Method fit

ann_class2/theano_ann.py:42–116

(self, X, Y, activation=T.nnet.relu, learning_rate=1e-3, mu=0.0, reg=0, epochs=100, batch_sz=None, print_period=100, show_fig=True)

        self.hidden_layer_sizes = hidden_layer_sizes

    def fit(self, X, Y, activation=T.nnet.relu, learning_rate=1e-3, mu=0.0, reg=0, epochs=100, batch_sz=None, 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 = HiddenLayer(M1, M2, activation)
            self.layers.append(h)
            M1 = M2
        
        # final layer
        K = len(set(Y))
        # print("K:", K)
        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

        # for momentum
        dparams = [theano.shared(np.zeros_like(p.get_value())) for p in self.params]

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

        rcost = reg*T.mean([(p*p).sum() for p in self.params])
        cost = -T.mean(T.log(p_y_given_x[T.arange(thY.shape[0]), thY])) #+ rcost
        prediction = T.argmax(p_y_given_x, axis=1)
        grads = T.grad(cost, self.params)

        # momentum only
        updates = [
            (p, p + mu*dp - learning_rate*g) for p, dp, g in zip(self.params, dparams, grads)
        ] + [
            (dp, mu*dp - learning_rate*g) for dp, g in zip(dparams, grads)
        ]

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

        self.predict_op = theano.function(
            inputs=[thX],
            outputs=prediction,
        )