Upload sweep9.py with huggingface_hub
Browse files
sweep9.py
ADDED
@@ -0,0 +1,388 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import transformers
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
import torch.optim as optim
|
6 |
+
from torch.utils.data import DataLoader
|
7 |
+
import numpy as np
|
8 |
+
import matplotlib.pyplot as plt
|
9 |
+
import json
|
10 |
+
from tqdm.auto import tqdm
|
11 |
+
import random
|
12 |
+
from scipy.signal import savgol_filter
|
13 |
+
import wandb
|
14 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
15 |
+
import os
|
16 |
+
from itertools import product
|
17 |
+
import pandas as pd
|
18 |
+
import multiprocessing as mp
|
19 |
+
from functools import partial
|
20 |
+
import logging
|
21 |
+
|
22 |
+
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
|
23 |
+
logger = logging.getLogger(__name__)
|
24 |
+
|
25 |
+
class Config:
|
26 |
+
def __init__(self):
|
27 |
+
self.model_name = "Qwen/Qwen2-0.5B"
|
28 |
+
self.data_dir = "dataset_chunks"
|
29 |
+
self.max_length = 1024
|
30 |
+
self.batch_size = 32
|
31 |
+
self.num_seeds = 10000
|
32 |
+
self.num_lr_steps = 10000
|
33 |
+
self.min_lr = 1e-8
|
34 |
+
self.max_lr = 10
|
35 |
+
self.hidden_dim_ratio = 0.5
|
36 |
+
self.dropout = 0.1
|
37 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
38 |
+
self.num_workers = mp.cpu_count()
|
39 |
+
|
40 |
+
class ImprovedAutoencoder(nn.Module):
|
41 |
+
def __init__(self, input_dim, hidden_dim, num_layers, dropout):
|
42 |
+
super().__init__()
|
43 |
+
self.encoder = nn.ModuleList([
|
44 |
+
nn.Linear(input_dim if i == 0 else hidden_dim, hidden_dim, dtype=torch.bfloat16)
|
45 |
+
for i in range(num_layers)
|
46 |
+
])
|
47 |
+
self.decoder = nn.ModuleList([
|
48 |
+
nn.Linear(hidden_dim, hidden_dim if i < num_layers - 1 else input_dim, dtype=torch.bfloat16)
|
49 |
+
for i in range(num_layers)
|
50 |
+
])
|
51 |
+
self.layer_norms = nn.ModuleList([
|
52 |
+
nn.LayerNorm(hidden_dim, dtype=torch.bfloat16)
|
53 |
+
for _ in range(num_layers * 2 - 1)
|
54 |
+
])
|
55 |
+
self.dropout = nn.Dropout(dropout)
|
56 |
+
|
57 |
+
def forward(self, x):
|
58 |
+
for enc, norm in zip(self.encoder, self.layer_norms[:len(self.encoder)]):
|
59 |
+
x = F.relu(norm(enc(x)))
|
60 |
+
x = self.dropout(x)
|
61 |
+
for dec, norm in zip(self.decoder[:-1], self.layer_norms[len(self.encoder):]):
|
62 |
+
x = F.relu(norm(dec(x)))
|
63 |
+
x = self.dropout(x)
|
64 |
+
x = self.decoder[-1](x)
|
65 |
+
return x
|
66 |
+
|
67 |
+
class TokenizedDataset(torch.utils.data.Dataset):
|
68 |
+
def __init__(self, file_paths):
|
69 |
+
self.data = []
|
70 |
+
for file_path in tqdm(file_paths, desc="Loading data chunks"):
|
71 |
+
chunk_data = torch.load(file_path)
|
72 |
+
logger.info(f"Loaded data from {file_path}")
|
73 |
+
logger.info(f"Type of loaded data: {type(chunk_data)}")
|
74 |
+
|
75 |
+
if isinstance(chunk_data, dict): # Handle dictionary format (if present)
|
76 |
+
logger.info(f"Keys in the dictionary: {chunk_data.keys()}")
|
77 |
+
logger.info(f"Shape of input_ids: {chunk_data['input_ids'].shape}")
|
78 |
+
self.data.append(chunk_data)
|
