Method fit

rnn_class/srn_language_tf.py:106–159

(self, X, epochs=500, show_fig=False)

    def fit(self, X, epochs=500, show_fig=False):
        N = len(X)
        D = self.D
        M = self.M
        V = self.V

        # initial weights
        We = init_weight(V, D).astype(np.float32)
        Wx = init_weight(D, M).astype(np.float32)
        Wh = init_weight(M, M).astype(np.float32)
        bh = np.zeros(M).astype(np.float32)
        h0 = np.zeros(M).astype(np.float32)
        Wo = init_weight(M, V).astype(np.float32)
        bo = np.zeros(V).astype(np.float32)

        # build tensorflow functions
        self.build(We, Wx, Wh, bh, h0, Wo, bo)

        # sentence input:
        # [START, w1, w2, ..., wn]
        # sentence target:
        # [w1,    w2, w3, ..., END]

        costs = []
        n_total = sum((len(sentence)+1) for sentence in X)
        for i in range(epochs):
            X = shuffle(X)
            n_correct = 0
            cost = 0
            for j in range(N):
                # problem! many words --> END token are overrepresented
                # result: generated lines will be very short
                # we will try to fix in a later iteration
                # BAD! magic numbers 0 and 1...
                input_sequence = [0] + X[j]
                output_sequence = X[j] + [1]

                # we set 0 to start and 1 to end
                _, c, p = self.session.run(
                    (self.train_op, self.cost, self.predict_op),
                    feed_dict={self.tfX: input_sequence, self.tfY: output_sequence}
                )
                # print "p:", p
                cost += c
                # print "j:", j, "c:", c/len(X[j]+1)
                for pj, xj in zip(p, output_sequence):
                    if pj == xj:
                        n_correct += 1
            print("i:", i, "cost:", cost, "correct rate:", (float(n_correct)/n_total))
            costs.append(cost)

        if show_fig:
            plt.plot(costs)
            plt.show()

    def predict(self, prev_words):
        # don't use argmax, so that we can sample