DNGPU_model.py 文件源码

def createLoss(self, x_in_indices, y_in, length):
        """perform loss calculation for one bin """

        # create mask
        mask_1 = tf.cast(tf.equal(x_in_indices, 0), tf.float32)
        mask_2 = tf.cast(tf.equal(y_in, 0), tf.float32)
        mask = tf.stack([1.0-mask_1*mask_2]*self.num_units,axis=2)

        # the input layer
        x_in = tf.one_hot(x_in_indices, self.n_input, dtype=tf.float32)
        cur = self.conv_linear(x_in, 1, self.n_input, self.num_units, 0.0, "input")
        cur = self.hard_tanh(cur, length)
        cur = self.dropout(cur)
        cur*=mask

        allMem = [cur] #execution trace

        #computation steps
        with vs.variable_scope("steps") as gruScope:
            for i in range(length):
                cur = self.DCGRU(cur, 3, "dcgru")
                cur *= mask
                allMem.append(cur)
                gruScope.reuse_variables()

        # output layer and loss
        allMem_tensor = tf.stack(allMem)
        prediction = self.conv_linear(cur, 1, self.num_units, self.n_classes, 0.0, "output")
        costVector = tf.nn.sparse_softmax_cross_entropy_with_logits(logits = prediction, labels = y_in)  # Softmax loss

        result = tf.argmax(prediction, 2)
        correct_pred = tf.equal(result, y_in)
        perItemCost = tf.reduce_mean(costVector, (1))
        cost = tf.reduce_mean(perItemCost)

        correct_pred = tf.cast(correct_pred, tf.float32)
        accuracy = tf.reduce_mean(correct_pred)

        return cost, accuracy, allMem_tensor, prediction, perItemCost, result