This example showcases the LeNet-based handwritten digits classification by WebNN API.
This example leverages the network topology of the LeNet example of Caffe*, the weights of the LeNet example of OpenVINO* and the MNIST dataset of mnist.js.
The following diagram illustrates the topology of lenet.prototxt.
The following table lists the corresponding WebNN ops and parameters for each layer.
| Layer (name:type) | WebNN op | Parameters |
|---|---|---|
| conv1: Convolution | nn.conv2d | input shape [1, 1, 28, 28], filter shape [20, 1, 5, 5], layout "nchw" |
| nn.add | bias shape [1, 20, 1, 1] | |
| pool1: Pooling | nn.maxPool2d | window shape [2, 2], strides [2, 2] |
| conv2: Convolution | nn.conv2d | filter shape [50, 20, 5, 5], layout "nchw" |
| nn.add | bias shape [1, 50, 1, 1] | |
| pool2: Pooling | nn.maxPool2d | window shape [2, 2], strides [2, 2] |
| ip1: InnerProduct | nn.matmul | transposed weights shape [800, 500] |
| nn.add | bias shape [1, 500] | |
| relu1: ReLU | nn.relu | |
| ip2: InnerProduct | nn.matmul | transposed weights shape [500, 10] |
| nn.add | bias shape [1, 10] | |
| prob: Softmax | nn.softmax | output shape [1, 10] |
Click the Device switch button to choose device preference for inference.
Click the Predict button to predict the digit shown in the canvas.
Click the Next button to pick up another digit from MNIST dataset.
Click the Clear button to clear the canvas and use mouse to draw a digit manually.
Predict the MNIST digit
Predict the mannually drawing digit
