RobbiePasquale
commited on
Commit
•
a8090dd
1
Parent(s):
70782ac
Update distill.py
Browse files- distill.py +838 -264
distill.py
CHANGED
@@ -1,264 +1,838 @@
|
|
1 |
-
import
|
2 |
-
import
|
3 |
-
import
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
import
|
9 |
-
|
10 |
-
import
|
11 |
-
import
|
12 |
-
|
13 |
-
|
14 |
-
from torch.utils.
|
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 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import math
|
3 |
+
import os
|
4 |
+
import sys
|
5 |
+
import json
|
6 |
+
import jsonlines
|
7 |
+
import copy
|
8 |
+
from typing import List, Optional, Tuple
|
9 |
+
|
10 |
+
import torch
|
11 |
+
import torch.nn as nn
|
12 |
+
import torch.nn.functional as F
|
13 |
+
import torch.optim as optim
|
14 |
+
from torch.utils.data import DataLoader, Dataset, random_split
|
15 |
+
from torch.cuda.amp import autocast, GradScaler
|
16 |
+
from torch.utils.tensorboard import SummaryWriter
|
17 |
+
|
18 |
+
from datasets import load_dataset
|
19 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
|
20 |
+
from tqdm import tqdm
|
21 |
+
|
22 |
+
# Set up device
|
23 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
24 |
+
|
25 |
+
# ======================================
|
26 |
+
# Import Custom Components from lightbulb_custom
|
27 |
+
# ======================================
|
28 |
+
from lightbulb_custom import (
|
29 |
+
RotaryPositionalEncoding,
|
30 |
+
MultiHeadAttention,
|
31 |
+
MoE,
|
32 |
+
TransformerBlock,
|
33 |
+
Transformer,
|
34 |
+
InfoNCE_Loss,
|
35 |
+
CovarianceRegularization,
|
36 |
+
DynamicsPerformanceLoss,
|
37 |
+
ThoughtConsistencyLoss,
|
38 |
+
PolicyValueJointLoss,
|
39 |
+
ActionDiversityReward,
|
40 |
+
ExpectedThoughtValueLoss,
|
41 |
+
ExplorationRegularization,
|
42 |
+
KL_DivergenceLoss,
|
43 |
+
ActionEncoder,
|
44 |
+
RepresentationNetwork,
|
45 |
+
DynamicsNetwork,
|
46 |
+
PredictionNetwork,
|
47 |
+
ThoughtNode,
|
48 |
+
MCTS,
|
49 |
+
State
|
50 |
+
)
|
51 |
+
|
52 |
+
# ==========================
|
53 |
+
# Custom Dataset Definition
|
54 |
+
# ==========================
|
55 |
+
class CustomDataset(Dataset):
|
56 |
+
def __init__(self, inputs, labels):
|
57 |
+
self.inputs = inputs
|
58 |
+
self.labels = labels
|
59 |
+
|
60 |
+
def __len__(self):
|
61 |
+
return len(self.inputs)
|
62 |
+
|
63 |
+
def __getitem__(self, idx):
|
64 |
+
return {'input_ids': self.inputs[idx], 'labels': self.labels[idx]}
|
65 |
+
|
66 |
+
# ================================
|
67 |
+
# Utility Functions for Data Loading
|
68 |
+
# ================================
|
69 |
+
def load_filtered_dataset(dataset_name: str, config: str, queries: Optional[List[str]] = None):
|
70 |
+
dataset = load_dataset(dataset_name, config)
|
71 |
+
if queries:
|
72 |
+
def filter_func(examples):
|
73 |
+
return [any(query.lower() in text.lower() for query in queries) for text in examples["text"]]
|
74 |
+
dataset = dataset.filter(filter_func, batched=True)
|
75 |
+
return dataset
|
76 |
+
|
77 |
+
def load_custom_data_from_files(file_paths):
|
78 |
+
custom_data = []
|
79 |
+
for file_path in file_paths:
|
80 |
+
if file_path.endswith('.json'):
|
81 |
+
with open(file_path, 'r') as f:
|
82 |
+
data = json.load(f)
|
83 |
+
if isinstance(data, list):
|
84 |
+
custom_data.extend(data)
|
85 |
+
else:
|
86 |
+
custom_data.append(data)
|
87 |
+
elif file_path.endswith('.jsonl'):
|
88 |
+
with jsonlines.open(file_path) as reader:
|
89 |
+
custom_data.extend(reader)
|
90 |
+
return custom_data
|
91 |
+
|
92 |
+
def preprocess_custom_data(data_list):
|
93 |
+
processed_data = []
|
94 |
+
for item in data_list:
|
95 |
+
# Check if the item is a string (JSON)
|
96 |
+
if isinstance(item, str):
|
97 |
+
try:
|
98 |
+
item = json.loads(item)
|
99 |
+
except json.JSONDecodeError:
|
100 |
+
print(f"Failed to parse JSON: {item[:100]}...") # Print first 100 chars for debugging
|
101 |
+
continue # Skip this item if it's not valid JSON
|
102 |
+
|
103 |
+
# Process query and content
|
104 |
+
query = item.get('query', '')
|
105 |
+
content = item.get('content', '')
|
106 |
+
if content == "RAG response generation failed.":
|
107 |
+
content = ""
|
108 |
+
|
109 |
+
# Combine query and content
|
110 |
+
combined_text = f"Query: {query} Content: {content}"
|
111 |
+
|
112 |
+
# Process numerical data (assuming these are available in the item dict)
|
113 |
+
episode_reward = item.get('episode_reward', 0)
|
114 |
+
loss = item.get('loss', 0)
|
115 |
+
cosine_similarity = item.get('cosine_similarity', 0)
|
116 |
+
rag_performance = item.get('rag_performance', 0)
|
117 |
+
