storytold/FineTrainers-Conditioning
2
0
Fork 0
mirror of https://github.com/storytold/FineTrainers-Conditioning.git synced 2026-10-17 00:23:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

updates

Browse Source
This commit is contained in:
sayakpaul
2024-11-18 15:28:37 +05:30
parent 9852c3d994
commit a40ccd2237
1 changed files with 12 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
training/dataset.py
+12 -17
View File
@@ -22,8 +22,8 @@ decord.bridge.set_bridge("torch")
logger = get_logger(__name__) logger = get_logger(__name__)
HEIGHT_BUCKETS = [256, 320, 384, 480, 512, 576, 720, 768, 960, 1024, 1280, 1536] HEIGHT_BUCKETS = [256, 320, 384, 480, 512, 576, 720, 768, 960, 1024, 1280, 1536]
WIDTH_BUCKETS = [256, 320, 384, 480, 512, 576, 720, 768, 848, 960, 1024, 1280, 1536] WIDTH_BUCKETS = [256, 320, 384, 480, 512, 576, 720, 768, 960, 1024, 1280, 1536]
FRAME_BUCKETS = [16, 24, 32, 48, 64, 80, 84] FRAME_BUCKETS = [16, 24, 32, 48, 64, 80]
class VideoDataset(Dataset): class VideoDataset(Dataset):
@@ -144,17 +144,16 @@ class VideoDataset(Dataset):
} }
else: else:
image, video, _ = self._preprocess_video(self.video_paths[index]) image, video, _ = self._preprocess_video(self.video_paths[index])
if video is not None: return {
return { "prompt": self.id_token + self.prompts[index],
"prompt": self.id_token + self.prompts[index], "image": image,
"image": image, "video": video,
"video": video, "video_metadata": {
"video_metadata": { "num_frames": video.shape[0],
"num_frames": video.shape[0], "height": video.shape[2],
"height": video.shape[2], "width": video.shape[3],
"width": video.shape[3], },
}, }
}
def _load_dataset_from_local_path(self) -> Tuple[List[str], List[str]]: def _load_dataset_from_local_path(self) -> Tuple[List[str], List[str]]:
if not self.data_root.exists(): if not self.data_root.exists():
@@ -276,13 +275,9 @@ class VideoDatasetWithResizing(VideoDataset):
else: else:
video_reader = decord.VideoReader(uri=path.as_posix()) video_reader = decord.VideoReader(uri=path.as_posix())
video_num_frames = len(video_reader) video_num_frames = len(video_reader)
nearest_frame_bucket = min( nearest_frame_bucket = min(
self.frame_buckets, key=lambda x: abs(x - min(video_num_frames, self.max_num_frames)) self.frame_buckets, key=lambda x: abs(x - min(video_num_frames, self.max_num_frames))
) )
if video_num_frames < nearest_frame_bucket:
# TODO: we could handle this by padding zero frames or duplicating the existing frames?
return None, None, None
frame_indices = list(range(0, video_num_frames, video_num_frames // nearest_frame_bucket)) frame_indices = list(range(0, video_num_frames, video_num_frames // nearest_frame_bucket))
frames = video_reader.get_batch(frame_indices) frames = video_reader.get_batch(frame_indices)