Spaces:
Running
on
Zero
Running
on
Zero
Antoni Bigata
commited on
Commit
·
4ef25d2
1
Parent(s):
4fd1a69
requirements
Browse files
app.py
CHANGED
@@ -614,257 +614,253 @@ def process_video(video_input, audio_input, max_num_seconds):
|
|
614 |
audio_input = DEFAULT_AUDIO_PATH
|
615 |
print(f"Using default audio: {DEFAULT_AUDIO_PATH}")
|
616 |
|
617 |
-
try:
|
618 |
-
|
619 |
-
|
620 |
-
|
621 |
-
|
622 |
-
|
623 |
-
|
624 |
-
|
625 |
-
|
626 |
-
|
627 |
-
|
628 |
-
|
629 |
-
|
630 |
-
|
631 |
-
|
632 |
-
|
633 |
-
|
634 |
-
|
635 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
636 |
|
637 |
-
|
638 |
-
|
639 |
-
|
640 |
-
|
641 |
-
len(
|
642 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
643 |
)
|
644 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
645 |
video_embedding = video_embedding[:min_len]
|
|
|
646 |
video_landmarks = video_landmarks[:min_len]
|
647 |
-
raw_audio = raw_audio[:min_len]
|
648 |
-
hubert_embedding = hubert_embedding[:min_len]
|
649 |
-
wavlm_embedding = wavlm_embedding[:min_len]
|
650 |
-
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
651 |
|
652 |
else:
|
653 |
-
#
|
654 |
-
|
655 |
-
print("Computing video embeddings and landmarks")
|
656 |
-
video_reader = decord.VideoReader(video_input)
|
657 |
-
decord.bridge.set_bridge("torch")
|
658 |
-
|
659 |
-
if not audio_cache_hit:
|
660 |
-
# Need to process audio to determine min_len
|
661 |
-
raw_audio = get_raw_audio(audio_input, 16000)
|
662 |
-
if len(raw_audio) == 0 or len(video_reader) == 0:
|
663 |
-
raise ValueError("Empty audio or video input")
|
664 |
-
|
665 |
-
min_len = min(len(raw_audio), len(video_reader))
|
666 |
-
|
667 |
-
# Store full audio in cache
|
668 |
-
cache["audio"]["path"] = audio_path_hash
|
669 |
-
cache["audio"]["raw_audio"] = raw_audio.clone()
|
670 |
-
|
671 |
-
# Create truncated copy for processing
|
672 |
-
raw_audio = raw_audio[:min_len]
|
673 |
-
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
674 |
-
else:
|
675 |
-
# Use cached audio - make a copy
|
676 |
-
if cache["audio"]["raw_audio"] is None:
|
677 |
-
raise ValueError("Cached audio is None")
|
678 |
-
|
679 |
-
raw_audio = cache["audio"]["raw_audio"].clone()
|
680 |
-
if len(raw_audio) == 0 or len(video_reader) == 0:
|
681 |
-
raise ValueError("Empty cached audio or video input")
|
682 |
-
|
683 |
-
min_len = min(len(raw_audio), len(video_reader))
|
684 |
-
|
685 |
-
# Create truncated copy for processing
|
686 |
-
raw_audio = raw_audio[:min_len]
|
687 |
-
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
688 |
-
|
689 |
-
# Compute video embeddings and landmarks - store full version in cache
|
690 |
-
video_embedding, video_frames = compute_video_embedding(
|
691 |
-
video_reader, len(video_reader)
|
692 |
-
)
|
693 |
-
video_landmarks = extract_video_landmarks(video_frames)
|
694 |
|
695 |
-
|
696 |
-
cache["video"]["
|
697 |
-
cache["video"]["
|
698 |
-
cache["video"]["
|
699 |
-
|
|
|
700 |
|
701 |
-
|
702 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
703 |
video_frames = video_frames[:min_len]
|
|
|
704 |
video_landmarks = video_landmarks[:min_len]
|
705 |
-
|
706 |
else:
|
707 |
-
#
|
708 |
-
|
709 |
-
|
710 |
-
|
711 |
-
|
712 |
-
|
713 |
-
|
714 |
-
|
715 |
-
|
716 |
-
|
717 |
-
|
718 |
-
|
719 |
-
|
720 |
-
|
721 |
-
|
722 |
-
|
723 |
-
|
724 |
-
|
725 |
-
|
