Question

我在尝试提供占位符时遇到问题，在train_epoch函数中使用feed_dict = {..}，它无法识别占位符

这是代码..

 class CNN(object):
    ###......
    def define_train_opeartions(self):
        X_data_train = tf.placeholder(dtype=tf.float32, shape=(None, self.height,self.width,self.chan),name='X_data_train')

        Y_data_train = tf.placeholder(dtype=tf.int32, shape=(None, self.n_classes),name='Y_data_train')  # Define this

        # Network prediction
        Y_net_train = self.inference(
            X_data_train,reuse=False)

        # Loss of train data tf.nn.softmax_cross_entropy_with_logits
        self.train_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y_data_train, logits=Y_net_train, name='train_loss'))

        # define learning rate decay method
        global_step = tf.Variable(0, trainable=False, name='global_step')
        # Define it--play with this
        learning_rate = 0.001

        # define the optimization algorithm
        # Define it --shall we try different type of optimizers
        optimizer = tf.train.AdamOptimizer(learning_rate)

        trainable = tf.trainable_variables()  # may be the weights??
        self.update_ops = optimizer.minimize(
            self.train_loss, var_list=trainable, global_step=global_step)

        # --- Validation computations
        X_data_valid = tf.placeholder(dtype=tf.float32, shape=(None, self.height, self.width, self.chan))  # Define this
        Y_data_valid = tf.placeholder(dtype=tf.int32, shape=(None, self.n_classes))  # Define this

        # Network prediction
        Y_net_valid = self.inference(X_data_valid,reuse=True)

        # Loss of validation data
        self.valid_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(
            labels=Y_data_valid, logits=Y_net_valid, name='valid_loss'))

然后我又有了另一个功能

def train_epoch(self, sess):
        train_loss = 0
        total_batches = 0
        keep_probability=0.2     #dropout probability
        n_batches = self.train_size / self.batch_size  # ??
        indx=0
        while (total_batches < n_batches):     # loop through train batches:
            X,Y=self.shuffling(self.Xtrain_in,self.Ytrain_in)  # shuffle X ,Y data
            Xbatch,Ybatch,indx=self.read_nxt_batch(X,Y,self.batch_size,indx)    # take the right batch
            mean_loss, _ = sess.run([self.train_loss, self.update_ops], feed_dict={X_data_train: Xbatch ,Y_data_train: Ybatch })
            if math.isnan(mean_loss):
                print('train cost is NaN')
                break
            train_loss += mean_loss
            total_batches += 1

        if total_batches > 0:
            train_loss /= total_batches

        return train_loss

错误消息：TypeError：无法将feed_dict键解释为Tensor：名称＆gt;'X_data_train'指的是Operation，而不是Tensor。张量名称的格式必须为op_name：output_index。

Answer 1

占位符张量名称与您为其指定的操作名称相同。那是造成错误的原因。给Op一个不同的名字：

X_data_train = tf.placeholder(dtype=tf.float32, shape=(None, self.height, self.width, self.chan), name='x_train_ph')

与Y_data_train相同。

未识别出张量流错误张量中的占位符

1 个答案: