Spaces:
Running on T4
Running on T4
Commit History
-
Memoise a video's features across the loader's repeated requests (single encode per job) Running
Verified signature76266ba -
Run feature extractors in-process (data.num_workers=0): forked DataLoader workers cannot use CUDA
Verified signature25c26a7 -
Batched loop: import tqdm
Verified signature6a629c2 -
Encode two clips per VJEPA2 forward (batched fp16 loop replaces per-clip loop)
Verified signatureed00044 -
Decode each video once (no per-frame ffmpeg seeks), VJEPA2 forward in fp16, trim at 16 fps
Verified signature1459d2a -
Fix WhisperX (keyword call, auto language, prewarm), fp16 video extractor, fp16 text model via accelerate
Verified signature6e48ea5 -
Add analyze_video_async: background analysis with callback to AdVisor, single GPU lock
Verified signaturea8ae6d3
-
Accept ad context for tailored insights (brand, CTA, USP)
Verified signaturea5fa6d8 -
Rewrite insights: data-driven analysis instead of generic templates
Verified signature13b1120 -
Fix: report seconds now match actual video time (removed HRF offset)
Verified signaturedc48354 -
Support up to 60s videos
Verified signatureb2bde86 -
Trim to 10s for faster processing (~20 encoder chunks instead of 30)
Verified signature06cbebe -
Creative strategy insights: translate brain regions to actionable ad advice
Verified signatureb90c03f -
Fix scoring: z-score normalization, artifact correction, better insights
Verified signature23e3186 -
Trim to 15s at 2fps 320px — ~4x faster encoding
Verified signaturef606156 -
Cap video at 30s + downscale to 480p for 10x faster processing
Verified signaturec66c3e7 -
Use full TRIBE v2 pipeline with WhisperX on T4 GPU
Verified signature77c97fe -
Complete rewrite: bypass WhisperX entirely, use audio_only mode
UserWhisperX runs in a uvx subprocess with hardcoded float16 that requires GPU. No amount of patching/env vars can fix a subprocess's hardcoded settings. Solution: Skip get_events_dataframe() entirely. Build the events DataFrame manually using only ExtractAudioFromVideo + ChunkEvents (no transcription). TRIBE v2 still predicts brain responses from video + audio features — word-level transcription adds language processing but isn't required. This makes the Space work on CPU Basic (free) without GPU.c99eb91 -
Fix: monkey-patch WhisperX to use int8 on CPU (float16 requires GPU)
User6de2052 -
Disable HF_HUB_ENABLE_HF_TRANSFER — uvx sandbox lacks the package
User9128f18 -
Fix atlas labels unpacking — handle all nilearn formats
Userc64d80c -
Fix: add error handling, ensure whisperx, set HF_TOKEN for Llama access
User2c73e57
-
Add TRIBE v2 neural engagement analyzer app
User7d1d2c9