ranking_model_performance = item.get('ranking_model_performance', 0)
|
118 |
+
|
119 |
+
# Create a dictionary with processed data
|
120 |
+
processed_item = {
|
121 |
+
'text': combined_text,
|
122 |
+
'episode_reward': episode_reward,
|
123 |
+
'loss': loss,
|
124 |
+
'cosine_similarity': cosine_similarity,
|
125 |
+
'rag_performance': rag_performance,
|
126 |
+
'ranking_model_performance': ranking_model_performance
|
127 |
+
}
|
128 |
+
|
129 |
+
processed_data.append(processed_item)
|
130 |
+
|
131 |
+
return processed_data
|
132 |
+
|
133 |
+
def load_custom_data(args, tokenizer, custom_data):
|
134 |
+
# Preprocess the custom data
|
135 |
+
processed_data = preprocess_custom_data(custom_data)
|
136 |
+
|
137 |
+
# Create a custom dataset
|
138 |
+
class CustomDatasetProcessed(torch.utils.data.Dataset):
|
139 |
+
def __init__(self, data, tokenizer, max_length):
|
140 |
+
self.data = data
|
141 |
+
self.tokenizer = tokenizer
|
142 |
+
self.max_length = max_length
|
143 |
+
|
144 |
+
def __len__(self):
|
145 |
+
return len(self.data)
|
146 |
+
|
147 |
+
def __getitem__(self, idx):
|
148 |
+
item = self.data[idx]
|
149 |
+
encoded = self.tokenizer.encode_plus(
|
150 |
+
item['text'],
|
151 |
+
max_length=self.max_length,
|
152 |
+
padding='max_length',
|
153 |
+
truncation=True,
|
154 |
+
return_tensors='pt'
|
155 |
+
)
|
156 |
+
return {
|
157 |
+
'input_ids': encoded['input_ids'].squeeze(),
|
158 |
+
'attention_mask': encoded['attention_mask'].squeeze(),
|
159 |
+
'episode_reward': torch.tensor(item['episode_reward'], dtype=torch.float),
|
160 |
+
'loss': torch.tensor(item['loss'], dtype=torch.float),
|
161 |
+
'cosine_similarity': torch.tensor(item['cosine_similarity'], dtype=torch.float),
|
162 |
+
'rag_performance': torch.tensor(item['rag_performance'], dtype=torch.float),
|
163 |
+
'ranking_model_performance': torch.tensor(item['ranking_model_performance'], dtype=torch.float)
|
164 |
+
}
|
165 |
+
|
166 |
+
# Create dataset and dataloader
|
167 |
+
dataset = CustomDatasetProcessed(processed_data, tokenizer, args.max_length)
|
168 |
+
|
169 |
+
# Split the dataset into train and eval
|
170 |
+
train_size = int(0.8 * len(dataset))
|
171 |
+
eval_size = len(dataset) - train_size
|
172 |
+
train_dataset, eval_dataset = random_split(dataset, [train_size, eval_size])
|
173 |
+
|
174 |
+
train_loader = DataLoader(
|
175 |
+
train_dataset,
|
176 |
+
batch_size=args.batch_size,
|
177 |
+
shuffle=True,
|
178 |
+
num_workers=4
|
179 |
+
)
|
180 |
+
eval_loader = DataLoader(
|
181 |
+
eval_dataset,
|
182 |
+
batch_size=args.batch_size,
|
183 |
+
shuffle=False,
|
184 |
+
num_workers=4
|
185 |
+
)
|
186 |
+
|
187 |
+
return train_loader, eval_loader
|
188 |
+
|
189 |
+
def prepare_data(tokenizer, dataset, max_length, batch_size):
|
190 |
+
# Tokenize the inputs and labels
|
191 |
+
tokenized_inputs = tokenizer(dataset["train"]["text"], return_tensors="pt", padding=True, truncation=True, max_length=max_length)
|
192 |
+
tokenized_labels = tokenizer(dataset["train"]["text"], return_tensors="pt", padding=True, truncation=True, max_length=max_length)
|
193 |
+
|
194 |
+
# Create custom dataset
|
195 |
+
custom_dataset = CustomDataset(tokenized_inputs["input_ids"], tokenized_labels["input_ids"])
|
196 |
+
|
197 |
+
# Split into training and validation sets
|
198 |
+
train_size = int(0.9 * len(custom_dataset))
|
199 |
+
val_size = len(custom_dataset) - train_size
|
200 |
+
train_dataset, val_dataset = random_split(custom_dataset, [train_size, val_size])
|
201 |
+
|
202 |
+
# Create DataLoaders
|
203 |
+
train_loader = DataLoader(
|
204 |
+
train_dataset,
|
205 |
+
shuffle=True,
|
206 |
+
batch_size=batch_size,
|
207 |
+
num_workers=4,
|
208 |
+
pin_memory=True
|
209 |
+
)
|
210 |
+
val_loader = DataLoader(
|
211 |
+
val_dataset,
|
212 |
+
shuffle=False,
|
213 |
+
batch_size=batch_size,
|
214 |
+
num_workers=4,
|
215 |
+
pin_memory=True
|
216 |
+
)
|
217 |
+
|
218 |
+
return train_loader, val_loader
|
219 |
+
|
220 |
+
# ==========================
|
221 |
+
# Training and Validation Functions
|
222 |
+
# ==========================
|
223 |
+
|
224 |
+
def save_all_models(transformer_model, representation_network, dynamics_network, prediction_network, action_encoder, save_dir, epoch):
|
225 |
+
"""
|
226 |
+
Save all models to the specified directory.
|
227 |
+
Args:
|
228 |
+
transformer_model (nn.Module): Transformer model.
|
229 |
+
representation_network (nn.Module): Representation network.
|
230 |
+
dynamics_network (nn.Module): Dynamics network.
|
231 |
+
prediction_network (nn.Module): Prediction network.