726 |
-
|
727 |
-
|
728 |
-
|
729 |
-
|
730 |
-
video_landmarks = cache["video"]["landmarks"].copy()
|
731 |
-
|
732 |
-
# Determine truncation length and create truncated copies
|
733 |
-
min_len = min(len(raw_audio), len(video_frames))
|
734 |
-
raw_audio = raw_audio[:min_len]
|
735 |
-
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
736 |
-
video_frames = video_frames[:min_len]
|
737 |
-
video_embedding = video_embedding[:min_len]
|
738 |
-
video_landmarks = video_landmarks[:min_len]
|
739 |
-
else:
|
740 |
-
# Both video and audio are cached - should not reach here
|
741 |
-
# as it's handled in the first if statement
|
742 |
-
pass
|
743 |
-
|
744 |
-
# Process audio if needed
|
745 |
-
if not audio_cache_hit:
|
746 |
-
print("Computing audio embeddings")
|
747 |
-
|
748 |
-
# Compute audio embeddings with the truncated audio
|
749 |
-
hubert_embedding = compute_hubert_embedding(raw_audio_reshape)
|
750 |
-
wavlm_embedding = compute_wavlm_embedding(raw_audio_reshape)
|
751 |
|
752 |
-
|
753 |
-
|
754 |
-
cache["audio"]["hubert_embedding"] = hubert_embedding.clone()
|
755 |
-
cache["audio"]["wavlm_embedding"] = wavlm_embedding.clone()
|
756 |
-
else:
|
757 |
-
# Use cached audio data - make copies
|
758 |
-
if (
|
759 |
-
cache["audio"]["hubert_embedding"] is None
|
760 |
-
or cache["audio"]["wavlm_embedding"] is None
|
761 |
-
):
|
762 |
-
raise ValueError(
|
763 |
-
"One or more audio embedding cache entries are None"
|
764 |
-
)
|
765 |
|
766 |
-
|
767 |
-
|
768 |
-
|
769 |
-
|
770 |
-
|
771 |
-
|
772 |
-
|
773 |
-
|
774 |
-
|
775 |
-
|
776 |
-
|
777 |
-
|
778 |
-
|
779 |
-
|
780 |
-
|
781 |
-
|
782 |
-
|
783 |
-
|
784 |
-
|
785 |
-
|
786 |
-
|
787 |
-
|
788 |
-
|
789 |
-
|
790 |
-
|
791 |
-
|
792 |
-
|
793 |
-
|
794 |
-
|
795 |
-
|
796 |
-
|
797 |
-
assert video_embedding.shape[0] == video_landmarks.shape[0], (
|
798 |
-
f"Video embedding length ({video_embedding.shape[0]}) doesn't match landmarks length ({video_landmarks.shape[0]})"
|
799 |
-
)
|
800 |
|
801 |
-
|
802 |
-
|
803 |
-
|
804 |
-
|
805 |
-
|
806 |
-
|
807 |
-
|
808 |
-
|
809 |
-
|
810 |
-
|
811 |
-
|
812 |
-
|
813 |
-
|
814 |
-
|
815 |
-
|
816 |
-
|
817 |
-
|
818 |
-
|
819 |
-
|
820 |
-
|
821 |
-
|
822 |
-
|
823 |
-
|
824 |
-
|
825 |
-
|
826 |
-
|
827 |
-
|
828 |
-
|
829 |
|
830 |
-
|
831 |
-
|
832 |
-
|
833 |
-
|
834 |
-
|
835 |
-
|
836 |
-
|
837 |
-
|
838 |
-
|
839 |
-
|
840 |
-
|
841 |
-
|
842 |
-
|
843 |
-
|
844 |
-
|
845 |
-
|
846 |
-
|
847 |
-
|
848 |
-
|
849 |
|
850 |
-
|
851 |
-
raw_audio[: complete_video.shape[0]], "f s -> () (f s)"
|
852 |
-
)
|
853 |
|
854 |
-
|
855 |
-
|
856 |
-
|
857 |
|
858 |
-
|
859 |
|
860 |
-
|
861 |
-
|
862 |
-
|
863 |
|
864 |
-
except Exception as e:
|
865 |
-
|
866 |
-
|
867 |
-
|
868 |
|
869 |
|
870 |
def get_max_duration(video_input, audio_input):
|
|
|
614 |
audio_input = DEFAULT_AUDIO_PATH
|
615 |
print(f"Using default audio: {DEFAULT_AUDIO_PATH}")
|
616 |
|
617 |
+
# try:
|
618 |
+
# Calculate hashes for cache keys
|
619 |
+
video_path_hash = video_input
|
620 |
+
audio_path_hash = audio_input
|
621 |
+
|
622 |
+
# Check if we need to recompute video embeddings
|
623 |
+
video_cache_hit = cache["video"]["path"] == video_path_hash
|
624 |
+
audio_cache_hit = cache["audio"]["path"] == audio_path_hash
|
625 |
+
|
626 |
+
if video_cache_hit and audio_cache_hit:
