README.md

---
license: cc
---

## Install following python libs

```
pip3 install tensorflow
pip3 install tensorflowjs
pip3 install tf2onnx
pip3 install onnxruntime
pip3 install pillow
pip3 install optimum[exporters]
```
Change to compatible version of numpy for tensorflow
```
pip3 uninstall numpy
pip3 install numpy==1.23.5
```

## Node Install 
Download install project dependencies.
```
npm install
```

### Summary of Commands:

- Run the Node training script to save the Layers Model.
- Convert tfjs_layers_model → tfjs_graph_model
- Convert graph model to onnx
- Validate onnx structure
- Test Model

# 1. Create Tensorflow model in node

This will loop through the training images taking base folder name as the label for the images to be associated against.
Once complete saved-model/model.json is created.

```
node generate.js
```

# 2. Convert Model

Convert from layers to graph model this is required to generate an onnx from tf2onnx

```
tensorflowjs_converter --input_format=tfjs_layers_model \  --output_format=tfjs_graph_model \  ./saved-model/layers-model/model.json \  ./saved-model/graph-model
```

# 3. Convert to ONNX Model

This will convert to a ONNX model to be used with transformers.js on web or nodejs.
```
python3 -m tf2onnx.convert --tfjs ./saved-model/graph-model/model.json --output ./saved-model/model.onnx
```

Unable to figure a way to use Optimum with tensorflow.js models atm..

# 4. Validate ONNX

Make sure the conversion worked and no issues
```
python3 validate_onnx.py
```

# 5. Test ONNX Model python

update the image path in the code to point to an image to confirm working as expected 
- I tested against one of the trained image that should give 1.
```
python3 test_image.py
```
Inference outputs: [array([[0., 1.]], dtype=float32)]

# 5. Test ONNX Model JS onnxruntime-node

update the image path in the code to point to an image to confirm working as expected 
```
node onnxruntime-node
```
Inference outputs: Tensor {
  cpuData: Float32Array(2) [ 0, 1 ],
  dataLocation: 'cpu',
  type: 'float32',
  dims: [ 1, 2 ],
  size: 2
}