|
232 |
+
action_encoder (nn.Module): Action encoder.
|
233 |
+
save_dir (str): Directory to save the models.
|
234 |
+
epoch (int): Current epoch number.
|
235 |
+
"""
|
236 |
+
os.makedirs(save_dir, exist_ok=True)
|
237 |
+
|
238 |
+
torch.save(transformer_model.state_dict(), os.path.join(save_dir, f'transformer_model_epoch_{epoch}.pt'))
|
239 |
+
torch.save(representation_network.state_dict(), os.path.join(save_dir, f'representation_network_epoch_{epoch}.pt'))
|
240 |
+
torch.save(dynamics_network.state_dict(), os.path.join(save_dir, f'dynamics_network_epoch_{epoch}.pt'))
|
241 |
+
torch.save(prediction_network.state_dict(), os.path.join(save_dir, f'prediction_network_epoch_{epoch}.pt'))
|
242 |
+
torch.save(action_encoder.state_dict(), os.path.join(save_dir, f'action_encoder_epoch_{epoch}.pt'))
|
243 |
+
|
244 |
+
print(f"All models saved for epoch {epoch}.")
|
245 |
+
|
246 |
+
def train_epoch_world_model(world_model_components, train_loader, optimizer, scheduler, scaler, args, model_transformer, state_dim, embed_dim, input_dim):
|
247 |
+
representation_network, dynamics_network, prediction_network, action_encoder, ppo_agent, model_transformer = world_model_components
|
248 |
+
representation_network.train()
|
249 |
+
dynamics_network.train()
|
250 |
+
prediction_network.train()
|
251 |
+
action_encoder.train()
|
252 |
+
ppo_agent.policy_network.train()
|
253 |
+
|
254 |
+
total_loss = 0.0
|
255 |
+
optimizer.zero_grad()
|
256 |
+
print(f"Starting World Model training epoch with {len(train_loader)} batches...")
|
257 |
+
|
258 |
+
for i, batch in enumerate(train_loader):
|
259 |
+
print(f"Processing batch {i+1}/{len(train_loader)}...")
|
260 |
+
|
261 |
+
# Move batches to the device
|
262 |
+
src_batch = batch['input_ids'].to(device)
|
263 |
+
tgt_batch = batch['labels'].to(device)
|
264 |
+
|
265 |
+
with torch.cuda.amp.autocast():
|
266 |
+
print("Forward pass through Transformer (frozen)...")
|
267 |
+
with torch.no_grad():
|
268 |
+
transformer_output = model_transformer(src_batch, tgt_batch[:, :-1])
|
269 |
+
|
270 |
+
# World Model - Representation
|
271 |
+
state_representation = representation_network(transformer_output)
|
272 |
+
|
273 |
+
# For simplicity, let's assume true actions are provided (e.g., next tokens)
|
274 |
+
true_actions = tgt_batch[:, :-1]
|
275 |
+
print(f"True actions shape: {true_actions.shape}")
|
276 |
+
action_sequences = true_actions
|
277 |
+
|
278 |
+
# Get action embeddings
|
279 |
+
action_embeddings = action_encoder(action_sequences)
|
280 |
+
print(f"Action embeddings shape: {action_embeddings.shape}")
|
281 |
+
|
282 |
+
# Apply dynamics network
|
283 |
+
predicted_next_state_batch = dynamics_network(state_representation, action_embeddings)
|
284 |
+
print(f"Predicted next state batch shape: {predicted_next_state_batch.shape}")
|
285 |
+
|
286 |
+
# Prediction Network - Policy logits and value
|
287 |
+
policy_logits, value_estimates = prediction_network(predicted_next_state_batch)
|
288 |
+
|
289 |
+
# Define true_policy and true_value as placeholders on the GPU
|
290 |
+
true_policy = F.one_hot(true_actions, num_classes=input_dim).float()
|
291 |
+
true_value = torch.zeros_like(value_estimates).to(device)
|
292 |
+
|
293 |
+
# Compute individual losses
|
294 |
+
ppo_loss = ppo_agent.compute_loss(
|
295 |
+
state_representation,
|
296 |
+
torch.zeros_like(true_actions, dtype=torch.float32).to(device),
|
297 |
+
true_actions,
|
298 |
+
torch.zeros_like(value_estimates, dtype=torch.float32).to(device),
|
299 |
+
torch.zeros_like(value_estimates, dtype=torch.float32).to(device)
|
300 |
+
)
|
301 |
+
|
302 |
+
info_nce = InfoNCE_Loss()(state_representation.reshape(-1, state_dim),
|
303 |
+
F.dropout(state_representation.reshape(-1, state_dim), p=0.1, training=True))
|
304 |
+
|
305 |
+
covariance = CovarianceRegularization()(predicted_next_state_batch.view(-1, predicted_next_state_batch.size(-1)))
|
306 |
+
dynamics_loss = DynamicsPerformanceLoss()(state_representation, predicted_next_state_batch)
|
307 |
+
|
308 |
+
perturbed_next_state = predicted_next_state_batch + torch.randn_like(predicted_next_state_batch) * 0.01
|
309 |
+
thought_loss = ThoughtConsistencyLoss()(predicted_next_state_batch, perturbed_next_state)
|
310 |
+
|
311 |
+
pv_loss = PolicyValueJointLoss()(policy_logits, true_policy, value_estimates.squeeze(-1), true_value.squeeze(-1))
|
312 |
+
action_diversity = ActionDiversityReward()(action_embeddings.view(-1, embed_dim))
|
313 |
+
|
314 |
+
mcts_best_values = torch.zeros(true_actions.size(0)).to(device)
|
315 |
+
etv = ExpectedThoughtValueLoss()(mcts_best_values)
|
316 |
+
|
317 |
+
visit_counts = torch.ones(true_actions.size(0), policy_logits.size(-1)).to(device)
|
318 |
+
exploration = ExplorationRegularization()(visit_counts)
|
319 |
+
|
320 |
+
old_policy = F.softmax(policy_logits.detach(), dim=-1)
|
321 |
+
new_policy = F.softmax(policy_logits, dim=-1)
|
322 |
+
kl_loss = KL_DivergenceLoss()(old_policy, new_policy)
|
323 |
+
|
324 |
+
# Total Loss
|
325 |
+
loss = (
|
326 |
+
ppo_loss +
|
327 |
+
info_nce +
|
328 |
+
covariance +
|
329 |
+
dynamics_loss +
|
330 |
+
thought_loss +
|
331 |
+
pv_loss +
|
332 |
+
action_diversity +
|
333 |
+
etv +
|
334 |
+
exploration +
|
335 |
+
kl_loss
|
336 |
+
)
|
337 |
+
loss = loss / args.accumulation_steps
|
338 |
+
|
339 |
+
print("Backward pass...")