79 |
+
elif isinstance(chunk_data, transformers.tokenization_utils_base.BatchEncoding):
|
80 |
+
logger.info(f"Keys in the BatchEncoding: {chunk_data.keys()}")
|
81 |
+
logger.info(f"Shape of input_ids: {chunk_data['input_ids'].shape}")
|
82 |
+
self.data.append(chunk_data) # Handle BatchEncoding format
|
83 |
+
else:
|
84 |
+
logger.warning(f"Unexpected data type: {type(chunk_data)}")
|
85 |
+
|
86 |
+
logger.info(f"Loaded {len(self.data)} chunks of data")
|
87 |
+
|
88 |
+
def __len__(self):
|
89 |
+
return sum(len(chunk['input_ids']) for chunk in self.data)
|
90 |
+
|
91 |
+
def __getitem__(self, idx):
|
92 |
+
for chunk in self.data:
|
93 |
+
if idx < len(chunk['input_ids']):
|
94 |
+
return {k: v[idx] for k, v in chunk.items()}
|
95 |
+
idx -= len(chunk['input_ids'])
|
96 |
+
raise IndexError("Index out of range")
|
97 |
+
|
98 |
+
def set_seed(seed):
|
99 |
+
random.seed(seed)
|
100 |
+
np.random.seed(seed)
|
101 |
+
torch.manual_seed(seed)
|
102 |
+
torch.cuda.manual_seed_all(seed)
|
103 |
+
|
104 |
+
def load_data(config):
|
105 |
+
logger.info(f"Looking for data in directory: {config.data_dir}")
|
106 |
+
chunk_files = [f for f in os.listdir(config.data_dir) if f.endswith('_tokenized.pt')]
|
107 |
+
logger.info(f"Found {len(chunk_files)} chunk files: {chunk_files}")
|
108 |
+
|
109 |
+
if not chunk_files:
|
110 |
+
raise ValueError(f"No tokenized data files found in {config.data_dir}")
|
111 |
+
|
112 |
+
chunk_files.sort(key=lambda x: int(x.split('_')[1]))
|
113 |
+
chunk_files = [os.path.join(config.data_dir, f) for f in chunk_files]
|
114 |
+
|
115 |
+
dataset = TokenizedDataset(chunk_files[:1]) # Load only the first chunk for now
|
116 |
+
logger.info(f"Created dataset with {len(dataset)} samples")
|
117 |
+
|
118 |
+
return DataLoader(dataset, batch_size=config.batch_size, shuffle=True, num_workers=config.num_workers)
|
119 |
+
|
120 |
+
def extract_hidden_states(batch, model):
|
121 |
+
with torch.no_grad():
|
122 |
+
outputs = model(**batch, output_hidden_states=True)
|
123 |
+
return outputs.hidden_states[0], outputs.hidden_states[-1]
|
124 |
+
|
125 |
+
class KLDivergenceLoss(nn.Module):
|
126 |
+
def forward(self, pred, target):
|
127 |
+
pred = F.log_softmax(pred, dim=-1)
|
128 |
+
target = F.softmax(target, dim=-1)
|
129 |
+
return F.kl_div(pred, target, reduction='batchmean', log_target=False)
|
130 |
+
|
131 |
+
def lr_finder(model, autoencoder, loss_fn, optimizer, train_loader, config):
|
132 |
+
model.eval()
|
133 |
+
autoencoder.train()
|
134 |
+
log_lrs, losses = [], []
|
135 |
+
best_loss, best_lr = float('inf'), None
|
136 |
+
|
137 |
+
pbar = tqdm(total=config.num_lr_steps, desc="LR Finder")
|
138 |
+
for batch_idx, batch in enumerate(train_loader):
|
139 |
+
if batch_idx >= config.num_lr_steps:
|
140 |
+
break
|
141 |
+
|
142 |
+
lr = config.min_lr * (config.max_lr / config.min_lr) ** (batch_idx / (config.num_lr_steps - 1))
|
143 |
+
optimizer.param_groups[0]['lr'] = lr
|
144 |
+
|
145 |
+
batch = {k: v.to(config.device) for k, v in batch.items()}
|
146 |
+
first_states, last_states = extract_hidden_states(batch, model)
|
147 |
+
|
