Spaces:
Paused
Paused
Commit History
-
Update requirements.txt
Verified signatureUse RayCaiNZ/tribev2 fork with relaxed torch pin822d49d -
Update requirements.txt
Verified signaturePin torch>=2.7.0 for Blackwell sm_120 supporte586b08
-
Accept real Whisper word timestamps in predict_video
New chunks_to_events_dataframe() creates Language events from Whisper's return_timestamps=word output, using real per-word [start, end] times instead of synthetic 0.35s/word scaling. predict_video now accepts an optional chunks_json parameter (3rd arg). When provided, the language arm gets accurate speech timing — critical for content where visual/audio streams are flat (e.g. single-shot monologue) and the only signal is when specific words land. Fallback: if chunks_json is empty, the existing synthetic scaling path (text_to_events_dataframe + linear scale to video duration) is used, so this is fully backward-compatible. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>65ae3cb
-
Emit per-timestep brain curves for retention alignment
Adds compute_per_timestep_scores(preds) that keeps the T axis instead of collapsing it, returning per-second attention/emotion/memory/overall arrays plus per-region curves. format_result and predict_multimodal now include a `per_timestep` field so downstream tools can align brain activation against real retention curves and localize drop-off seconds. No endpoint signature changes; no extra GPU cost. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>048d65d -
Force DataLoader num_workers=0 under ZeroGPU daemon worker
ZeroGPU runs @spaces.GPU in a daemon subprocess which cannot spawn child processes. Torch DataLoader's default num_workers>0 triggers 'daemonic processes are not allowed to have children'. Monkey-patch DataLoader.__init__ to force num_workers=0 around predict calls. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>071f4c4 -
Pin python 3.12 — tribev2 requires >=3.11, ZeroGPU default was 3.10
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>cde1623
-
Fix event schema: use filepath field, extract audio to wav
neuralset's event coercion expects 'filepath' for Visual/Audio and 'word' for Language, not 'value'. Also extracts audio from video to a .wav file because soundfile.info cannot read mp4 containers. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>5ce277b
-
Fix: capitalize event types for neuralset standardize_events
neuralset/events/utils.py rejects lowercase event types with 'Legacy uncapitalized event' error. Changed visual→Visual, audio→Audio, language→Language. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>da73ed2 -
fix: bypass whisperx for video — construct events directly
predict_video now builds events dataframe without calling whisperx (which fails with float16 on ZeroGPU subprocess). Uses: - Visual scene events sampled every 0.5s from video duration - Audio event covering full video duration - Language events from optional pre-transcribed text Accepts optional transcript parameter so caller can provide Whisper output directly instead of TRIBEv2 trying to run whisperx.a4f81c4 -
feat: add multimodal support — audio, video, and combined input tabs
TRIBEv2 Space now supports 4 input modes: - Text: direct text-to-events (bypasses gTTS/whisperx) - Audio: voiceover/podcast via Wav2Vec - Video: full multimodal via V-JEPA2 + Wav2Vec + LLaMA - Multimodal: combine text + audio + video for richest prediction Each tab has its own @spaces.GPU allocation (120-300s). Legacy predict endpoint maintained for backward compatibility.b14b3bd -
Add timeline column to events dataframe
1f07d3e -
Add duration column to events dataframe
6e6f220 -
Fix: use start/stop columns for TRIBEv2 events
ffc65f2 -
Fix: use 'type' column name for TRIBEv2 events dataframe
eb8c001 -
Bypass gTTS+whisperx: construct events dataframe directly from text
5b68a36 -
Add edge-tts fallback for gTTS rate limits, retry logic for text input
b5db78a -
Fix: disable hf_transfer requirement for whisperx model download
5f48704 -
Add uv dependency for TRIBEv2 text processing (uvx)
1811124 -
TRIBEv2 brain prediction API with ZeroGPU
76a616a -
initial commit
Verified signature7604fb5