|
627 |
+
print("Using cached video and audio computations")
|
628 |
+
# Make copies of cached data to avoid modifying cache
|
629 |
+
video_embedding = cache["video"]["embedding"].clone()
|
630 |
+
video_frames = cache["video"]["frames"].clone()
|
631 |
+
video_landmarks = cache["video"]["landmarks"].copy()
|
632 |
+
raw_audio = cache["audio"]["raw_audio"].clone()
|
633 |
+
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
634 |
+
hubert_embedding = cache["audio"]["hubert_embedding"].clone()
|
635 |
+
wavlm_embedding = cache["audio"]["wavlm_embedding"].clone()
|
636 |
+
|
637 |
+
# Ensure all data is truncated to the same length if needed
|
638 |
+
min_len = min(
|
639 |
+
len(video_frames),
|
640 |
+
len(raw_audio),
|
641 |
+
len(hubert_embedding),
|
642 |
+
len(wavlm_embedding),
|
643 |
+
)
|
644 |
+
video_frames = video_frames[:min_len]
|
645 |
+
video_embedding = video_embedding[:min_len]
|
646 |
+
video_landmarks = video_landmarks[:min_len]
|
647 |
+
raw_audio = raw_audio[:min_len]
|
648 |
+
hubert_embedding = hubert_embedding[:min_len]
|
649 |
+
wavlm_embedding = wavlm_embedding[:min_len]
|
650 |
+
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
651 |
+
|
652 |
+
else:
|
653 |
+
# Process video if needed
|
654 |
+
if not video_cache_hit:
|
655 |
+
print("Computing video embeddings and landmarks")
|
656 |
+
video_reader = decord.VideoReader(video_input)
|
657 |
+
decord.bridge.set_bridge("torch")
|
658 |
+
|
659 |
+
if not audio_cache_hit:
|
660 |
+
# Need to process audio to determine min_len
|
661 |
+
raw_audio = get_raw_audio(audio_input, 16000)
|
662 |
+
if len(raw_audio) == 0 or len(video_reader) == 0:
|
663 |
+
raise ValueError("Empty audio or video input")
|
664 |
+
|
665 |
+
min_len = min(len(raw_audio), len(video_reader))
|
666 |
+
|
667 |
+
# Store full audio in cache
|
668 |
+
cache["audio"]["path"] = audio_path_hash
|
669 |
+
cache["audio"]["raw_audio"] = raw_audio.clone()
|
670 |
+
|
671 |
+
# Create truncated copy for processing
|
672 |
+
raw_audio = raw_audio[:min_len]
|
673 |
+
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
674 |
+
else:
|
675 |
+
# Use cached audio - make a copy
|
676 |
+
if cache["audio"]["raw_audio"] is None:
|
677 |
+
raise ValueError("Cached audio is None")
|
678 |
|
679 |
+
raw_audio = cache["audio"]["raw_audio"].clone()
|
680 |
+
if len(raw_audio) == 0 or len(video_reader) == 0:
|
681 |
+
raise ValueError("Empty cached audio or video input")
|
682 |
+
|
683 |
+
min_len = min(len(raw_audio), len(video_reader))
|
684 |
+
|
685 |
+
# Create truncated copy for processing
|
686 |
+
raw_audio = raw_audio[:min_len]
|
687 |
+
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
688 |
+
|
689 |
+
# Compute video embeddings and landmarks - store full version in cache
|
690 |
+
video_embedding, video_frames = compute_video_embedding(
|
691 |
+
video_reader, len(video_reader)
|
692 |
)
|
693 |
+
video_landmarks = extract_video_landmarks(video_frames)
|
694 |
+
|
695 |
+
# Update video cache with full versions
|
696 |
+
cache["video"]["path"] = video_path_hash
|
697 |
+
cache["video"]["embedding"] = video_embedding
|
698 |
+
cache["video"]["frames"] = video_frames
|
699 |
+
cache["video"]["landmarks"] = video_landmarks
|
700 |
+
|
701 |
+
# Create truncated copies for processing
|
702 |
video_embedding = video_embedding[:min_len]
