5 days ago
-
3D full finetuning: NDHWC train_graph.json + torch_params.json, structured weight_map to_onnx, 3D runtime; 2D spike peak_mb corrected
6 days ago
-
Browser finetuning: manifest finetune block (head; full/tfjs for 2D students), initializer offsets, weight maps, torch_version/exporter/opset, source_repo for 2D students
8 days ago
-
README: download links, fp32 vs fp16 WebGPU speeds
9 days ago
-
Model card: both graph layouts + browser WebGPU benchmark (Apple GPU, real EM)
-
All models: export Conv3d as exact Conv2d over depth-folded batch (~16x faster on onnxruntime-web WebGPU); drop experimental/
-
Unet3dIso: export 3D ConvTranspose as exact 1x1x1 Conv + depth-to-space (WebGPU EP has no 3D ConvTranspose)
-
Add exp_mito ONNX web models (6 setups x 3 block sizes x fp32/fp16/int8w/int8w_fp16)
