JS Library Performance Comparison

we chose about seven popular neuron network JavaScript library to run training and inference task on mnist dataset.

Network Construction

Dataset: mnist dataset on http://yann.lecun.com/exdb/mnist/

Train Input: 3200 images * 28 * 28 pixels * 1 channels

Test Input: 1000 images * 28 * 28 pixels * 1 channels

Model: a simple DNN model (with a little modification) in Keras Official Examples

Structure:

• Activation function: relu
• Optimizer: sgd (or library’s default optimizer)
• Loss function: cross entropy
• No dropout layers

URL

example: http://localhost:8000/tfjs-dnn-mnist/tensorflowjs.html?backend=cpu&traintime=10000&infertime=1000&hiddenlayernum=1&hiddenlayersize=64&batchsize=64

1. backend
2. processtime(in ms)
3. hiddenlayernum
4. hiddenlayersize
5. batchsize(for training)

Message

tfjs dnn backend hiddenLayerNum hiddenLayerSize batchSize avgTrainTime LoadTime warmupTime avgInferTime