Spaces:
Running
Running
Commit
·
d78dede
1
Parent(s):
64a70c0
investigate bugs in Finetrainers
Browse files- finetrainers/dataset.py +165 -66
- finetrainers/trainer.py +28 -0
- training/cogvideox/dataset.py +2 -2
finetrainers/dataset.py
CHANGED
@@ -15,6 +15,9 @@ from torchvision import transforms
|
|
15 |
from torchvision.transforms import InterpolationMode
|
16 |
from torchvision.transforms.functional import resize
|
17 |
|
|
|
|
|
|
|
18 |
|
19 |
# Must import after torch because this can sometimes lead to a nasty segmentation fault, or stack smashing error
|
20 |
# Very few bug reports but it happens. Look in decord Github issues for more relevant information.
|
@@ -30,6 +33,22 @@ from .constants import ( # noqa
|
|
30 |
)
|
31 |
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
logger = get_logger(__name__)
|
34 |
|
35 |
|
@@ -229,20 +248,48 @@ class ImageOrVideoDataset(Dataset):
|
|
229 |
return image
|
230 |
|
231 |
def _preprocess_video(self, path: Path) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
232 |
-
|
233 |
Loads a single video, or latent and prompt embedding, based on initialization parameters.
|
234 |
-
|
235 |
Returns a [F, C, H, W] video tensor.
|
236 |
"""
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
|
245 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
|
247 |
|
248 |
class ImageOrVideoDatasetWithResizing(ImageOrVideoDataset):
|
@@ -264,35 +311,60 @@ class ImageOrVideoDatasetWithResizing(ImageOrVideoDataset):
|
|
264 |
return image
|
265 |
|
266 |
def _preprocess_video(self, path: Path) -> torch.Tensor:
|
267 |
-
|
268 |
-
|
269 |
-
#print(f"ImageOrVideoDatasetWithResizing: self.resolution_buckets = ", self.resolution_buckets)
|
270 |
-
#print(f"ImageOrVideoDatasetWithResizing: self.max_num_frames = ", self.max_num_frames)
|
271 |
-
#print(f"ImageOrVideoDatasetWithResizing: video_num_frames = ", video_num_frames)
|
272 |
-
|
273 |
-
video_buckets = [bucket for bucket in self.resolution_buckets if bucket[0] <= video_num_frames]
|
274 |
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
296 |
|
297 |
def _find_nearest_resolution(self, height, width):
|
298 |
nearest_res = min(self.resolution_buckets, key=lambda x: abs(x[1] - height) + abs(x[2] - width))
|
@@ -338,35 +410,62 @@ class ImageOrVideoDatasetWithResizeAndRectangleCrop(ImageOrVideoDataset):
|
|
338 |
return arr
|
339 |
|
340 |
def _preprocess_video(self, path: Path) -> torch.Tensor:
|
341 |
-
|
342 |
-
|
343 |
-
print(f"ImageOrVideoDatasetWithResizeAndRectangleCrop: self.resolution_buckets = ", self.resolution_buckets)
|
344 |
-
print(f"ImageOrVideoDatasetWithResizeAndRectangleCrop: self.max_num_frames = ", self.max_num_frames)
|
345 |
-
print(f"ImageOrVideoDatasetWithResizeAndRectangleCrop: video_num_frames = ", video_num_frames)
|
346 |
|
347 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
348 |
|
349 |
-
if not video_buckets:
|
350 |
-
_, h, w = self.resolution_buckets[0]
|
351 |
-
video_buckets = [(1, h, w)]
|
352 |
-
|
353 |
-
nearest_frame_bucket = min(
|
354 |
-
video_buckets,
|
355 |
-
key=lambda x: abs(x[0] - min(video_num_frames, self.max_num_frames)),
|
356 |
-
default=video_buckets[0],
|
357 |
-
)[0]
|
358 |
-
|
359 |
-
frame_indices = list(range(0, video_num_frames, video_num_frames // nearest_frame_bucket))
|
360 |
-
|
361 |
-
frames = video_reader.get_batch(frame_indices)
|
362 |
-
frames = frames[:nearest_frame_bucket].float()
|
363 |
-
frames = frames.permute(0, 3, 1, 2).contiguous()
|
364 |
-
|
365 |
-
nearest_res = self._find_nearest_resolution(frames.shape[2], frames.shape[3])
|
366 |
-
frames_resized = self._resize_for_rectangle_crop(frames, nearest_res)
|
367 |
-
frames = torch.stack([self.video_transforms(frame) for frame in frames_resized], dim=0)
|
368 |
-
return frames
|
369 |
-
|
370 |
def _find_nearest_resolution(self, height, width):
|
371 |
nearest_res = min(self.resolutions, key=lambda x: abs(x[1] - height) + abs(x[2] - width))
|
372 |
return nearest_res[1], nearest_res[2]