|
703 |
+
video_frames = video_frames[:min_len]
|
704 |
video_landmarks = video_landmarks[:min_len]
|
|
|
|
|
|
|
|
|
705 |
|
706 |
else:
|
707 |
+
# Use cached video data - make copies
|
708 |
+
print("Using cached video computations")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
709 |
|
710 |
+
if (
|
711 |
+
cache["video"]["embedding"] is None
|
712 |
+
or cache["video"]["frames"] is None
|
713 |
+
or cache["video"]["landmarks"] is None
|
714 |
+
):
|
715 |
+
raise ValueError("One or more video cache entries are None")
|
716 |
|
717 |
+
if not audio_cache_hit:
|
718 |
+
# New audio with cached video
|
719 |
+
raw_audio = get_raw_audio(audio_input, 16000)
|
720 |
+
if len(raw_audio) == 0:
|
721 |
+
raise ValueError("Empty audio input")
|
722 |
+
|
723 |
+
# Store full audio in cache
|
724 |
+
cache["audio"]["path"] = audio_path_hash
|
725 |
+
cache["audio"]["raw_audio"] = raw_audio.clone()
|
726 |
+
|
727 |
+
# Make copies of video data
|
728 |
+
video_embedding = cache["video"]["embedding"].clone()
|
729 |
+
video_frames = cache["video"]["frames"].clone()
|
730 |
+
video_landmarks = cache["video"]["landmarks"].copy()
|
731 |
+
|
732 |
+
# Determine truncation length and create truncated copies
|
733 |
+
min_len = min(len(raw_audio), len(video_frames))
|
734 |
+
raw_audio = raw_audio[:min_len]
|
735 |
+
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
736 |
video_frames = video_frames[:min_len]
|
737 |
+
video_embedding = video_embedding[:min_len]
|
738 |
video_landmarks = video_landmarks[:min_len]
|
|
|
739 |
else:
|
740 |
+
# Both video and audio are cached - should not reach here
|
741 |
+
# as it's handled in the first if statement
|
742 |
+
pass
|
743 |
+
|
744 |
+
# Process audio if needed
|
745 |
+
if not audio_cache_hit:
|
746 |
+
print("Computing audio embeddings")
|
747 |
+
|
748 |
+
# Compute audio embeddings with the truncated audio
|
749 |
+
hubert_embedding = compute_hubert_embedding(raw_audio_reshape)
|
750 |
+
wavlm_embedding = compute_wavlm_embedding(raw_audio_reshape)
|
751 |
+
|
752 |
+
# Update audio cache with full embeddings
|
753 |
+
# Note: raw_audio was already cached above
|
754 |
+
cache["audio"]["hubert_embedding"] = hubert_embedding.clone()
|
755 |
+
cache["audio"]["wavlm_embedding"] = wavlm_embedding.clone()
|
756 |
+
else:
|
757 |
+
# Use cached audio data - make copies
|
758 |
+
if (
|
759 |
+
cache["audio"]["hubert_embedding"] is None
|
760 |
+
or cache["audio"]["wavlm_embedding"] is None
|
761 |
+
):
|
762 |
+
raise ValueError("One or more audio embedding cache entries are None")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
763 |
|
764 |
+
hubert_embedding = cache["audio"]["hubert_embedding"].clone()
|
765 |
+
wavlm_embedding = cache["audio"]["wavlm_embedding"].clone()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
766 |
|
767 |
+
# Make sure embeddings match the truncated video length if needed
|
768 |
+
if "min_len" in locals() and (
|
769 |
+
min_len < len(hubert_embedding) or min_len < len(wavlm_embedding)
|
770 |
+
):
|
771 |
+
hubert_embedding = hubert_embedding[:min_len]
|
772 |
+
wavlm_embedding = wavlm_embedding[:min_len]
|
773 |
+
|
774 |
+
# Apply max_num_seconds limit if specified
|
775 |
+
if max_num_seconds > 0:
|
776 |
+
# Convert seconds to frames (assuming 25 fps)
|
777 |
+