|
340 |
+
scaler.scale(loss).backward()
|
341 |
+
|
342 |
+
if (i + 1) % args.accumulation_steps == 0 or (i + 1) == len(train_loader):
|
343 |
+
print("Gradient clipping...")
|
344 |
+
scaler.unscale_(optimizer)
|
345 |
+
torch.nn.utils.clip_grad_norm_(
|
346 |
+
[param for group in optimizer.param_groups for param in group['params']],
|
347 |
+
args.max_grad_norm
|
348 |
+
)
|
349 |
+
|
350 |
+
print("Optimizer step...")
|
351 |
+
scaler.step(optimizer)
|
352 |
+
scaler.update()
|
353 |
+
|
354 |
+
print("Zeroing gradients...")
|
355 |
+
optimizer.zero_grad()
|
356 |
+
|
357 |
+
print("Updating learning rate...")
|
358 |
+
scheduler.step()
|
359 |
+
|
360 |
+
total_loss += loss.item() * args.accumulation_steps
|
361 |
+
|
362 |
+
# Print individual losses and total loss for this batch
|
363 |
+
print(f"Batch {i+1} completed. Losses:")
|
364 |
+
print(f" PPO Loss: {ppo_loss.item():.4f}")
|
365 |
+
print(f" InfoNCE Loss: {info_nce.item():.4f}")
|
366 |
+
print(f" Covariance Loss: {covariance.item():.4f}")
|
367 |
+
print(f" Dynamics Loss: {dynamics_loss.item():.4f}")
|
368 |
+
print(f" Thought Consistency Loss: {thought_loss.item():.4f}")
|
369 |
+
print(f" Policy-Value Loss: {pv_loss.item():.4f}")
|
370 |
+
print(f" Action Diversity Loss: {action_diversity.item():.4f}")
|
371 |
+
print(f" Expected Thought Value Loss: {etv.item():.4f}")
|
372 |
+
print(f" Exploration Loss: {exploration.item():.4f}")
|
373 |
+
print(f" KL Divergence Loss: {kl_loss.item():.4f}")
|
374 |
+
print(f" Total Loss: {loss.item():.4f}")
|
375 |
+
|
376 |
+
avg_loss = total_loss / len(train_loader)
|
377 |
+
print(f"World Model training epoch completed. Average loss: {avg_loss:.4f}")
|
378 |
+
return avg_loss
|
379 |
+
|
380 |
+
def train_step(teacher, student, data_loader, optimizer, criterion, scaler, temperature=2.0):
|
381 |
+
teacher.eval()
|
382 |
+
student.train()
|
383 |
+
total_loss = 0
|
384 |
+
|
385 |
+
for batch in tqdm(data_loader, desc="Training"):
|
386 |
+
inputs = batch["input_ids"].to(device)
|
387 |
+
labels = batch["labels"].to(device)
|
388 |
+
|
389 |
+
with autocast():
|
390 |
+
with torch.no_grad():
|
391 |
+
teacher_outputs = teacher(inputs).logits
|
392 |
+
teacher_logits = teacher_outputs / temperature
|
393 |
+
|
394 |
+
student_outputs = student(inputs).logits
|
395 |
+
student_logits = student_outputs / temperature
|
396 |
+
|
397 |
+
# Compute KL Divergence Loss
|
398 |
+
loss = criterion(nn.functional.log_softmax(student_logits, dim=-1), nn.functional.softmax(teacher_logits, dim=-1))
|
399 |
+
loss = loss * (temperature ** 2) # Scale loss by temperature squared
|
400 |
+
|
401 |
+
scaler.scale(loss).backward()
|
402 |
+
scaler.step(optimizer)
|
403 |
+
scaler.update()
|
404 |
+
optimizer.zero_grad()
|
405 |
+
|
406 |
+
total_loss += loss.item()
|
407 |
+
|
408 |
+
avg_loss = total_loss / len(data_loader)
|
409 |
+
return avg_loss
|
410 |
+
|
411 |
+
def validate(teacher, student, data_loader, criterion, temperature=2.0):
|
412 |
+
teacher.eval()
|
413 |
+
student.eval()
|
414 |
+
total_loss = 0
|
415 |
+
|
416 |
+
with torch.no_grad():
|
417 |
+
for batch in tqdm(data_loader, desc="Validation"):
|
418 |
+
inputs = batch["input_ids"].to(device)
|
419 |
+
labels = batch["labels"].to(device)
|
420 |
+
|
421 |
+
teacher_outputs = teacher(inputs).logits
|
422 |
+
teacher_logits = teacher_outputs / temperature
|
423 |
+
|
424 |
+
student_outputs = student(inputs).logits
|
425 |
+
student_logits = student_outputs / temperature
|
426 |
+
|
427 |
+
loss = criterion(nn.functional.log_softmax(student_logits, dim=-1), nn.functional.softmax(teacher_logits, dim=-1))
|
428 |
+
loss = loss * (temperature ** 2)
|
429 |
+
|
430 |
+
total_loss += loss.item()
|
431 |
+
|
432 |
+
avg_loss = total_loss / len(data_loader)
|
433 |
+
return avg_loss
|
434 |
+
|
435 |
+
def save_checkpoint(state, save_dir, epoch):
|
436 |
+
os.makedirs(save_dir, exist_ok=True)
|
437 |
+
checkpoint_path = os.path.join(save_dir, f'checkpoint_epoch_{epoch}.pt')
|
438 |
+
torch.save(state, checkpoint_path)
|
439 |
+
print(f"Checkpoint saved at {checkpoint_path}")
|
440 |
+
|
441 |
+
# ==========================
|
442 |
+
# Inference Functions
|
443 |
+
# ==========================
|
444 |
+
|
445 |
+
def infer(query, world_model_components, root_thought_node, tokenizer, max_length=2000, inference_mode='world_model', beam_size=5, n_tokens_predict=3, mcts_iterations=10, exploration_constant=1.414):
|
446 |
+
"""
|
447 |
+
Perform inference given a query, utilizing the Tree of Thought and MCTS with multi-token beam search.
