examples['question']]
model_inputs = tokenizer(
inputs,
max_length=128,
truncation=True,
padding='max_length',
return_tensors='pt'
)
# Prepare labels
answers = [answer['value'] if len(answer['value']) > 0 else "" for answer in examples['answer']]
labels = tokenizer(
answers,
max_length=128,
truncation=True,
padding='max_length',
return_tensors='pt'
)
model_inputs['labels'] = labels['input_ids']
# Convert attention_mask to boolean type
model_inputs['attention_mask'] = model_inputs['attention_mask'].bool()
return model_inputs
# Apply preprocessing
tokenized_dataset = dataset.map(preprocess_function, batched=True)
# Remove unnecessary columns
columns_to_remove = ['question', 'question_id', 'question_source', 'entity_pages', 'search_results', 'answer']
tokenized_dataset = tokenized_dataset.remove_columns(columns_to_remove)
# ============================
# 2. Dataset Splitting
# ============================
# Split the dataset into training and validation sets (90% train, 10% validation)
train_size = int(0.9 * len(tokenized_dataset))
val_size = len(tokenized_dataset) - train_size
train_dataset, val_dataset = random_split(tokenized_dataset, [train_size, val_size])
# Initialize DataCollator
data_collator = DataCollatorWithPadding(tokenizer, padding=True)
# Create DataLoaders
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True, collate_fn=data_collator)
val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False, collate_fn=data_collator)
# ============================
# 3. Model Initialization
# ============================
# Load CoLT5 model
model = CoLT5(num_layers=6, dim=512).to('cuda')
model.load_state_dict(torch.load('./checkpoints_926/best_colt5.pth'))
# ============================
# 4. Training Setup
# ============================
# Define Optimizer and Scheduler
optimizer = torch.optim.AdamW(model.parameters(), lr=5e-5)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=2, verbose=True)
# Define Loss Function
loss_fn = nn.CrossEntropyLoss(ignore_index=tokenizer.pad_token_id)
# Directory for checkpoints
checkpoint_dir = './checkpoints_927'
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
# Initialize lists to track loss
train_losses = []
val_losses = []
# Early Stopping Parameters
patience = 5
best_val_loss = float('inf')
counter = 0
# ============================
# 5. Training Loop with Validation and Loss Tracking
# ============================
epochs = 10 # Increased number of epochs for better convergence
for epoch in range(epochs):
# Training Phase
model.train()
epoch_train_loss = 0
loop = tqdm(train_loader, leave=True, desc=f"Epoch {epoch+1}/{epochs}")
for batch in loop:
input_ids = batch['input_ids'].to('cuda')
mask = batch['attention_mask'].to('cuda')
labels = batch['labels'].to('cuda')
# Initialize decoder_input_ids with pad tokens and shift labels
decoder_input_ids = torch.full((labels.size(0), labels.size(1)), tokenizer.pad_token_id, dtype=torch.long).to('cuda')
decoder_input_ids[:, 1:] = labels[:, :-1] # Shift labels for decoder input
decoder_input_ids[:, 0] = tokenizer.pad_token_id # Ensure the first token is pad
decoder_mask = labels != tokenizer.pad_token_id
decoder_mask[:, 1:] = decoder_mask[:,:-1]
decoder_mask[:,0] = True
decoder_mask = decoder_mask.to('cuda')
optimizer.zero_grad()
# Forward pass
logits = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids, mask=mask, decoder_mask=decoder_mask)
# Compute loss
loss = loss_fn(logits.view(-1, logits.size(-1)), labels.view(-1))
# Backward pass
loss.backward()
# Gradient Clipping (optional but recommended)
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
# Optimizer step
optimizer.step()
# Accumulate loss
epoch_train_loss += loss.item()
# Update progress bar
loop.set_postfix(loss=loss.item())
# Compute average training loss for the epoch
avg_train_loss = epoch_train_loss / len(train_loader)
train_losses.append(avg_train_loss)
# Validation Phase
model.eval()
epoch_val_loss = 0
with torch.no_grad():
for batch in tqdm(val_loader, desc="Validation", leave=False):
input_ids = batch['input_ids'].to('cuda')
mask = batch['attention_mask'].to('cuda')
labels = batch['labels'].to('cuda')
# Initialize decoder_input_ids with pad tokens and shift labels
decoder_input_ids = torch.full((labels.size(0), labels.size(1)), tokenizer.pad_token_id, dtype=torch.long).to('cuda')
decoder_input_ids[:, 1:] = labels[:, :-1]
decoder_input_ids[:, 0] = tokenizer.pad_token_id
decoder_mask = labels != tokenizer.pad_token_id
decoder_mask[:, 1:] = decoder_mask[:,:-1]
decoder_mask[:,0] = True
decoder_mask = decoder_mask.to('cuda')
# Forward pass
logits = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids, mask=mask, decoder_mask=decoder_mask)
# Compute loss
loss = loss_fn(logits.view(-1, logits.size(-1)), labels.view(-1))
epoch_val_loss += loss.item()
# Compute average validation loss for the epoch
avg_val_loss = epoch_val_loss / len(val_loader)
val_losses.append(avg_val_loss)
# Scheduler step based on validation loss
scheduler.step(avg_val_loss)
# Early Stopping Check
if avg_val_loss < best_val_loss:
best_val_loss = avg_val_loss
counter = 0
# Save the best model
torch.save(model.state_dict(), f'{checkpoint_dir}/best_colt5.pth')
print(f"Epoch {epoch + 1} improved. Saving best model.")
else:
counter += 1
print(f"Epoch {epoch + 1} did not improve.")
if counter >= patience:
print("Early stopping triggered.")
break
# Save a checkpoint at the end of each epoch
torch.save(model.state_dict(), f'{checkpoint_dir}/colt5_epoch_{epoch+1}.pth')
print(f"Epoch {epoch + 1} completed. Training Loss: {avg_train_loss:.4f}, Validation Loss: {avg_val_loss:.4f}")
# ============================
# 6. Save the Final Model
# ============================
# Save the final model
# model.save_pretrained('./colt5_triviaqa_model')
# ============================
# 7. Plotting the Loss Curves
# ============================
# Plot the loss curves
plt.figure(figsize=(10, 5))
plt.plot(range(1, len(train_losses) + 1), train_losses, label='Training Loss')
plt.plot(range(1, len(val_losses) + 1), val_losses, label='Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.title('Training and Validation Loss over Epochs')
plt.legend()
plt.grid(True)
plt.savefig('loss_curve.png') # Save the plot as an image file
plt.show()