Repository Issue Activity (beta)

#1205 Documentation proposal: stress-test OpenCLIP similarity under JPEG compression and downscaling

#1201 Audio-text retrieval eval (AudioCaps/Clotho) and audio front-end unit tests

#1012 error when use try to use int8 operations with OpenCLIP.

#921 How to persist int8 model?

#1199 CI silently broken by transformers>=5.14.0 (GenerationMixin.generate() now requires PreTrainedModel-only get_experts_implementation()); venv cache is masking it on main

#516 Dimension mismatch when using Coca for VQA task

#458 Issue with number of tokens for CoCa

#554 CoCa v2: fixes and improvements

#761 Gradient accumulation may requires scaling before backward

#1191 Roadmap of new release?

#1041 Request for SigLIP2 Training Code and Pretrained Weights

#1144 [Bug/Question] Unexpected behavior and redundant computation with gather_with_grad and local_loss flag combinations

#304 New feature: text/text contrastive

#521 Entry Not Found Error for JSON file when using OpenCLIP ViT-H/14

#436 Scaling validation to large datasets

#957 Apply scale loss when performing accum_freq

#822 oscillations in the loss

#1167 Validation loss behaves abnormally while downstream classification metrics are normal when fine-tuning CLIP on chest X-ray images

#1162 All open_clip_torch failed to load convnext clip

#384 MuLaN

#849 CPU memory leak when training with CsvDataset-like dataset

#310 FDSP

#1148 factory.load_state_dict: dead except TypeError fallback drops weights_only

#1149 openai.py: torch.load fallback missing weights_only=True

#1150 pt_load: training --resume path loads pickle without weights_only