-
Notifications
You must be signed in to change notification settings - Fork 29
/
helper.py
490 lines (384 loc) · 17.3 KB
/
helper.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
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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
import torch
import torch.nn as nn
import torch.optim as optim
from collections import defaultdict
import numpy as np
import wandb
import os
from torchvision.utils import save_image
import torch.nn.functional as F
import torch
import matplotlib.pyplot as plt
import os
import io
from PIL import Image
from mpl_toolkits.mplot3d import Axes3D
from rich.console import Console
from rich.traceback import install
from rich.progress import track
console = Console(width=3000)
def handle_training_error(e: Exception, batch_idx: int = None, extra_info: str = None):
"""Enhanced error handling with Rich output"""
console.print("\n[bold red]Training Error Occurred![/bold red]")
if batch_idx is not None:
console.print(f"[yellow]Batch Index:[/yellow] {batch_idx}")
if extra_info:
console.print(f"[yellow]Additional Info:[/yellow] {extra_info}")
console.print("[bold yellow]Exception Details:[/bold yellow]")
console.print_exception()
def resize_for_wandb(tensor_image, size=128):
"""
Resize a tensor image for wandb logging.
Args:
tensor_image: Input tensor of shape [C, H, W] or [B, C, H, W]
size: Target size (both height and width)
Returns:
Resized tensor of shape [C, size, size]
"""
if tensor_image.dim() == 4:
tensor_image = tensor_image[0] # Take first image if batched
return F.interpolate(
tensor_image.unsqueeze(0), # Add batch dimension
size=(size, size),
mode='bilinear',
align_corners=False
).squeeze(0) # Remove batch dimension
def consistent_sub_sample(tensor1, tensor2, sub_sample_size):
"""
Consistently sub-sample two tensors with the same random offset.
Args:
tensor1 (torch.Tensor): First input tensor of shape (B, C, H, W)
tensor2 (torch.Tensor): Second input tensor of shape (B, C, H, W)
sub_sample_size (tuple): Desired sub-sample size (h, w)
Returns:
tuple: Sub-sampled versions of tensor1 and tensor2
"""
assert tensor1.shape == tensor2.shape, "Input tensors must have the same shape"
assert tensor1.ndim == 4, "Input tensors should have 4 dimensions (B, C, H, W)"
batch_size, channels, height, width = tensor1.shape
sub_h, sub_w = sub_sample_size
assert height >= sub_h and width >= sub_w, "Sub-sample size should not exceed the tensor dimensions"
offset_x = torch.randint(0, height - sub_h + 1, (1,)).item()
offset_y = torch.randint(0, width - sub_w + 1, (1,)).item()
tensor1_sub = tensor1[..., offset_x:offset_x+sub_h, offset_y:offset_y+sub_w]
tensor2_sub = tensor2[..., offset_x:offset_x+sub_h, offset_y:offset_y+sub_w]
return tensor1_sub, tensor2_sub
def plot_loss_landscape(model, loss_fns, dataloader, num_points=20, alpha=1.0):
# Store original parameters
original_params = [p.clone() for p in model.parameters()]
# Calculate two random directions
direction1 = [torch.randn_like(p) for p in model.parameters()]
direction2 = [torch.randn_like(p) for p in model.parameters()]
# Normalize directions
norm1 = torch.sqrt(sum(torch.sum(d**2) for d in direction1))
norm2 = torch.sqrt(sum(torch.sum(d**2) for d in direction2))
direction1 = [d / norm1 for d in direction1]
direction2 = [d / norm2 for d in direction2]
# Create grid
x = np.linspace(-alpha, alpha, num_points)
y = np.linspace(-alpha, alpha, num_points)
X, Y = np.meshgrid(x, y)
# Calculate loss for each point and each loss function
Z = {f'loss_{i}': np.zeros_like(X) for i in range(len(loss_fns))}
Z['total_loss'] = np.zeros_like(X)
for i in range(num_points):
for j in range(num_points):
# Update model parameters
for p, d1, d2 in zip(model.parameters(), direction1, direction2):
p.data = p.data + X[i,j] * d1 + Y[i,j] * d2