|
448 |
+
Args:
|
449 |
+
query (str): The input query or prompt.
|
450 |
+
world_model_components (tuple): Tuple containing the model components.
|
451 |
+
root_thought_node (ThoughtNode): The root node of the Tree of Thought.
|
452 |
+
tokenizer (transformers.PreTrainedTokenizer): The tokenizer used.
|
453 |
+
max_length (int): Maximum length for the generated sequence.
|
454 |
+
inference_mode (str): Inference mode ('world_model', 'without_world_model', 'world_model_tree_of_thought')
|
455 |
+
beam_size (int): Size of the beam for beam search
|
456 |
+
n_tokens_predict (int): Number of tokens to predict at each step
|
457 |
+
mcts_iterations (int): Number of MCTS iterations
|
458 |
+
exploration_constant (float): Exploration constant for MCTS
|
459 |
+
Returns:
|
460 |
+
List[str] or str: The sequence of actions (thoughts) selected or generated text.
|
461 |
+
"""
|
462 |
+
if inference_mode != 'world_model':
|
463 |
+
print("Inference mode other than 'world_model' not implemented yet.")
|
464 |
+
return ""
|
465 |
+
|
466 |
+
representation_network, dynamics_network, prediction_network, action_encoder, ppo_agent, model_transformer = world_model_components
|
467 |
+
|
468 |
+
# Tokenize and encode the query
|
469 |
+
input_ids = tokenizer.encode(query, return_tensors='pt').to(device)
|
470 |
+
attention_mask = (input_ids != tokenizer.pad_token_id).long()
|
471 |
+
|
472 |
+
# Use the world model components
|
473 |
+
with torch.no_grad():
|
474 |
+
transformer_output = model_transformer(input_ids, input_ids)
|
475 |
+
# Get the initial state representation
|
476 |
+
initial_representation = representation_network(transformer_output) # Shape: (batch_size=1, seq_len, state_dim)
|
477 |
+
initial_representation = initial_representation[:, -1, :].unsqueeze(1) # Shape: (batch_size=1, 1, state_dim)
|
478 |
+
initial_state = State(
|
479 |
+
representation=initial_representation,
|
480 |
+
dynamics_network=dynamics_network,
|
481 |
+
action_encoder=action_encoder,
|
482 |
+
thought_node=root_thought_node
|
483 |
+
)
|
484 |
+
# Use MCTS with Tree of Thought and multi-token beam search
|
485 |
+
mcts = MCTS(prediction_network, dynamics_network, action_encoder, num_iterations=mcts_iterations, exploration_constant=exploration_constant)
|
486 |
+
|
487 |
+
current_state = initial_state
|
488 |
+
thought_sequence = []
|
489 |
+
|
490 |
+
for _ in range(max_length // n_tokens_predict):
|
491 |
+
best_actions = mcts.search_with_beam(current_state)
|
492 |
+
|
493 |
+
thought_sequence.extend(best_actions)
|
494 |
+
|
495 |
+
# Apply the best actions to get the next state
|
496 |
+
for action in best_actions:
|
497 |
+
current_state = current_state.apply_action(action)
|
498 |
+
|
499 |
+
# Check if we've reached a leaf node (no further actions)
|
500 |
+
if len(current_state.thought_node.children) == 0:
|
501 |
+
break
|
502 |
+
|
503 |
+
return thought_sequence
|
504 |
+
|
505 |
+
# ==========================
|
506 |
+
# Main Training Function
|
507 |
+
# ==========================
|
508 |
+
|
509 |
+
def distill_model(
|
510 |
+
teacher_model_name: str,
|
511 |
+
student_model_name: str,
|
512 |
+
dataset_name: str,
|
513 |
+
config: str,
|
514 |
+
distill_full_model: bool = True,
|
515 |
+
query_terms: Optional[List[str]] = None,
|
516 |
+
num_epochs: int = 3,
|
517 |
+
batch_size: int = 4,
|
518 |
+
max_length: int = 128,
|
519 |
+
learning_rate: float = 5e-5,
|
520 |
+
temperature: float = 2.0,
|
521 |
+
save_path: str = "./distilled_model",
|
522 |
+
log_dir: str = "./logs",
|
523 |
+
checkpoint_dir: str = "./checkpoints",
|
524 |
+
early_stopping_patience: int = 3,
|
525 |
+
accumulation_steps: int = 1,
|
526 |
+
max_grad_norm: float = 1.0,
|
527 |
+
weight_decay: float = 0.01
|
528 |
+
):
|
529 |
+
# Initialize TensorBoard writer
|
530 |
+
writer = SummaryWriter(log_dir=log_dir)
|
531 |
+
|
532 |
+
# Load tokenizer
|
533 |
+
print("Loading tokenizer...")