|
|
|
15 |
from torchvision.transforms import InterpolationMode
|
16 |
from torchvision.transforms.functional import resize
|
17 |
|
18 |
+
import gc
|
19 |
+
import time
|
20 |
+
import resource
|
21 |
|
22 |
# Must import after torch because this can sometimes lead to a nasty segmentation fault, or stack smashing error
|
23 |
# Very few bug reports but it happens. Look in decord Github issues for more relevant information.
|
|
|
33 |
)
|
34 |
|
35 |
|
36 |
+
# Decord is causing us some issues!
|
37 |
+
# Let's try to increase file descriptor limits to avoid this error:
|
38 |
+
#
|
39 |
+
# decord._ffi.base.DECORDError: Resource temporarily unavailable
|
40 |
+
try:
|
41 |
+
soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
|
42 |
+
logger.info(f"Current file descriptor limits: soft={soft}, hard={hard}")
|
43 |
+
|
44 |
+
# Try to increase to hard limit if possible
|
45 |
+
if soft < hard:
|
46 |
+
resource.setrlimit(resource.RLIMIT_NOFILE, (hard, hard))
|
47 |
+
new_soft, new_hard = resource.getrlimit(resource.RLIMIT_NOFILE)
|
48 |
+
logger.info(f"Updated file descriptor limits: soft={new_soft}, hard={new_hard}")
|
49 |
+
except Exception as e:
|
50 |
+
logger.warning(f"Could not check or update file descriptor limits: {e}")
|
51 |
+
|
52 |
logger = get_logger(__name__)
|
53 |
|
54 |
|
|
|
248 |
return image
|
249 |
|
250 |
def _preprocess_video(self, path: Path) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
251 |
+
"""
|
252 |
Loads a single video, or latent and prompt embedding, based on initialization parameters.
|
|
|
253 |
Returns a [F, C, H, W] video tensor.
|
254 |
"""
|
255 |
+
max_retries = 3
|
256 |
+
retry_delay = 1.0 # seconds
|
257 |
+
|
258 |
+
for attempt in range(max_retries):
|
259 |
+
try:
|
260 |
+
# Create video reader
|
261 |
+
video_reader = decord.VideoReader(uri=path.as_posix())
|
262 |
+
video_num_frames = len(video_reader)
|
263 |
+
|
264 |
+
# Process frames
|
265 |
+
indices = list(range(0, video_num_frames, video_num_frames // self.max_num_frames))
|
266 |
+
frames = video_reader.get_batch(indices)
|
267 |
+
frames = frames[: self.max_num_frames].float()
|
268 |
+
frames = frames.permute(0, 3, 1, 2).contiguous()
|
269 |
+
frames = torch.stack([self.video_transforms(frame) for frame in frames], dim=0)
|
270 |
+
|
271 |
+
# Explicitly clean up resources
|
272 |
+
del video_reader
|
273 |
+
|
274 |
+
# Force garbage collection occasionally
|
275 |
+
if random.random() < 0.05: # 5% chance
|
276 |
+
gc.collect()
|
277 |
+
|
278 |
+
return frames
|
279 |
+
|
280 |
+
except decord._ffi.base.DECORDError as e:
|
281 |
+
# Log the error
|
282 |
+
error_msg = str(e)
|
283 |
+
if "Resource temporarily unavailable" in error_msg and attempt < max_retries - 1:
|
284 |
+
logger.warning(f"Retry {attempt+1}/{max_retries} loading video {path}: {error_msg}")
|
285 |
+
|
286 |
+
# Clean up and wait before retrying
|
287 |
+
gc.collect()
|
288 |
+
time.sleep(retry_delay * (attempt + 1)) # Increasing backoff
|
289 |
+
else:
|
290 |
+
# Either not a resource error or we've run out of retries
|
291 |
+
logger.error(f"Failed to load video {path} after {attempt+1} attempts: {error_msg}")
|
292 |
+
raise RuntimeError(f"Failed to load video after {max_retries} attempts: {error_msg}")
|
293 |
|
294 |
|
295 |
class ImageOrVideoDatasetWithResizing(ImageOrVideoDataset):