# Calculate loss for each loss function
total_loss = 0
num_batches = 0
for batch in dataloader:
inputs, targets = batch
outputs = model(inputs)
for k, loss_fn in enumerate(loss_fns):
loss = loss_fn(outputs, targets)
Z[f'loss_{k}'][i,j] += loss.item()
total_loss += loss.item()
num_batches += 1
# Average the losses
for k in range(len(loss_fns)):
Z[f'loss_{k}'][i,j] /= num_batches
Z['total_loss'][i,j] = total_loss / num_batches
# Reset model parameters
for p, orig_p in zip(model.parameters(), original_params):
p.data = orig_p.clone()
# Plot the loss landscapes
figs = []
for loss_key in Z.keys():
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111, projection='3d')
surf = ax.plot_surface(X, Y, Z[loss_key], cmap='viridis')
ax.set_xlabel('Direction 1')
ax.set_ylabel('Direction 2')
ax.set_zlabel('Loss')
ax.set_title(f'Loss Landscape - {loss_key}')
fig.colorbar(surf)
figs.append(fig)
# Save the plots to buffers
bufs = []
for fig in figs:
buf = io.BytesIO()
plt.savefig(buf, format='png')
buf.seek(0)
bufs.append(buf)
plt.close(fig)
return bufs
def log_loss_landscape(model, loss_fns, dataloader, step):
# Generate the loss landscape plots
bufs = plot_loss_landscape(model, loss_fns, dataloader)
# Log the plots to wandb
log_dict = {
f"loss_landscape_{i}": wandb.Image(buf, caption=f"Loss Landscape - Loss {i}")
for i, buf in enumerate(bufs[:-1])
}
log_dict["loss_landscape_total"] = wandb.Image(bufs[-1], caption="Loss Landscape - Total Loss")
log_dict["step"] = step
wandb.log(log_dict)
# Global variable to store the current table structure
current_table_columns = None
def log_grad_flow(named_parameters, global_step):
global current_table_columns
grads = []
layers = []
for n, p in named_parameters:
if p.requires_grad and "bias" not in n and p.grad is not None:
layers.append(n)
grads.append(p.grad.abs().mean().item())
if not grads:
print("No valid gradients found for logging.")
return
# Normalize gradients
max_grad = max(grads)
if max_grad == 0:
print("👿👿👿 Warning: All gradients are zero. 👿👿👿")
normalized_grads = grads # Use unnormalized grads if max is zero
raise ValueError(f"👿👿👿 Warning: All gradients are zero. 👿👿👿")
else:
normalized_grads = [g / max_grad for g in grads]
# Create the matplotlib figure
plt.figure(figsize=(12, 6))
plt.bar(range(len(grads)), normalized_grads, alpha=0.5)
plt.xticks(range(len(grads)), layers, rotation="vertical")
plt.xlabel("Layers")
plt.ylabel("Gradient Magnitude")
plt.title(f"Gradient Flow (Step {global_step})")
if max_grad == 0:
plt.title(f"Gradient Flow (Step {global_step}) - All Gradients Zero")
plt.tight_layout()
# Save the figure to a bytes buffer
buf = io.BytesIO()
plt.savefig(buf, format='png')
buf.seek(0)
# Create a wandb.Image from the buffer
img = wandb.Image(Image.open(buf))
plt.close()
# Calculate statistics
stats = {
"max_gradient": max_grad,
"min_gradient": min(grads),
"mean_gradient": np.mean(grads),
"median_gradient": np.median(grads),
"gradient_variance": np.var(grads),
}
# Check for gradient issues
issues = check_gradient_issues(grads, layers)
# Log everything
log_dict = {
"gradient_flow_plot": img,
**stats,
"gradient_issues": wandb.Html(issues),
"step": global_step
}
# Log other metrics
wandb.log(log_dict)
def check_gradient_issues(grads, layers):
issues = []
mean_grad = np.mean(grads)
std_grad = np.std(grads)
for layer, grad in zip(layers, grads):
if grad > mean_grad + 3 * std_grad:
issues.append(f"🔥 Potential exploding gradient in {layer}: {grad:.2e}")
elif grad < mean_grad - 3 * std_grad:
issues.append(f"🥶 Potential vanishing gradient in {layer}: {grad:.2e}")
if issues:
return "<br>".join(issues)
else:
return "✅ No significant gradient issues detected"
def count_model_params(model, trainable_only=False, verbose=False):
"""
Count the number of parameters in a PyTorch model, distinguishing between system native and custom modules.
