anudit commited on
Commit
146ae88
·
verified ·
1 Parent(s): 6634a2a

Add BiRefNet_lite-2K fp16 ONNX (static 2560x1440, low-memory graph)

Browse files
Files changed (2) hide show
  1. README.md +47 -0
  2. onnx/model_fp16.onnx +3 -0
README.md ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ library_name: onnx
4
+ pipeline_tag: image-segmentation
5
+ base_model: ZhengPeng7/BiRefNet_lite-2K
6
+ tags:
7
+ - background-removal
8
+ - dichotomous-image-segmentation
9
+ - onnx
10
+ - webgpu
11
+ - birefnet
12
+ ---
13
+
14
+ # BiRefNet_lite-2K — ONNX fp16 (static 2560×1440, low-memory graph)
15
+
16
+ ONNX export of [ZhengPeng7/BiRefNet_lite-2K](https://huggingface.co/ZhengPeng7/BiRefNet_lite-2K)
17
+ (Swin-T backbone, 44M params, trained for 2560×1440), made for in-browser inference with
18
+ [`birefnet-webgpu`](https://github.com/anudit/birefnet-webgpu).
19
+
20
+ | file | precision | input | output | size |
21
+ |---|---|---|---|---|
22
+ | `onnx/model_fp16.onnx` | fp16 weights, fp32 I/O | `input_image` `[1,3,1440,2560]` (ImageNet-normalised RGB) | `output_image` `[1,1,1440,2560]` logits (apply sigmoid) | ~99 MB |
23
+
24
+ The input size is **fixed**: resize (stretch, no aspect preservation — this matches training) your image
25
+ to 2560×1440, run, then resize the mask back.
26
+
27
+ ## What is different from a plain `torch.onnx.export`
28
+
29
+ A naive export needs ~11 GB of activations at this resolution, which is impossible in a browser tab.
30
+ This graph computes the same function with ~0.45 GB peak live activations (~1.1 GB allocated on GPU in
31
+ the WebGPU engine):
32
+
33
+ * Deformable convolutions use the standard ONNX **`DeformConv`** op (opset 19+) instead of an im2col gather chain.
34
+ * The 5-branch ASPP concat (1280 ch) followed by a 1×1 conv is rewritten as a sum of five 1×1 convs.
35
+ * The full-resolution decoder tail (bilinear upsample → concat → 1×1 conv, and the `ipt_blk1` branch) is
36
+ algebraically collapsed into a 1×1 conv at low resolution plus one 5×5 conv on the input image.
37
+ This is exact except for a ~1 px outer ring where zero-padding interacts with the folded convolutions
38
+ (≤ 3e-5 sigmoid difference on a test image).
39
+ * GELU is exported as the `Gelu` op (opset 20).
40
+
41
+ Validated against the original PyTorch model on a 2560×1440 image: fp32 graph IoU 0.999996, fp16 graph IoU 0.9995 (ORT CPU).
42
+ Requires a runtime that implements `DeformConv` and `Gelu` (verified with onnxruntime on CPU; the `birefnet-webgpu` engine on WebGPU).
43
+
44
+ ## License / credits
45
+
46
+ Weights and architecture: Zheng Peng et al., [BiRefNet](https://github.com/ZhengPeng7/BiRefNet) (MIT).
47
+ Please check the upstream card for the training-data (DIS5K) terms before commercial use.
onnx/model_fp16.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d03dc5de5f3295dfcf4b84829fbbcd52bc3adfac890965c7316b99d40739a02
3
+ size 98728016