您必须为占位符张量“Placeholder”提供一个值，其数据类型为浮点型和形状

新的错误案例

我改变了我的代码。
x = tf.placeholder(tf.float32, [None, 222], name='inputX')

错误案例 1
tensorFreq = tf.convert_to_tensor(basicFreq, tf.float32)
newY = tf.gather(tensorFreq, maxPredict) * 60
loss = tf.abs(y - tf.Variable(newY))

ValueError：initial_value 必须指定一个形状：Tensor("mul:0", shape=(?,), dtype=float32)

错误案例 2
tensorFreq = tf.convert_to_tensor(basicFreq, tf.float32)
newY = tf.gather(tensorFreq, maxPredict) * 60
loss = tf.abs(y - newY)

回溯（最近一次调用）：文件“D:/PycharmProject/DetectionSignal/TEST_FCL_StackOverflow.py”，第 127 行，trainStep = opt.minimize(loss) 文件“C:\Users\Heewony\Anaconda3\envs\TSFW_pycharm\lib \site-packages\tensorflow\python\training\optimizer.py”，第 407 行，最小化（[str(v) for _，v in grads_and_vars]，loss））ValueError：没有为任何变量提供梯度，检查你的图表对于不支持梯度的操作，变量之间 [tf.Variable 'Variable:0' shape=(222, 1024) dtype=float32_ref, tf.Variable 'Variable_1:0' shape=(1024,) dtype=float32_re, .. ....... tf.Variable 'Variable_5:0' shape=(222,) dtype=float32_ref] 和损失 Tensor("Abs:0", dtype=float32)。

应该有两件事要改变。错误案例 0。您不需要重塑层之间的流程。您可以None在第一个维度上使用来传递动态批量大小。错误案例 1.您可以直接使用 newY 作为 NN 的输出。您只使用 tf.Variable 来定义权重或偏差。错误案例 2.似乎 tensorflow 没有梯度下降实现，也tf.abs()没有tf.gather()。对于回归问题，均方误差通常就足够了。在这里，我如何重写您的代码。我没有你的 matlab 部分，所以我无法调试你的 python/matlab 接口：模型：def Model_FCL(inputX):    # Fully Connected Layer 1    fcW1 = tf.get_variable('w1', shape=[222, 1024], initializer=tf.initializer.truncated_normal())    fcb1 = tf.get_variable('b1', shape=[222], initializer=tf.initializer.truncated_normal())    # fcb1 = tf.get_variable('b1', shape=[None, 222], trainable=False, initializer=tf.constant_initializer(valueThatYouWant)) # if you want to fix your bias constant    fch1 = tf.nn.relu(tf.matmul(inputX, fcW1) + fcb1, name='relu1')    # Fully Connected Layer 2    fcW2 = tf.get_variable('w2', shape=[1024, 1024], initializer=tf.initializer.truncated_normal())    fcb2 = tf.get_variable('b2', shape=[222], initializer=tf.initializer.truncated_normal())    # fcb2 = tf.get_variable('b2', shape=[None, 222], trainable=False, initializer=tf.constant_initializer(valueThatYouWant)) # if you want to fix your bias constant    fch2 = tf.nn.relu(tf.matmul(fch1, fcW2) + fcb2, name='relu2')    # Output Layer    fcW3 = tf.get_variable('w3', shape=[1024, 222], initializer=tf.initializer.truncated_normal())    fcb3 = tf.get_variable('b3', shape=[222], initializer=tf.initializer.truncated_normal())    # fcb2 = tf.get_variable('b2', shape=[None, 222], trainable=False, initializer=tf.constant_initializer(valueThatYouWant)) # if you want to fix your bias constant    logits = tf.add(tf.matmul(fch2, fcW3), fcb3)    predictY = tf.nn.softmax(logits)  #I'm not sure that it will learn if you do softmax then abs/MSE    return predictY, logits图形：with myGraph.as_default():    # define input data & output data 입력받기 위한 placeholder    # put None(dynamic batch size) not -1 at the first dimension so that you can change your batch size    x = tf.placeholder(tf.float32, shape=[None, 222], name='inputX')  # Signal size = [1, 222]    y = tf.placeholder(tf.float32, shape=[None], name='trueY')  # Float value size = [1]    ...    predictY, logits = Model_FCL(x)  # Predict Signal, size = [1, 222]    maxPredict = tf.argmax(predictY, 1, name='maxPredict')  # Find max index of Predict Signal    tensorFreq = tf.convert_to_tensor(basicFreq, tf.float32)    newY = tf.gather(tensorFreq, maxPredict) * 60   # Find the value that corresponds to the Freq array index    loss = tf.losses.mean_squared_error(labels=y, predictions=newY)  # maybe use MSE for regression problem    # loss = tf.abs(y - newY)  # Calculate absolute (true Y - predict Y) #tensorflow doesn't have gradient descent implementation for tf.abs    opt = tf.train.AdamOptimizer(learning_rate=0.0001)    trainStep = opt.minimize(loss)

您必须为占位符张量“Placeholder”提供一个值，其数据类型为浮点型和形状

新的错误案例

开发环境

2回答