Method fit

ann_class2/batch_norm_theano.py:115–194

(self, X, Y, Xtest, Ytest, activation=T.nnet.relu, learning_rate=1e-2, mu=0.9, epochs=15, batch_sz=100, print_period=100, show_fig=True)

    self.hidden_layer_sizes = hidden_layer_sizes

  def fit(self, X, Y, Xtest, Ytest, activation=T.nnet.relu, learning_rate=1e-2, mu=0.9, epochs=15, batch_sz=100, print_period=100, show_fig=True):
    X = X.astype(np.float32)
    Y = Y.astype(np.int32)

    # initialize hidden layers
    N, D = X.shape
    self.layers = []
    M1 = D
    for M2 in self.hidden_layer_sizes:
      h = HiddenLayerBatchNorm(M1, M2, activation)
      self.layers.append(h)
      M1 = M2
      
    # final layer
    K = len(set(Y))
    h = HiddenLayer(M1, K, T.nnet.softmax)
    self.layers.append(h)

    if batch_sz is None:
      batch_sz = N

    # collect params for later use
    self.params = []
    for h in self.layers:
      self.params += h.params

    # note! we will need to build the output differently
    # for train and test (prediction)

    # set up theano functions and variables
    thX = T.matrix('X')
    thY = T.ivector('Y')

    # for training
    p_y_given_x = self.forward(thX, is_training=True)

    cost = -T.mean(T.log(p_y_given_x[T.arange(thY.shape[0]), thY]))
    prediction = T.argmax(p_y_given_x, axis=1)
    grads = T.grad(cost, self.params)

    # momentum only
    updates = momentum_updates(cost, self.params, learning_rate, mu)
    for layer in self.layers[:-1]:
      updates += layer.running_update

    train_op = theano.function(
      inputs=[thX, thY],
      outputs=[cost, prediction],
      updates=updates,
    )

    # for testing
    test_p_y_given_x = self.forward(thX, is_training=False)
    test_prediction = T.argmax(test_p_y_given_x, axis=1)

    self.predict = theano.function(
      inputs=[thX],
      outputs=test_prediction,