148 |
+
optimizer.zero_grad()
|
149 |
+
reconstructed = autoencoder(first_states)
|
150 |
+
loss = loss_fn(reconstructed, last_states)
|
151 |
+
loss.backward()
|
152 |
+
optimizer.step()
|
153 |
+
|
154 |
+
if loss < best_loss:
|
155 |
+
best_loss = loss.item()
|
156 |
+
best_lr = lr
|
157 |
+
|
158 |
+
log_lrs.append(lr)
|
159 |
+
losses.append(loss.item())
|
160 |
+
|
161 |
+
pbar.update(1)
|
162 |
+
pbar.set_postfix({"Loss": f"{loss.item():.4f}", "LR": f"{lr:.2e}"})
|
163 |
+
|
164 |
+
pbar.close()
|
165 |
+
return log_lrs, losses, best_lr, best_loss
|
166 |
+
|
167 |
+
def run_experiment(config, model, train_loader, num_layers, seed):
|
168 |
+
set_seed(seed)
|
169 |
+
input_dim = model.config.hidden_size
|
170 |
+
hidden_dim = int(input_dim * config.hidden_dim_ratio)
|
171 |
+
autoencoder = ImprovedAutoencoder(input_dim, hidden_dim, num_layers, config.dropout).to(config.device)
|
172 |
+
|
173 |
+
loss_fn = KLDivergenceLoss()
|
174 |
+
optimizer = optim.AdamW(autoencoder.parameters(), lr=config.min_lr)
|
175 |
+
|
176 |
+
log_lrs, losses = [], []
|
177 |
+
best_loss, best_lr = float('inf'), None
|
178 |
+
|
179 |
+
pbar = tqdm(total=config.num_lr_steps, desc=f"LR Finder (Layers: {num_layers}, Seed: {seed})")
|
180 |
+
for batch_idx, batch in enumerate(train_loader):
|
181 |
+
if batch_idx >= config.num_lr_steps:
|
182 |
+
break
|
183 |
+
|
184 |
+
lr = config.min_lr * (config.max_lr / config.min_lr) ** (batch_idx / (config.num_lr_steps - 1))
|
185 |
+
optimizer.param_groups[0]['lr'] = lr
|
186 |
+
|
187 |
+
batch = {k: v.to(config.device) for k, v in batch.items()}
|
188 |
+
first_states, last_states = extract_hidden_states(batch, model)
|
189 |
+
|
190 |
+
optimizer.zero_grad()
|
191 |
+
reconstructed = autoencoder(first_states)
|
192 |
+
loss = loss_fn(reconstructed, last_states)
|
193 |
+
loss.backward()
|
194 |
+
optimizer.step()
|
195 |
+
|
196 |
+
if loss < best_loss:
|
197 |
+
best_loss = loss.item()
|
198 |
+
best_lr = lr
|
199 |
+
|
200 |
+
log_lrs.append(lr)
|
201 |
+
losses.append(loss.item())
|
202 |
+
|
203 |
+
# Log to wandb at every step
|
204 |
+
wandb.log({
|
205 |
+
"loss": loss.item(),
|
206 |
+
"lr": lr,
|
207 |
+
"batch_idx": batch_idx,
|
208 |
+
"num_layers": num_layers,
|
209 |
+
"seed": seed,
|
210 |
+
"best_loss": best_loss,
|
211 |
+
"best_lr": best_lr
|
212 |
+
})
|
213 |
+
|
214 |
+
pbar.update(1)
|
215 |
+
pbar.set_postfix({"Loss": f"{loss.item():.4f}", "LR": f"{lr:.2e}"})
|
216 |
+
|
217 |
+
pbar.close()
|
218 |
+
|
219 |
+
result = {
|
220 |
+
'seed': seed,
|
221 |
+
'num_layers': num_layers,
|
222 |
+
'hidden_dim_ratio': config.hidden_dim_ratio,
|
223 |
+
'dropout': config.dropout,
|
224 |
+
'final_loss': losses[-1],
|
225 |
+
'final_lr': log_lrs[-1],
|
226 |
+
'best_lr': best_lr,
|
227 |
+
'best_loss': best_loss
|
228 |
+
}
|
229 |
+
|
230 |
+
logger.info(f"Experiment completed: {result}")
|
231 |
+
return result
|
232 |
+
|
233 |
+
def main():
|
234 |
+
config = Config()
|
235 |
+
wandb.init(project="qwen-autoencoder-lr-finder", config=config.__dict__)
|
236 |
+
|
237 |
+
logger.info("Loading Qwen model and tokenizer...")