|
534 |
+
tokenizer = AutoTokenizer.from_pretrained(teacher_model_name)
|
535 |
+
if tokenizer.pad_token is None:
|
536 |
+
tokenizer.pad_token = tokenizer.eos_token
|
537 |
+
print("Tokenizer loaded successfully.")
|
538 |
+
|
539 |
+
# Load teacher model
|
540 |
+
print("Loading teacher model...")
|
541 |
+
teacher = AutoModelForCausalLM.from_pretrained(teacher_model_name).to(device)
|
542 |
+
print("Teacher model loaded successfully.")
|
543 |
+
|
544 |
+
if distill_full_model:
|
545 |
+
# Full World Model Distillation
|
546 |
+
print(f"Starting Full World Model Distillation into '{student_model_name}'.")
|
547 |
+
|
548 |
+
# Load or instantiate student model
|
549 |
+
print(f"Attempting to load student model '{student_model_name}'...")
|
550 |
+
try:
|
551 |
+
student = AutoModelForCausalLM.from_pretrained(student_model_name).to(device)
|
552 |
+
print(f"Student model '{student_model_name}' loaded successfully.")
|
553 |
+
except (OSError, ValueError) as e:
|
554 |
+
print(f"Student model '{student_model_name}' not found. Instantiating a new student model.")
|
555 |
+
# Instantiate a smaller pre-trained model as the student, e.g., distilgpt2
|
556 |
+
try:
|
557 |
+
student = AutoModelForCausalLM.from_pretrained('distilgpt2').to(device)
|
558 |
+
# Save the instantiated student model with the desired name
|
559 |
+
student.save_pretrained(save_path)
|
560 |
+
tokenizer.save_pretrained(save_path)
|
561 |
+
print(f"New student model '{student_model_name}' instantiated and saved to '{save_path}'.")
|
562 |
+
except Exception as inst_e:
|
563 |
+
print(f"Failed to instantiate and save student model: {inst_e}")
|
564 |
+
sys.exit(1)
|
565 |
+
|
566 |
+
# Optionally freeze teacher model parameters
|
567 |
+
for param in teacher.parameters():
|
568 |
+
param.requires_grad = False
|
569 |
+
|
570 |
+
# Load and prepare dataset
|
571 |
+
print(f"Loading full dataset '{dataset_name}' with config '{config}'...")
|
572 |
+
dataset = load_dataset(dataset_name, config)
|
573 |
+
train_loader, val_loader = prepare_data(tokenizer, dataset, max_length, batch_size)
|
574 |
+
print("Data loaded and preprocessed successfully.")
|
575 |
+
|
576 |
+
# Define optimizer, scheduler, and scaler for mixed precision
|
577 |
+
optimizer = optim.AdamW(student.parameters(), lr=learning_rate, weight_decay=weight_decay)
|
578 |
+
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_epochs)
|
579 |
+
scaler = GradScaler()
|
580 |
+
|
581 |
+
# Define loss criterion
|
582 |
+
criterion = nn.KLDivLoss(reduction="batchmean")
|
583 |
+
|
584 |
+
best_val_loss = float('inf')
|
585 |
+
epochs_no_improve = 0
|
586 |
+
|
587 |
+
# Training loop
|
588 |
+
for epoch in range(1, num_epochs + 1):
|
589 |
+
print(f"\nEpoch {epoch}/{num_epochs}")
|
590 |
+
print("-" * 20)
|
591 |
+
|
592 |
+
# Training
|
593 |
+
train_loss = train_step(teacher, student, train_loader, optimizer, criterion, scaler, temperature)
|
594 |
+
print(f"Training Loss: {train_loss:.4f}")
|
595 |
+
writer.add_scalar("Loss/Train", train_loss, epoch)
|
596 |
+
|
597 |
+
# Validation
|
598 |
+
val_loss = validate(teacher, student, val_loader, criterion, temperature)
|
599 |
+
print(f"Validation Loss: {val_loss:.4f}")
|
600 |
+
writer.add_scalar("Loss/Validation", val_loss, epoch)
|
601 |
+
|
602 |
+
# Check for improvement
|
603 |
+
if val_loss < best_val_loss:
|
604 |
+
best_val_loss = val_loss
|
605 |
+
epochs_no_improve = 0
|
606 |
+
# Save the best model
|
607 |
+
save_checkpoint({
|
608 |
+
'epoch': epoch,
|
609 |
+
'model_state_dict': student.state_dict(),
|
610 |
+
'optimizer_state_dict': optimizer.state_dict(),
|
611 |
+
'scheduler_state_dict': scheduler.state_dict(),
|
612 |
+
'scaler_state_dict': scaler.state_dict(),
|
613 |
+
'best_val_loss': best_val_loss
|
614 |
+
}, checkpoint_dir, epoch)
|
615 |
+
# Save the model as the best one
|
616 |
+
student.save_pretrained(save_path)
|
617 |
+
tokenizer.save_pretrained(save_path)
|
618 |
+
print(f"Best model saved at epoch {epoch}")
|
619 |
+
else:
|
620 |
+
epochs_no_improve += 1
|
621 |
+
print(f"No improvement in validation loss for {epochs_no_improve} epoch(s)")
|
622 |
+
if epochs_no_improve >= early_stopping_patience:
|
623 |
+
print("Early stopping triggered")
|
624 |
+
break
|
625 |
+
|
626 |
+
# Step the scheduler
|
627 |
+
scheduler.step()
|
628 |
+
|
629 |
+
writer.close()
|
630 |
+
print("\nFull World Model Distillation completed.")
