xclip-onnx

Two ONNX graphs exported from microsoft/xclip-base-patch16-kinetics-600 (MIT license, code and weights) at revision e4921c41fc296102aae210d43d4127c5e3e51928, for use through wasi:nn where no PyTorch runtime is available. Used by ffrwd/describe.

There is no official or community ONNX export of X-CLIP; these are produced by scripts/export_xclip.py in that repository, opset 17, no dynamic axes, and validated against the PyTorch model on export (max absolute diff ~1e-6, float32 rounding noise).

Files

file	tower	input	output
video_tower.onnx	get_video_features	pixel_values [1, 8, 3, 224, 224] fp32	video_embeds [1, 512]
text_tower.onnx	get_text_features	input_ids, attention_mask, both [1, 77] i64	text_embeds [1, 512]

Both towers' output is the model's pooler_output field, which X-CLIP overwrites with the projected embedding (not the tower's own pooled hidden state) - the field the export script pulls explicitly.

License

MIT, following the source model.