|
238 |
+
model = AutoModelForCausalLM.from_pretrained(config.model_name, torch_dtype=torch.bfloat16).to(config.device)
|
239 |
+
tokenizer = AutoTokenizer.from_pretrained(config.model_name)
|
240 |
+
|
241 |
+
logger.info("Loading data...")
|
242 |
+
train_loader = load_data(config)
|
243 |
+
|
244 |
+
logger.info("Starting experiments...")
|
245 |
+
results = []
|
246 |
+
for num_layers in range(4, 9):
|
247 |
+
for seed in range(1, config.num_seeds + 1):
|
248 |
+
# Start a new wandb run for each experiment
|
249 |
+
with wandb.init(project="qwen-autoencoder-lr-finder",
|
250 |
+
config=config.__dict__,
|
251 |
+
group=f"layers_{num_layers}",
|
252 |
+
name=f"seed_{seed}",
|
253 |
+
job_type="experiment",
|
254 |
+
reinit=True):
|
255 |
+
|
256 |
+
result = run_experiment(config, model, train_loader, num_layers, seed)
|
257 |
+
results.append(result)
|
258 |
+
|
259 |
+
# Save results after each experiment
|
260 |
+
with open('lr_finder_results.jsonl', 'a') as f:
|
261 |
+
json.dump(result, f)
|
262 |
+
f.write('\n')
|
263 |
+
|
264 |
+
# Log final results to wandb
|
265 |
+
wandb.log(result)
|
266 |
+
|
267 |
+
logger.info("Creating visualizations...")
|
268 |
+
plot_results(results)
|
269 |
+
create_heatmap(results)
|
270 |
+
create_parallel_coordinates_plot(results)
|
271 |
+
create_3d_scatter(results)
|
272 |
+
|
273 |
+
logger.info("Experiment completed. Check WandB for detailed results and visualizations.")
|
274 |
+
|
275 |
+
if __name__ == "__main__":
|
276 |
+
main()
|
277 |
+
def run_experiments_sequential(config, model, train_loader):
|
278 |
+
results = []
|
279 |
+
for num_layers in tqdm(range(4, 9), desc="Number of Layers"):
|
280 |
+
for seed in tqdm(range(1, config.num_seeds + 1), desc="Seeds", leave=False):
|
281 |
+
result = run_experiment(config, model, train_loader, num_layers, seed)
|
282 |
+
results.append(result)
|
283 |
+
return results
|
284 |
+
|
285 |
+
def plot_results(results):
|
286 |
+
fig, axs = plt.subplots(3, 2, figsize=(20, 30))
|
287 |
+
fig.suptitle('Learning Rate Finder Results')
|
288 |
+
|
289 |
+
for i, num_layers in enumerate(range(4, 9)):
|
290 |
+
layer_results = [r for r in results if r['num_layers'] == num_layers]
|
291 |
+
best_lrs = [r['Best'] for r in layer_results]
|
292 |
+
best_losses = [r['best_loss'] for r in layer_results]
|
293 |
+
|
294 |
+
axs[i // 2, i % 2].scatter(best_lrs, best_losses, alpha=0.5)
|
295 |
+
axs[i // 2, i % 2].set_xlabel('Best Learning Rate')
|
296 |
+
axs[i // 2, i % 2].set_ylabel('Best Loss')
|
297 |
+
axs[i // 2, i % 2].set_title(f'{num_layers} Layers')
|
298 |
+
axs[i // 2, i % 2].set_xscale('log')
|
299 |
+
axs[i // 2, i % 2].set_yscale('log')
|
300 |
+
|
301 |
+
plt.tight_layout()
|
302 |
+
wandb.log({"lr_loss_relationships": wandb.Image(plt)})
|
303 |
+
plt.close()
|
304 |
+
|
305 |
+
def create_heatmap(results):
|
306 |
+
layer_counts = len(set(r['num_layers'] for r in results))
|
307 |
+
seed_counts = len(set(r['seed'] for r in results))
|
308 |
+
|
309 |
+
heatmap_data = np.zeros((layer_counts, seed_counts))
|
310 |
+
for r in results:
|
311 |
+