|
631 |
+
|
632 |
+
else:
|
633 |
+
# Standard Language Model Distillation
|
634 |
+
print(f"Starting Standard Language Model Distillation into '{student_model_name}'.")
|
635 |
+
|
636 |
+
if not query_terms:
|
637 |
+
print("Error: --query_terms must be provided for standard language model distillation.")
|
638 |
+
sys.exit(1)
|
639 |
+
|
640 |
+
# Load or instantiate student model
|
641 |
+
print(f"Attempting to load student model '{student_model_name}'...")
|
642 |
+
try:
|
643 |
+
student = AutoModelForCausalLM.from_pretrained(student_model_name).to(device)
|
644 |
+
print(f"Student model '{student_model_name}' loaded successfully.")
|
645 |
+
except (OSError, ValueError) as e:
|
646 |
+
print(f"Student model '{student_model_name}' not found. Instantiating a new student model.")
|
647 |
+
# Instantiate a smaller pre-trained model as the student, e.g., distilgpt2
|
648 |
+
try:
|
649 |
+
student = AutoModelForCausalLM.from_pretrained('distilgpt2').to(device)
|
650 |
+
# Save the instantiated student model with the desired name
|
651 |
+
student.save_pretrained(save_path)
|
652 |
+
tokenizer.save_pretrained(save_path)
|
653 |
+
print(f"New student model '{student_model_name}' instantiated and saved to '{save_path}'.")
|
654 |
+
except Exception as inst_e:
|
655 |
+
print(f"Failed to instantiate and save student model: {inst_e}")
|
656 |
+
sys.exit(1)
|
657 |
+
|
658 |
+
# Optionally freeze teacher model parameters
|
659 |
+
for param in teacher.parameters():
|
660 |
+
param.requires_grad = False
|
661 |
+
|
662 |
+
# Load and prepare custom dataset
|
663 |
+
print(f"Loading custom data files: {query_terms}")
|
664 |
+
custom_data = load_custom_data_from_files(query_terms)
|
665 |
+
train_loader, val_loader = load_custom_data(
|
666 |
+
args=argparse.Namespace(max_length=max_length),
|
667 |
+
tokenizer=tokenizer,
|
668 |
+
custom_data=custom_data
|
669 |
+
)
|
670 |
+
print("Custom data loaded and preprocessed successfully.")
|
671 |
+
|
672 |
+
# Define optimizer, scheduler, and scaler for mixed precision
|
673 |
+
optimizer = optim.AdamW(student.parameters(), lr=learning_rate, weight_decay=weight_decay)
|
674 |
+
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_epochs)
|
675 |
+
scaler = GradScaler()
|
676 |
+
|
677 |
+
# Define loss criterion
|
678 |
+
criterion = nn.KLDivLoss(reduction="batchmean")
|
679 |
+
|
680 |
+
best_val_loss = float('inf')
|
681 |
+
epochs_no_improve = 0
|
682 |
+
|
683 |
+
# Training loop
|
684 |
+
for epoch in range(1, num_epochs + 1):
|
685 |
+
print(f"\nEpoch {epoch}/{num_epochs}")
|
686 |
+
print("-" * 20)
|
687 |
+
|
688 |
+
# Training
|
689 |
+
train_loss = train_step(teacher, student, train_loader, optimizer, criterion, scaler, temperature)
|
690 |
+
print(f"Training Loss: {train_loss:.4f}")
|
691 |
+
writer.add_scalar("Loss/Train", train_loss, epoch)
|
692 |
+
|
693 |
+
# Validation
|
694 |
+
val_loss = validate(teacher, student, val_loader, criterion, temperature)
|
695 |
+
print(f"Validation Loss: {val_loss:.4f}")
|
696 |
+
writer.add_scalar("Loss/Validation", val_loss, epoch)
|
697 |
+
|
698 |
+
# Check for improvement
|
699 |
+
if val_loss < best_val_loss:
|
700 |
+
best_val_loss = val_loss
|
701 |
+
epochs_no_improve = 0
|
702 |
+
# Save the best model
|
703 |
+
save_checkpoint({
|
704 |
+
'epoch': epoch,
|
705 |
+
'model_state_dict': student.state_dict(),
|
706 |
+
'optimizer_state_dict': optimizer.state_dict(),
|
707 |
+
'scheduler_state_dict': scheduler.state_dict(),
|
708 |
+
'scaler_state_dict': scaler.state_dict(),
|
709 |
+
'best_val_loss': best_val_loss
|
710 |
+
}, checkpoint_dir, epoch)
|
711 |
+
# Save the model as the best one
|
712 |
+
student.save_pretrained(save_path)
|
713 |
+
tokenizer.save_pretrained(save_path)
|
714 |
+
print(f"Best model saved at epoch {epoch}")
|
715 |
+
else:
|
716 |
+
epochs_no_improve += 1
|
717 |
+
print(f"No improvement in validation loss for {epochs_no_improve} epoch(s)")
|
718 |
+
if epochs_no_improve >= early_stopping_patience:
|
719 |
+
print("Early stopping triggered")
|
720 |
+
break
|
721 |
+
|
722 |
+
# Step the scheduler
|
723 |
+
scheduler.step()
|
724 |
+
|
725 |
+
writer.close()
|
726 |
+
print("\nStandard Language Model Distillation completed.")
|
727 |
+
|
728 |
+
# ==========================
|
729 |
+
# Argument Parsing
|
730 |
+
# ==========================
|
731 |
+
|
732 |
+
def parse_args():
|
733 |
+
parser = argparse.ArgumentParser(description="Distill a large LLM into a smaller one or a full language world model.")
