I can't use ConvLayers in deepqlearn.js

[soswow]

There is a problem with this line: https://github.com/karpathy/convnetjs/blob/4c3358a315b4d71f31a0d532eb5d1700e9e592ee/build/deepqlearn.js#L142

the shape of the input vector is basically hardcoded here when given to the network's forward function. Which makes it all break apart in line https://github.com/karpathy/convnetjs/blob/4c3358a315b4d71f31a0d532eb5d1700e9e592ee/src/convnet_layers_dotproducts.js#L67 Where we try to check for existence of those dimensions.

[soswow]

My current workaround is to do this:

deepqlearn.Brain.prototype.policy = function(s) {
  const svol_sx =  this.value_net.layers[0].out_sx;
  const svol_sy =  this.value_net.layers[0].out_sy;
  const svol_depth =  this.value_net.layers[0].out_depth;
  var svol = new convnetjs.Vol(svol_sx, svol_sy, svol_depth);
  svol.w = s;
  var action_values = this.value_net.forward(svol);
  var maxk = 0;
  var maxval = action_values.w[0];
  for(var k=1;k<this.num_actions;k++) {
    if(action_values.w[k] > maxval) { maxk = k; maxval = action_values.w[k]; }
  }
  return {action:maxk, value:maxval};
};```