Need help with my image classification net

[Guoliang Cao]

Hi,

I'm trying to use convnetjs to classify 320x240 grayscale image as Rock/Paper/Scissors/Unknown. Here is how I constructed the model

    var layer_defs = [];

    // input layer (all volumes are 3D)

    layer_defs.push({type:'input', out_sx:320, out_sy:240, out_depth:1});

    // output of below layer will be 160x120x8

    layer_defs.push({type:'conv', sx:5, filters:8, stride:2, pad:2, activation:'relu'});

    // output of below layer will be 80x60x8

    layer_defs.push({type:'pool', sx:2, stride:2});

    // output of below layer will be 40x30x8

    layer_defs.push({type:'conv', sx:5, filters:8, stride:2, pad:2, activation:'sigmoid'});

    // output of below layer will be 20x15x8

    layer_defs.push({type:'pool', sx:2, stride:2});

    // a softmax classifier predicting probabilities for four classes: 0,1,2,3

    layer_defs.push({type:'softmax', num_classes:4});

    this.model = new convnetjs.Net();

    this.model.makeLayers(layer_defs);

    this.trainer = new convnetjs.Trainer(this.model, {

      l2_decay:0.001,

      learning_rate: 0.02,

    });

After I train the model for ~100 examples, when I use the model to predict, no matter what image I give, the prediction weights are the same. Anyone know what is wrong with my model?

Thank you for your help!

Guoliang Cao

[vladn...@gmail.com]

You need thousands of examples if you don't want to overfit. 100 examples is way to little for a proper training. Also, how many iterations of training are you running?

[jjrod...@anda.pe]

Your model has too much overfitting use more data and apply normalization after convolution. Create synthetic data if you don't have enough data