layer_idx = r['num_layers'] - 4
|
312 |
+
seed_idx = r['seed'] - 1
|
313 |
+
heatmap_data[layer_idx, seed_idx] = r['best_loss']
|
314 |
+
|
315 |
+
plt.figure(figsize=(20, 10))
|
316 |
+
plt.imshow(heatmap_data, aspect='auto', cmap='viridis')
|
317 |
+
plt.colorbar(label='Best Loss')
|
318 |
+
plt.xlabel('Seed')
|
319 |
+
plt.ylabel('Number of Layers')
|
320 |
+
plt.title('Heatmap of Best Loss across Layers and Seeds')
|
321 |
+
plt.tight_layout()
|
322 |
+
wandb.log({"loss_heatmap": wandb.Image(plt)})
|
323 |
+
plt.close()
|
324 |
+
|
325 |
+
def create_parallel_coordinates_plot(results):
|
326 |
+
df = pd.DataFrame(results)
|
327 |
+
|
328 |
+
plt.figure(figsize=(20, 10))
|
329 |
+
pd.plotting.parallel_coordinates(df, 'num_layers', colormap='viridis')
|
330 |
+
plt.title('Parallel Coordinates Plot of Hyperparameters')
|
331 |
+
plt.tight_layout()
|
332 |
+
wandb.log({"parallel_coordinates": wandb.Image(plt)})
|
333 |
+
plt.close()
|
334 |
+
|
335 |
+
def create_3d_scatter(results):
|
336 |
+
fig = plt.figure(figsize=(15, 15))
|
337 |
+
ax = fig.add_subplot(111, projection='3d')
|
338 |
+
|
339 |
+
for num_layers in range(4, 9):
|
340 |
+
layer_results = [r for r in results if r['num_layers'] == num_layers]
|
341 |
+
x = [r['Best'] for r in layer_results]
|
342 |
+
y = [r['best_loss'] for r in layer_results]
|
343 |
+
z = [r['seed'] for r in layer_results]
|
344 |
+
ax.scatter(x, y, z, label=f'{num_layers} Layers')
|
345 |
+
|
346 |
+
ax.set_xlabel('Best Learning Rate')
|
347 |
+
ax.set_ylabel('Best Loss')
|
348 |
+
ax.set_zlabel('Seed')
|
349 |
+
ax.set_xscale('log')
|
350 |
+
ax.set_yscale('log')
|
351 |
+
ax.legend()
|
352 |
+
plt.title('3D Scatter Plot of Best LR, Loss, and Seed')
|
353 |
+
plt.tight_layout()
|
354 |
+
wandb.log({"3d_scatter": wandb.Image(plt)})
|
355 |
+
plt.close()
|
356 |
+
|
357 |
+
def main():
|
358 |
+
mp.set_start_method('spawn')
|
359 |
+
config = Config()
|
360 |
+
wandb.init(project="qwen-autoencoder-lr-finder", config=config.__dict__)
|
361 |
+
|
362 |
+
logger.info("Loading Qwen model and tokenizer...")
|
363 |
+
model = AutoModelForCausalLM.from_pretrained(config.model_name, torch_dtype=torch.bfloat16).to(config.device)
|
364 |
+
tokenizer = AutoTokenizer.from_pretrained(config.model_name)
|
365 |
+
|
366 |
+
logger.info("Loading data...")
|
367 |
+
train_loader = load_data(config)
|
368 |
+
|
369 |
+
logger.info("Starting experiments...")
|
370 |
+
results = run_experiments_sequential(config, model, train_loader)
|
371 |
+
|
372 |
+
logger.info("Saving results...")
|
373 |
+
with open('lr_finder_results.jsonl', 'w') as f:
|
374 |
+
for result in results:
|
375 |
+
json.dump(result, f)
|
376 |
+
f.write('\n')
|
377 |
+
|
378 |
+
logger.info("Creating visualizations...")
|
379 |
+
plot_results(results)
|
380 |
+
create_heatmap(results)
|
381 |
+
create_parallel_coordinates_plot(results)
|
382 |
+
create_3d_scatter(results)
|
383 |
+
|
384 |
+
logger.info("Experiment completed. Check WandB for detailed results and visualizations.")
|
385 |
+
wandb.finish()
|
386 |
+
|
387 |
+
if __name__ == "__main__":
|
388 |
+
main()
|