DeblurGANV2Demo
š
1
ClearIfy Image Project
Dan PRO
Daankular
AI & ML interests
None yet
Recent Activity
liked a Space about 1 hour ago
AIDemoProject/DeblurGANV2Demo liked a model about 2 hours ago
FireRedTeam/FireRed-Image-Edit-1.0 liked a dataset about 3 hours ago
ajehsmihba/aesthetic-female-portraitsOrganizations
liked a
Space about 1 hour ago
liked a
model about 2 hours ago
liked a
dataset about 3 hours ago
reacted to matteospanio's post with š about 6 hours ago
Post
4326
š¶ Released mule-torch ā an unofficial PyTorch port of MULE (SF-NFNet-F0), SiriusXM/Pandora's music-audio embedding model (McCallum et al., ISMIR 2022).
No retraining: I re-implemented the architecture in pure PyTorch and transferred the original TensorFlow weights, then checked it layer by layer against the genuine TF pipeline.
ā End-to-end clip-embedding cosine 0.9999999 vs the original
ā ONNX backbone parity < 1e-6
ā 62.35M params (paper: ~62.4M)
ā Batched, GPU-native, ONNX-exportable ā none of which the original
š¤ Weights: matteospanio/mule
š» Code: https://github.com/matteospanio/mule-torch
š¦ PyPI: https://pypi.org/project/mule-torch/
The fun bug: parity was perfect through every conv but the block output was anti-correlated (cos = ā1). Cause: the learnable skip-init gains couldn't be mapped by layer name (Keras scrambles the order) ā they had to be recovered from the graph.
ā ļø Unofficial, community port ā not affiliated with or endorsed by the original authors. All credit to them; please cite the paper. Weights inherit CC-BY-NC-4.0.
No retraining: I re-implemented the architecture in pure PyTorch and transferred the original TensorFlow weights, then checked it layer by layer against the genuine TF pipeline.
ā End-to-end clip-embedding cosine 0.9999999 vs the original
ā ONNX backbone parity < 1e-6
ā 62.35M params (paper: ~62.4M)
ā Batched, GPU-native, ONNX-exportable ā none of which the original
Analysis pipeline doespip install mule-torchfrom mule_torch import MuleModel
emb = MuleModel.from_pretrained()(waveform) # (B, T)@16kHz -> (B, 1728)š¤ Weights: matteospanio/mule
š» Code: https://github.com/matteospanio/mule-torch
š¦ PyPI: https://pypi.org/project/mule-torch/
The fun bug: parity was perfect through every conv but the block output was anti-correlated (cos = ā1). Cause: the learnable skip-init gains couldn't be mapped by layer name (Keras scrambles the order) ā they had to be recovered from the graph.
ā ļø Unofficial, community port ā not affiliated with or endorsed by the original authors. All credit to them; please cite the paper. Weights inherit CC-BY-NC-4.0.
updated a
Space about 14 hours ago
liked 2
Spaces about 15 hours ago
liked a
Space about 16 hours ago
liked a
dataset 1 day ago
liked a
model 1 day ago
reacted to AxionLab-official's post with š„ 1 day ago
Post
10416
THIS IS CRAZY! THE MODEL ON THE IMAGE(Supra-50M-Reasoning) answered correctly and its QUANTIZED IN 2BIT! THE RESPONSE IS CORRECT, IN A 15MB SIZE FILE!
- 13 replies
i did not get to use it once today
š 1
4
#1 opened 3 days ago
by 
Myzombielover
30 min of ZeroGPU gone
#2 opened 2 days ago
by Daankular
liked a
model 2 days ago
liked 2
models 4 days ago
liked 2
Spaces 5 days ago
liked a
model 5 days ago