max_frames = int(max_num_seconds * 25)
|
778 |
+
|
779 |
+
# Truncate all data to max_frames
|
780 |
+
video_embedding = video_embedding[:max_frames]
|
781 |
+
video_frames = video_frames[:max_frames]
|
782 |
+
video_landmarks = video_landmarks[:max_frames]
|
783 |
+
hubert_embedding = hubert_embedding[:max_frames]
|
784 |
+
wavlm_embedding = wavlm_embedding[:max_frames]
|
785 |
+
raw_audio = raw_audio[:max_frames]
|
786 |
+
raw_audio_reshape = rearrange(raw_audio, "f s -> (f s)")
|
787 |
+
|
788 |
+
# Validate shapes before proceeding
|
789 |
+
assert video_embedding.shape[0] == hubert_embedding.shape[0], (
|
790 |
+
f"Video embedding length ({video_embedding.shape[0]}) doesn't match Hubert embedding length ({hubert_embedding.shape[0]})"
|
791 |
+
)
|
792 |
+
assert video_embedding.shape[0] == wavlm_embedding.shape[0], (
|
793 |
+
f"Video embedding length ({video_embedding.shape[0]}) doesn't match WavLM embedding length ({wavlm_embedding.shape[0]})"
|
794 |
+
)
|
795 |
+
assert video_embedding.shape[0] == video_landmarks.shape[0], (
|
796 |
+
f"Video embedding length ({video_embedding.shape[0]}) doesn't match landmarks length ({video_landmarks.shape[0]})"
|
797 |
+
)
|
|
|
|
|
|
|
798 |
|
799 |
+
print(f"Hubert embedding shape: {hubert_embedding.shape}")
|
800 |
+
print(f"WavLM embedding shape: {wavlm_embedding.shape}")
|
801 |
+
print(f"Video embedding shape: {video_embedding.shape}")
|
802 |
+
print(f"Video landmarks shape: {video_landmarks.shape}")
|
803 |
+
|
804 |
+
# Create pipeline inputs for models
|
805 |
+
(
|
806 |
+
interpolation_chunks,
|
807 |
+
keyframe_chunks,
|
808 |
+
audio_interpolation_chunks,
|
809 |
+
audio_keyframe_chunks,
|
810 |
+
emb_cond,
|
811 |
+
masks_keyframe_chunks,
|
812 |
+
masks_interpolation_chunks,
|
813 |
+
to_remove,
|
814 |
+
audio_interpolation_idx,
|
815 |
+
audio_keyframe_idx,
|
816 |
+
) = create_pipeline_inputs(
|
817 |
+
hubert_embedding,
|
818 |
+
wavlm_embedding,
|
819 |
+
14,
|
820 |
+
video_embedding,
|
821 |
+
video_landmarks,
|
822 |
+
overlap=1,
|
823 |
+
add_zero_flag=True,
|
824 |
+
mask_arms=None,
|
825 |
+
nose_index=28,
|
826 |
+
)
|
827 |
|
828 |
+
complete_video = sample(
|
829 |
+
audio_keyframe_chunks,
|
830 |
+
keyframe_chunks,
|
831 |
+
masks_keyframe_chunks,
|
832 |
+
to_remove,
|
833 |
+
audio_keyframe_idx,
|
834 |
+
14,
|
835 |
+
"cuda",
|
836 |
+
emb_cond,
|
837 |
+
[],
|
838 |
+
3,
|
839 |
+
3,
|
840 |
+
audio_interpolation_idx,
|
841 |
+
audio_interpolation_chunks,
|
842 |
+
masks_interpolation_chunks,
|
843 |
+
interpolation_chunks,
|
844 |
+
keyframe_model,
|
845 |
+
interpolation_model,
|
846 |
+
)
|
847 |
|
848 |
+
complete_audio = rearrange(raw_audio[: complete_video.shape[0]], "f s -> () (f s)")
|
|
|
|
|
849 |
|
850 |
+
# 4. Convert frames to video and combine with audio
|
851 |
+
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_video:
|
852 |
+
output_path = temp_video.name
|
853 |
|
854 |
+
print("Saving video to", output_path)
|
855 |
|
856 |
+
save_audio_video(complete_video, audio=complete_audio, save_path=output_path)
|
857 |
+
torch.cuda.empty_cache()
|
858 |
+
return output_path
|
859 |
|
860 |
+
# except Exception as e:
|
861 |
+
# raise e
|
862 |
+
# print(f"Error processing video: {str(e)}")
|
863 |
+
# return None
|
864 |
|
865 |
|
866 |
def get_max_duration(video_input, audio_input):
|