Args:
model (nn.Module): The PyTorch model to analyze.
trainable_only (bool): If True, count only trainable parameters. Default is False.
verbose (bool): If True, print detailed breakdown of parameters. Default is False.
Returns:
float: Total number of (trainable) parameters in millions.
dict: Breakdown of parameters by layer type.
"""
total_params = 0
trainable_params = 0
param_counts = defaultdict(int)
# List of PyTorch native modules
native_modules = set([name for name, obj in nn.__dict__.items() if isinstance(obj, type)])
for name, module in model.named_modules():
for param_name, param in module.named_parameters():
if param.requires_grad:
trainable_params += param.numel()
total_params += param.numel()
# Count parameters for each layer type
layer_type = module.__class__.__name__
if layer_type in native_modules:
layer_type = f"Native_{layer_type}"
else:
layer_type = f"Custom_{layer_type}"
param_counts[layer_type] += param.numel()
if verbose:
native_counts = {k: v for k, v in param_counts.items() if k.startswith("Native_")}
custom_counts = {k: v for k, v in param_counts.items() if k.startswith("Custom_")}
native_total = sum(native_counts.values())
custom_total = sum(custom_counts.values())
print("-" * 55)
print(f"{'☕ Native Modules Total':<30} {native_total:<15,d} {native_total/total_params*100:.2f}%")
print("-" * 55)
# Print native modules
for i, (layer_type, count) in enumerate(sorted(native_counts.items(), key=lambda x: x[1], reverse=True), 1):
percentage = count / total_params * 100
print(f" {i}. ⅀ {layer_type[7:]:<23} {count:<15,d} {percentage:.2f}%")
print("-" * 55)
print(f"{'🍄 Custom Modules Total':<30} {custom_total:<15,d} {custom_total/total_params*100:.2f}%")
print("-" * 55)
# Print custom modules
for i, (layer_type, count) in enumerate(sorted(custom_counts.items(), key=lambda x: x[1], reverse=True), 1):
percentage = count / total_params * 100
print(f" {i}. {layer_type[7:]:<23} {count:<15,d} {percentage:.2f}%")
print(f"{'Layer Type':<30} {'Parameter Count':<15} {'% of Total':<10}")
print("-" * 55)
print(f"{'Total':<30} {total_params:<15,d} 100.00%")
print(f"{'Trainable':<30} {trainable_params:<15,d} {trainable_params/total_params*100:.2f}%")
if trainable_only:
return trainable_params / 1e6, dict(param_counts)
else:
return total_params / 1e6, dict(param_counts)
def normalize(tensor):
mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1).to(tensor.device)
std = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1).to(tensor.device)
return (tensor - mean) / std
def sample_recon(model, data, accelerator, output_path, num_samples=1):
model.eval()
with torch.no_grad():
try:
x_reconstructed,x_current, x_reference = data
batch_size = x_reconstructed.size(0)
num_samples = min(num_samples, batch_size)
# Select a subset of images if batch_size > num_samples
x_reconstructed = x_reconstructed[:num_samples]
x_reference = x_reference[:num_samples]
x_current = x_current[:num_samples]
# Prepare frames for saving (2 rows: clamped reconstructed and original reference)
frames = torch.cat((x_reconstructed,x_current, x_reference), dim=0)
# Ensure we have a valid output directory
if output_path:
output_dir = os.path.dirname(output_path)
if not output_dir:
output_dir = '.'
os.makedirs(output_dir, exist_ok=True)
# Save frames as a grid (2 rows, num_samples columns)
save_image(accelerator.gather(frames), output_path, nrow=num_samples, padding=2, normalize=False)
# accelerator.print(f"Saved sample reconstructions to {output_path}")
else:
accelerator.print("Warning: No output path provided. Skipping image save.")
