训练CNN：ValueError：没有为任何变量提供渐变

问题描述

我想用以下几层训练模型：

embedding_dim = 80

model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(64,3,padding='same',activation='relu',input_shape=(50,120,3)))
model.add(tf.keras.layers.MaxPool2D(padding='same'))

model.add(tf.keras.layers.Conv2D(128,activation='relu'))
model.add(tf.keras.layers.MaxPool2D(padding='same'))

model.add(tf.keras.layers.Conv2D(256,activation='relu'))

model.add(tf.keras.layers.Conv2D(256,activation='relu'))

model.add(tf.keras.layers.Conv2D(512,2,strides=(2,4),activation='relu'))

model.add(tf.keras.layers.Lambda(add_timing_signal_nd))

model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(embedding_dim))

之后，我跑

model.compile(optimizer='adam',loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True,reduction='none'),metrics=['accuracy'])

model.fit(image_dataset,epochs=10,validation_data=val_dataset)

我收到以下错误

ValueError: No gradients provided for any variable: ['conv2d/kernel:0','conv2d/bias:0','conv2d_1/kernel:0','conv2d_1/bias:0','conv2d_2/kernel:0','conv2d_2/bias:0','conv2d_3/kernel:0','conv2d_3/bias:0','conv2d_4/kernel:0','conv2d_4/bias:0','conv2d_5/kernel:0','conv2d_5/bias:0','conv2d_6/kernel:0','conv2d_6/bias:0','dense/kernel:0','dense/bias:0'].

为进一步说明，add_timing_signa_nd定义如下

def add_timing_signal_nd(x,min_timescale=1.0,max_timescale=1.0e4):
    """
    Args:
        x: a Tensor with shape [batch,d1 ... dn,channels]
        min_timescale: a float
        max_timescale: a float
    Returns:
        a Tensor the same shape as x.
    """
    static_shape = x.get_shape().as_list()
    num_dims = len(static_shape) - 2
    channels = tf.shape(x)[-1]
    num_timescales = channels // (num_dims * 2)
    log_timescale_increment = (
            math.log(float(max_timescale) / float(min_timescale)) /
            (tf.cast(num_timescales,dtype=tf.float32) - 1))
    inv_timescales = min_timescale * tf.exp(
            tf.cast(tf.range(num_timescales),dtype=tf.float32) * -log_timescale_increment)
    for dim in xrange(num_dims):
        length = tf.shape(x)[dim + 1]
        position = tf.cast(tf.range(length),dtype=tf.float32)
        scaled_time = tf.expand_dims(position,1) * tf.expand_dims(
                inv_timescales,0)
        signal = tf.concat([tf.sin(scaled_time),tf.cos(scaled_time)],axis=1)
        prepad = dim * 2 * num_timescales
        postpad = channels - (dim + 1) * 2 * num_timescales
        signal = tf.pad(signal,[[0,0],[prepad,postpad]])
        for _ in xrange(1 + dim):
            signal = tf.expand_dims(signal,0)
        for _ in xrange(num_dims - 1 - dim):
            signal = tf.expand_dims(signal,-2)
        x += signal
    return x

如果对输入大小有帮助，请参见以下

(3,50,3)
(3,3)
(1,3)

此外，我正计划在训练将权重用于另一个问题后提取权重。

谢谢！

解决方法

我猜想image_dataset或val_dataset不正确。按照您的代码，我模拟了一些数据（包括标签）进行训练，并且可以正常运行。

image_dataset = np.random.uniform(0,1,(3000,50,120,3))
image_dataset_y = np.random.uniform(0,embedding_dim,)).astype(np.int)
val_dataset = np.random.uniform(0,(300,3))
val_dataset_y = np.random.uniform(0,)).astype(np.int)
model.fit(image_dataset,image_dataset_y,batch_size=30,epochs=10,validation_data=(val_dataset,val_dataset_y))