Spaces:
Running on Zero
Running on Zero
Commit History
-
docs: note that the Jan-2025 checkpoint is outdated; point to mippia.com for the production service Running
a520582 -
fix: pin torch/torchaudio 2.8.0 for ZeroGPU compatibility
b154efd
-
fix: honor torchaudio.load(channels_first=False) in patched loader
slslslrhfembeat_this calls torchaudio.load(path, channels_first=False) expecting (time, channels). Our librosa-based replacement always returned (channels, time), so beat_this's signal.mean(1) collapsed the time axis to a tiny vector and torch.stft crashed with padding (512, 512) at dimension 2 of input [1, 1, 1]. Local repro confirms: 10 s mono and 10 s stereo files now both pass through File2Beats successfully.a255c88 -
fix: reject audio shorter than 1.5s with a readable message
slslslrhfemRoot cause confirmed by local repro: beat_this resamples to 22050 Hz and runs torch.stft with n_fft=1024 / pad=512. Any clip shorter than ~1 s crashes with 'Padding size should be less than the corresponding input dimension, but got: padding (512, 512) at dimension 2 of input [1, 1, X]'. The previous 1024-sample guard was too lax (it triggered only below ~43 ms at 24 kHz). Bump to 1.5 s, expressed in seconds so the message makes sense to the user.897dd20 -
debug: surface full traceback in inference error UI
slslslrhfemNeed full stack trace to locate where the [1,1,1] tensor + 512 padding crash happens. Logs to console as well as the rendered output card.8314770 -
fix: use librosa.load as the single audio loader path
slslslrhfemsoundfile silently returned 1-frame results for some compressed inputs, producing [1,1,1] tensors that crashed downstream STFT padding with 'Padding size should be less than the corresponding input dimension'. librosa.load goes through soundfile -> audioread fallback and is robust across wav/flac/mp3/m4a/ogg. Also add an explicit too-short guard so edge inputs surface a readable message instead of a padding error.4986065 -
fix: bump torch/torchaudio to 2.6.0 (CVE-2025-32434 mitigation)
slslslrhfemtorch 2.5.1 is rejected by transformers/lightning torch.load safety check due to CVE-2025-32434. 2.6.0 is the minimum patched version and still uses CUDA 12.4 wheels (libcudart.so.12), staying compatible with the HF Space base image.13d5fc5 -
fix: pin torch/torchaudio to 2.5.1 (CUDA 12 compatible)
slslslrhfemThe unpinned >=2.0.0 was resolving to torch 2.9+ which links against libcudart.so.13 (CUDA 13). HF Space base image only ships CUDA 12, so import torchaudio failed with 'libcudart.so.13: cannot open shared object file'. Pin to 2.5.1 which uses cu121 wheels.c2fe02d -
fix: bypass torchaudio backend dispatch for torchaudio 2.9+
slslslrhfemtorchaudio.load now routes through load_with_torchcodec and ignores the backend= kwarg, so the previous soundfile-backend patch failed with 'TorchCodec is required'. Replace torchaudio.load / torchaudio.info with direct soundfile-based implementations to bypass backend dispatch entirely. Falls back to librosa for formats soundfile cannot decode.d75df89
-
fix: guard torchaudio.info patch with hasattr check for torchaudio 2.9+
mineeuk824783f -
fix: remove torchcodec, add ffmpeg system pkg, monkey-patch torchaudio to use soundfile backend
mineeuk67d992f -
fix: correct hardware name to zero-a10g
mippiaa7ad714 -
fix: add torchcodec dep, switch to zero-a10g-small GPU for speed
mippiae6fef3d -
fix: auto-detect CUDA/CPU device in get_segments_from_wav
mippia5ee2577 -
fix: use local beat_this checkpoint to avoid download failure in HF Space
mippiad209d6f -
fix: pin transformers>=4.46.0 and huggingface_hub to resolve is_offline_mode ImportError
mippia5291112
-
change color for darkmode
slslslrhfem8a4fa1b
-
change probability func
slslslrhfem693d2c7 -
change probability func
slslslrhfemc410f34 -
change probability func
slslslrhfema2657b4 -
change checkpoints
slslslrhfem6da214d
-
fix some model
slslslrhfem508344f -
fix some model
slslslrhfem617e256 -
add hf token
slslslrhfem035eb8a
-
commit for ICASSP 2026 github repository
slslslrhfembe8ccdd
-
change description
Seonghyeon Go216b804 -
change description
Seonghyeon Go01e4765 -
change desing and result
Seonghyeon Gof62dfd2 -
change import index
Seonghyeon Goacb0b29 -
delete examples
Seonghyeon Go50d7dc5 -
fix some typo
Seonghyeon Go8f94ca9 -
move audio files to Git LFS
Seonghyeon Go7ce7d3b -
add spaces decorator
Seonghyeon Goe836611 -
initial commit for AIGM
Seonghyeon Go0ede85b -
initial segment transformer
Seonghyeon Goc3c908f -
initial commit
seonghyeon gof1a54b8