# Log images to wandb
wandb_images = []
for i in range(num_samples):
wandb_images.extend([
wandb.Image(x_reconstructed[i].cpu().detach().numpy().transpose(1, 2, 0), caption=f"x_reconstructed {i}"),
wandb.Image(x_current[i].cpu().detach().numpy().transpose(1, 2, 0), caption=f"x_current {i}"),
wandb.Image(x_reference[i].cpu().detach().numpy().transpose(1, 2, 0), caption=f"x_reference {i}")
])
wandb.log({"Sample Reconstructions": wandb_images})
return frames
except RuntimeError as e:
print(f"🔥 e:{e}")
return None
def monitor_gradients(model, epoch, batch_idx, log_interval=10):
"""
Monitor gradients of the model parameters.
:param model: The neural network model
:param epoch: Current epoch number
:param batch_idx: Current batch index
:param log_interval: How often to log gradient statistics
"""
if batch_idx % log_interval == 0:
grad_stats = defaultdict(list)
for name, param in model.named_parameters():
if param.grad is not None:
grad_norm = param.grad.norm().item()
grad_stats['norm'].append(grad_norm)
if torch.isnan(param.grad).any():
print(f"NaN gradient detected in {name}")
if torch.isinf(param.grad).any():
print(f"Inf gradient detected in {name}")
grad_stats['names'].append(name)
if grad_stats['norm']:
avg_norm = np.mean(grad_stats['norm'])
max_norm = np.max(grad_stats['norm'])
min_norm = np.min(grad_stats['norm'])
print(f"Epoch {epoch}, Batch {batch_idx}")
print(f"Gradient norms - Avg: {avg_norm:.4f}, Max: {max_norm:.4f}, Min: {min_norm:.4f}")
# Identify layers with unusually high or low gradients
threshold_high = avg_norm * 10 # Adjust this multiplier as needed
threshold_low = avg_norm * 0.1 # Adjust this multiplier as needed
for name, norm in zip(grad_stats['names'], grad_stats['norm']):
if norm > threshold_high:
print(f"High gradient in {name}: {norm:.4f}")
elif norm < threshold_low:
print(f"Low gradient in {name}: {norm:.4f}")
else:
print("No gradients to monitor")
# helper for gradient vanishing / explosion
def hook_fn(name):
def hook(grad):
if torch.isnan(grad).any():
# print(f"🔥 NaN gradient detected in {name}")
return torch.zeros_like(grad) # Replace NaN with zero
elif torch.isinf(grad).any():
# print(f"🔥 Inf gradient detected in {name}")
return torch.clamp(grad, -1e6, 1e6) # Clamp infinite values
#else:
# You can add more conditions or logging here
# grad_norm = grad.norm().item()
# print(f"Gradient norm for {name}: {grad_norm}")
return grad
return hook
def add_gradient_hooks(model):
for name, param in model.named_parameters():
if param.requires_grad:
param.register_hook(hook_fn(name))
def visualize_latent_token(token, save_path):
"""
Visualize a 1D latent token as a colorful bar.
Args:
token (torch.Tensor): A 1D tensor representing the latent token.
save_path (str): Path to save the visualization.
"""
# Ensure the token is on CPU and convert to numpy
token_np = token.cpu().detach().numpy()
# Create a figure and axis
fig, ax = plt.subplots(figsize=(10, 0.5))
# Normalize the token values to [0, 1] for colormap
token_normalized = (token_np - token_np.min()) / (token_np.max() - token_np.min())
# Create a colorful representation
cmap = plt.get_cmap('viridis')
colors = cmap(token_normalized)
# Plot the token as a colorful bar
ax.imshow(colors.reshape(1, -1, 4), aspect='auto')
# Remove axes
ax.set_xticks([])
ax.set_yticks([])
# Add a title
plt.title(f"Latent Token (dim={len(token_np)})")
# Save the figure
plt.savefig(save_path, bbox_inches='tight', pad_inches=0)
plt.close()