|
|
|
311 |
return image
|
312 |
|
313 |
def _preprocess_video(self, path: Path) -> torch.Tensor:
|
314 |
+
max_retries = 3
|
315 |
+
retry_delay = 1.0 # seconds
|
|
|
|
|
|
|
|
|
|
|
316 |
|
317 |
+
for attempt in range(max_retries):
|
318 |
+
try:
|
319 |
+
# Create video reader
|
320 |
+
video_reader = decord.VideoReader(uri=path.as_posix())
|
321 |
+
video_num_frames = len(video_reader)
|
322 |
+
|
323 |
+
# Find appropriate bucket for the video
|
324 |
+
video_buckets = [bucket for bucket in self.resolution_buckets if bucket[0] <= video_num_frames]
|
325 |
+
|
326 |
+
if not video_buckets:
|
327 |
+
_, h, w = self.resolution_buckets[0]
|
328 |
+
video_buckets = [(1, h, w)]
|
329 |
+
|
330 |
+
nearest_frame_bucket = min(
|
331 |
+
video_buckets,
|
332 |
+
key=lambda x: abs(x[0] - min(video_num_frames, self.max_num_frames)),
|
333 |
+
default=video_buckets[0],
|
334 |
+
)[0]
|
335 |
+
|
336 |
+
# Extract and process frames
|
337 |
+
frame_indices = list(range(0, video_num_frames, video_num_frames // nearest_frame_bucket))
|
338 |
+
frames = video_reader.get_batch(frame_indices)
|
339 |
+
frames = frames[:nearest_frame_bucket].float()
|
340 |
+
frames = frames.permute(0, 3, 1, 2).contiguous()
|
341 |
+
|
342 |
+
nearest_res = self._find_nearest_resolution(frames.shape[2], frames.shape[3])
|
343 |
+
frames_resized = torch.stack([resize(frame, nearest_res) for frame in frames], dim=0)
|
344 |
+
frames = torch.stack([self.video_transforms(frame) for frame in frames_resized], dim=0)
|
345 |
+
|
346 |
+
# Explicitly clean up resources
|
347 |
+
del video_reader
|
348 |
+
|
349 |
+
# Force garbage collection occasionally
|
350 |
+
if random.random() < 0.05: # 5% chance
|
351 |
+
gc.collect()
|
352 |
+
|
353 |
+
return frames
|
354 |
+
|
355 |
+
except decord._ffi.base.DECORDError as e:
|
356 |
+
# Log the error
|
357 |
+
error_msg = str(e)
|
358 |
+
if "Resource temporarily unavailable" in error_msg and attempt < max_retries - 1:
|
359 |
+
logger.warning(f"Retry {attempt+1}/{max_retries} loading video {path}: {error_msg}")
|
360 |
+
|
361 |
+
# Clean up and wait before retrying
|
362 |
+
gc.collect()
|
363 |
+
time.sleep(retry_delay * (attempt + 1)) # Increasing backoff
|
364 |
+
else:
|
365 |
+
# Either not a resource error or we've run out of retries
|
366 |
+
logger.error(f"Failed to load video {path} after {attempt+1} attempts: {error_msg}")
|
367 |
+
raise RuntimeError(f"Failed to load video after {max_retries} attempts: {error_msg}")
|
368 |
|
369 |
def _find_nearest_resolution(self, height, width):
|
370 |
nearest_res = min(self.resolution_buckets, key=lambda x: abs(x[1] - height) + abs(x[2] - width))
|
|
|
410 |
return arr
|
411 |
|
412 |
def _preprocess_video(self, path: Path) -> torch.Tensor:
|
413 |
+
max_retries = 3
|
414 |
+
retry_delay = 1.0 # seconds
|
|
|
|
|
|
|
415 |
|
416 |
+
for attempt in range(max_retries):
|
417 |
+
try:
|
418 |
+
# Create video reader
|
419 |
+
video_reader = decord.VideoReader(uri=path.as_posix())
|
420 |
+
video_num_frames = len(video_reader)
|
421 |
+
|
422 |
+
# Find appropriate bucket for the video
|
423 |
+
video_buckets = [bucket for bucket in self.resolution_buckets if bucket[0] <= video_num_frames]
|
424 |
+
|
425 |
+
if not video_buckets:
|
426 |
+
_, h, w = self.resolution_buckets[0]
