Add ONNX model (fp32).

README.md

@@ -1,3 +1,56 @@
 ---
+library_name: transformers
+pipeline_tag: image-segmentation
+tags:
+  - image-segmentation
+  - mask-generation
+  - transformers.js
 license: apache-2.0
+language:
+  - en
 ---
+# U-2-Net
+
+## Model Description
+U-2-Net is a deep learning model designed for image segmentation tasks, particularly for generating detailed masks. It leverages a unique architecture with nested U-blocks that allow the model to capture both high-level semantic features and fine-grained details. U-2-Net has demonstrated high performance in a variety of segmentation tasks, making it a versatile choice for applications such as background removal, object detection, and medical image analysis.
+
+## Usage
+Perform mask generation with `BritishWerewolf/U-2-Net`.
+
+### Example
+```javascript
+import { AutoModel, AutoProcessor, RawImage } from '@huggingface/transformers';
+
+const img_url = 'https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png';
+const image = await RawImage.read(img_url);
+
+const processor = await AutoProcessor.from_pretrained('BritishWerewolf/U-2-Net');
+const processed = await processor(image);
+
+const model = await AutoModel.from_pretrained('BritishWerewolf/U-2-Net', {
+    dtype: 'fp32',
+});
+
+const output = await model({ input: processed.pixel_values });
+// {
+//   mask: Tensor {
+//     dims: [ 1, 320, 320 ],
+//     type: 'uint8',
+//     data: Uint8Array(102400) [ ... ],
+//     size: 102400
+//   }
+// }
+```
+
+## Model Architecture
+The U-2-Net model is built upon a nested U-structure, where each U-block consists of multiple convolutional layers, pooling, and up-sampling operations. The architecture features a combination of down-sampling and up-sampling paths, enabling the model to learn features at different scales. This design allows the U-2-Net to produce accurate and high-resolution segmentation maps. The key components of the architecture include Residual U-blocks (RSU) that enhance feature representation and ensure efficient information flow through the network.
+
+### Inference
+To use the model for inference, you can follow the example provided above. The `AutoProcessor` and `AutoModel` classes from the `transformers` library make it easy to load the model and processor.
+
+## Credits
+* [`rembg`](https://github.com/danielgatis/rembg) for the ONNX model.
+* The authors of the original U-2-Net model can be credited at https://github.com/xuebinqin/U-2-Net.
+
+## Licence
+This model is licensed under the Apache License 2.0 to match the original U-2-Net model.

config.json

@@ -0,0 +1,20 @@
+{
+  "_name_or_path": "u2net",
+  "model_type": "u2net",
+  "architectures": [
+    "U2NetModel"
+  ],
+  "input_name": ["input.1"],
+  "input_shape": [1, 3, 320, 320],
+  "output_composite": "1959",
+  "output_names": [
+    "1959",
+    "1960",
+    "1961",
+    "1962",
+    "1963",
+    "1964",
+    "1965"
+  ],
+  "output_shape": [1, 320, 320]
+}

onnx/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d10d2f3bb75ae3b6d527c77944fc5e7dcd94b29809d47a739a7a728a912b491
+size 175997641

preprocessor_config.json

@@ -0,0 +1,27 @@
+{
+  "processor_class": "U2NetProcessor",
+  "image_processor_type": "U2NetImageProcessor",
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_pad": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "keep_aspect_ratio": true,
+  "image_mean": [
+    0.485,
+    0.456,
+    0.406
+  ],
+  "image_std": [
+    0.229,
+    0.224,
+    0.225
+  ],
+  "pad_size": {
+    "width": 320,
+    "height": 320
+  },
+  "size": {
+    "longest_edge": 320
+  }
+}