|
734 |
+
|
735 |
+
# Required arguments
|
736 |
+
parser.add_argument("--teacher_model_name", type=str, required=True, help="Name of the teacher model")
|
737 |
+
parser.add_argument("--student_model_name", type=str, required=True, help="Name of the student model")
|
738 |
+
|
739 |
+
# Dataset arguments
|
740 |
+
parser.add_argument("--dataset_name", type=str, required=True, help="Name of the dataset")
|
741 |
+
parser.add_argument("--config", type=str, default=None, help="Dataset configuration (e.g., 'wikitext-2-raw-v1')")
|
742 |
+
|
743 |
+
# Mode selection
|
744 |
+
parser.add_argument("--distill_full_model", action="store_true", help="Whether to distill into the full language world model")
|
745 |
+
|
746 |
+
# For standard distillation
|
747 |
+
parser.add_argument("--query_terms", type=str, nargs="+", help="Paths to custom data files for standard language model distillation")
|
748 |
+
|
749 |
+
# Training hyperparameters
|
750 |
+
parser.add_argument("--num_epochs", type=int, default=3, help="Number of epochs")
|
751 |
+
parser.add_argument("--batch_size", type=int, default=4, help="Batch size")
|
752 |
+
parser.add_argument("--max_length", type=int, default=128, help="Maximum sequence length")
|
753 |
+
parser.add_argument("--learning_rate", type=float, default=5e-5, help="Learning rate")
|
754 |
+
parser.add_argument("--temperature", type=float, default=2.0, help="Distillation temperature")
|
755 |
+
|
756 |
+
# Saving and logging
|
757 |
+
parser.add_argument("--save_path", type=str, default="./distilled_model", help="Path to save the distilled model")
|
758 |
+
parser.add_argument("--log_dir", type=str, default="./logs", help="Directory for TensorBoard logs")
|
759 |
+
parser.add_argument("--checkpoint_dir", type=str, default="./checkpoints", help="Directory to save checkpoints")
|
760 |
+
|
761 |
+
# Early stopping
|
762 |
+
parser.add_argument("--early_stopping_patience", type=int, default=3, help="Early stopping patience")
|
763 |
+
|
764 |
+
# Gradient accumulation and optimization
|
765 |
+
parser.add_argument("--accumulation_steps", type=int, default=1, help="Gradient accumulation steps")
|
766 |
+
parser.add_argument("--max_grad_norm", type=float, default=1.0, help="Maximum gradient norm for clipping")
|
767 |
+
parser.add_argument("--weight_decay", type=float, default=0.01, help="Weight decay for optimizer")
|
768 |
+
|
769 |
+
return parser.parse_args()
|
770 |
+
|
771 |
+
# ==========================
|
772 |
+
# Main Function
|
773 |
+
# ==========================
|
774 |
+
|
775 |
+
def main():
|
776 |
+
args = parse_args()
|
777 |
+
print("Arguments parsed successfully.")
|
778 |
+
|
779 |
+
# Create save directories
|
780 |
+
os.makedirs(args.save_path, exist_ok=True)
|
781 |
+
os.makedirs(args.log_dir, exist_ok=True)
|
782 |
+
os.makedirs(args.checkpoint_dir, exist_ok=True)
|
783 |
+
print(f"Save directory created: {args.save_path}")
|
784 |
+
print(f"Log directory created: {args.log_dir}")
|
785 |
+
print(f"Checkpoint directory created: {args.checkpoint_dir}")
|
786 |
+
|
787 |
+
# Handle dataset loading based on distillation mode
|
788 |
+
if args.distill_full_model:
|
789 |
+
# Full World Model Distillation
|
790 |
+
distill_model(
|
791 |
+
teacher_model_name=args.teacher_model_name,
|
792 |
+
student_model_name=args.student_model_name,
|
793 |
+
dataset_name=args.dataset_name,
|
794 |
+
config=args.config,
|
795 |
+
distill_full_model=args.distill_full_model,
|
796 |
+
query_terms=args.query_terms, # Not used in this mode
|
797 |
+
num_epochs=args.num_epochs,
|
798 |
+
batch_size=args.batch_size,
|
799 |
+
max_length=args.max_length,
|
800 |
+
learning_rate=args.learning_rate,
|
801 |
+
temperature=args.temperature,
|
802 |
+
save_path=args.save_path,
|
803 |
+
log_dir=args.log_dir,
|
804 |
+
checkpoint_dir=args.checkpoint_dir,
|
805 |
+
early_stopping_patience=args.early_stopping_patience,
|
806 |
+
accumulation_steps=args.accumulation_steps,
|
807 |
+
max_grad_norm=args.max_grad_norm,
|
808 |
+
weight_decay=args.weight_decay
|
809 |
+
)
|
810 |
+
else:
|
811 |
+
# Standard Language Model Distillation
|
812 |
+
distill_model(
|
813 |
+
teacher_model_name=args.teacher_model_name,
|
814 |
+
student_model_name=args.student_model_name,
|
815 |
+
dataset_name=args.dataset_name,
|
816 |
+
config=args.config,
|
817 |
+
distill_full_model=args.distill_full_model,
|
818 |
+
query_terms=args.query_terms,
|
819 |
+
num_epochs=args.num_epochs,
|
820 |
+
batch_size=args.batch_size,
|
821 |
+
max_length=args.max_length,
|
822 |
+
learning_rate=args.learning_rate,
|
823 |
+
temperature=args.temperature,
|
824 |
+
save_path=args.save_path,
|
825 |
+
log_dir=args.log_dir,
|
826 |
+
checkpoint_dir=args.checkpoint_dir,
|
827 |
+
early_stopping_patience=args.early_stopping_patience,
|
828 |
+
accumulation_steps=args.accumulation_steps,
|
829 |
+
max_grad_norm=args.max_grad_norm,
|
830 |
+
weight_decay=args.weight_decay
|
831 |
+
)
|
832 |
+
|
833 |
+
|
834 |
+
|
835 |
+
if __name__ == "__main__":
|
836 |
+
main()
|
837 |
+
|
838 |
+
|