|
427 |
+
video_buckets = [(1, h, w)]
|
428 |
+
|
429 |
+
nearest_frame_bucket = min(
|
430 |
+
video_buckets,
|
431 |
+
key=lambda x: abs(x[0] - min(video_num_frames, self.max_num_frames)),
|
432 |
+
default=video_buckets[0],
|
433 |
+
)[0]
|
434 |
+
|
435 |
+
# Extract and process frames
|
436 |
+
frame_indices = list(range(0, video_num_frames, video_num_frames // nearest_frame_bucket))
|
437 |
+
frames = video_reader.get_batch(frame_indices)
|
438 |
+
frames = frames[:nearest_frame_bucket].float()
|
439 |
+
frames = frames.permute(0, 3, 1, 2).contiguous()
|
440 |
+
|
441 |
+
# Fix: Change self.resolutions to self.resolution_buckets to match the class attribute
|
442 |
+
nearest_res = self._find_nearest_resolution(frames.shape[2], frames.shape[3])
|
443 |
+
frames_resized = self._resize_for_rectangle_crop(frames, nearest_res)
|
444 |
+
frames = torch.stack([self.video_transforms(frame) for frame in frames_resized], dim=0)
|
445 |
+
|
446 |
+
# Explicitly clean up resources
|
447 |
+
del video_reader
|
448 |
+
|
449 |
+
# Force garbage collection occasionally
|
450 |
+
if random.random() < 0.05: # 5% chance
|
451 |
+
gc.collect()
|
452 |
+
|
453 |
+
return frames
|
454 |
+
|
455 |
+
except decord._ffi.base.DECORDError as e:
|
456 |
+
# Log the error
|
457 |
+
error_msg = str(e)
|
458 |
+
if "Resource temporarily unavailable" in error_msg and attempt < max_retries - 1:
|
459 |
+
logger.warning(f"Retry {attempt+1}/{max_retries} loading video {path}: {error_msg}")
|
460 |
+
|
461 |
+
# Clean up and wait before retrying
|
462 |
+
gc.collect()
|
463 |
+
time.sleep(retry_delay * (attempt + 1)) # Increasing backoff
|
464 |
+
else:
|
465 |
+
# Either not a resource error or we've run out of retries
|
466 |
+
logger.error(f"Failed to load video {path} after {attempt+1} attempts: {error_msg}")
|
467 |
+
raise RuntimeError(f"Failed to load video after {max_retries} attempts: {error_msg}")
|
468 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
469 |
def _find_nearest_resolution(self, height, width):
|
470 |
nearest_res = min(self.resolutions, key=lambda x: abs(x[1] - height) + abs(x[2] - width))
|
471 |
return nearest_res[1], nearest_res[2]
|
finetrainers/trainer.py
CHANGED
@@ -2,6 +2,7 @@ import json
|
|
2 |
import logging
|
3 |
import math
|
4 |
import os
|
|
|
5 |
import random
|
6 |
from datetime import datetime, timedelta
|
7 |
from pathlib import Path
|
@@ -549,6 +550,20 @@ class Trainer:
|
|
549 |
def train(self) -> None:
|
550 |
logger.info("Starting training")
|
551 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
552 |
memory_statistics = get_memory_statistics()
|
553 |
logger.info(f"Memory before training start: {json.dumps(memory_statistics, indent=4)}")
|
554 |
|
@@ -816,9 +831,15 @@ class Trainer:
|
|
816 |
progress_bar.set_postfix(logs)
|
817 |
accelerator.log(logs, step=global_step)
|
818 |
|
|
|
|
|
|
|
|
|
819 |
if global_step >= self.state.train_steps:
|
820 |
break
|
821 |
|
|
|
|
|
822 |
if num_loss_updates > 0:
|
823 |
epoch_loss /= num_loss_updates
|
824 |
accelerator.log({"epoch_loss": epoch_loss}, step=global_step)
|
@@ -833,6 +854,13 @@ class Trainer:
|
|
833 |
if should_run_validation:
|
834 |
self.validate(global_step)
|
835 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
836 |
accelerator.wait_for_everyone()
|
