2348493e231cefa9230ea785562005cea43508cb,basenji/rnn.py,RNN,build,#RNN#Any#,15

Before Change

                    // convolution params
                    stdev = 1./np.sqrt(self.cnn_filters[li]*seq_depth)
                    kernel = tf.Variable(tf.random_uniform([1, self.cnn_filter_sizes[li], seq_depth, self.cnn_filters[li]], minval=-stdev, maxval=stdev), name="kernel")
                    biases = tf.Variable(tf.zeros([self.cnn_filters[li]]), name="bias")

                    // convolution
                    conv = tf.nn.conv2d(cinput, kernel, [1, 1, 1, 1], padding="SAME")

After Change

                    conv = tf.nn.conv2d(cinput, kernel, [1, 1, 1, 1], padding="SAME")

                    // batch normalization
                    conv = tf.batch_norm(conv, center=True, scale=True, activation_fn=tf.nn.relu, is_training=self.is_training, update_collections=None, name="bnconv%d"%li)

                    // nonlinearity
                    // cinput = tf.nn.relu(tf.nn.bias_add(conv, biases), name="conv%d"%li)

Instances