837 |
if accelerator.is_main_process:
|
838 |
transformer = unwrap_model(accelerator, self.transformer)
|
|
|
2 |
import logging
|
3 |
import math
|
4 |
import os
|
5 |
+
import gc
|
6 |
import random
|
7 |
from datetime import datetime, timedelta
|
8 |
from pathlib import Path
|
|
|
550 |
def train(self) -> None:
|
551 |
logger.info("Starting training")
|
552 |
|
553 |
+
|
554 |
+
# Add these lines at the beginning
|
555 |
+
if hasattr(resource, 'RLIMIT_NOFILE'):
|
556 |
+
try:
|
557 |
+
soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
|
558 |
+
logger.info(f"Current file descriptor limits in trainer: soft={soft}, hard={hard}")
|
559 |
+
# Try to increase to hard limit if possible
|
560 |
+
if soft < hard:
|
561 |
+
resource.setrlimit(resource.RLIMIT_NOFILE, (hard, hard))
|
562 |
+
new_soft, new_hard = resource.getrlimit(resource.RLIMIT_NOFILE)
|
563 |
+
logger.info(f"Updated file descriptor limits: soft={new_soft}, hard={new_hard}")
|
564 |
+
except Exception as e:
|
565 |
+
logger.warning(f"Could not check or update file descriptor limits: {e}")
|
566 |
+
|
567 |
memory_statistics = get_memory_statistics()
|
568 |
logger.info(f"Memory before training start: {json.dumps(memory_statistics, indent=4)}")
|
569 |
|
|
|
831 |
progress_bar.set_postfix(logs)
|
832 |
accelerator.log(logs, step=global_step)
|
833 |
|
834 |
+
if global_step % 100 == 0: # Every 100 steps
|
835 |
+
# Force garbage collection to clean up any lingering resources
|
836 |
+
gc.collect()
|
837 |
+
|
838 |
if global_step >= self.state.train_steps:
|
839 |
break
|
840 |
|
841 |
+
|
842 |
+
|
843 |
if num_loss_updates > 0:
|
844 |
epoch_loss /= num_loss_updates
|
845 |
accelerator.log({"epoch_loss": epoch_loss}, step=global_step)
|
|
|
854 |
if should_run_validation:
|
855 |
self.validate(global_step)
|
856 |
|
857 |
+
if epoch % 3 == 0: # Every 3 epochs
|
858 |
+
logger.info("Performing periodic resource cleanup")
|
859 |
+
free_memory()
|
860 |
+
gc.collect()
|
861 |
+
torch.cuda.empty_cache()
|
862 |
+
torch.cuda.synchronize(accelerator.device)
|
863 |
+
|
864 |
accelerator.wait_for_everyone()
|
865 |
if accelerator.is_main_process:
|
866 |
transformer = unwrap_model(accelerator, self.transformer)
|
training/cogvideox/dataset.py
CHANGED
@@ -57,7 +57,7 @@ class VideoDataset(Dataset):
|
|
57 |
self.random_flip = random_flip
|
58 |
self.image_to_video = image_to_video
|
59 |
|
60 |
-
self.
|
61 |
(f, h, w) for h in self.height_buckets for w in self.width_buckets for f in self.frame_buckets
|
62 |
]
|
63 |
|
@@ -295,7 +295,7 @@ class VideoDatasetWithResizing(VideoDataset):
|
|
295 |
return image, frames, None
|
296 |
|
297 |
def _find_nearest_resolution(self, height, width):
|
298 |
-
nearest_res = min(self.
|
299 |
return nearest_res[1], nearest_res[2]
|
300 |
|
301 |
|
|
|
57 |
self.random_flip = random_flip
|
58 |
self.image_to_video = image_to_video
|
59 |
|
60 |
+
self.resolution_buckets = [
|
61 |
(f, h, w) for h in self.height_buckets for w in self.width_buckets for f in self.frame_buckets
|
62 |
]
|
63 |
|
|
|
295 |
return image, frames, None
|
296 |
|
297 |
def _find_nearest_resolution(self, height, width):
|
298 |
+
nearest_res = min(self.resolution_buckets, key=lambda x: abs(x[1] - height) + abs(x[2] - width))
|
299 |
return nearest_res[1], nearest_res[2]
|
300 |
|
301 |
|