获取 keras 自定义损失 function 内部的训练数据形状

Question

I have written the below custom loss function, where I need to create a factor by dividing the input shape with the output shape.

def distance_loss(x,y):
    x_shape = K.int_shape(x)[1]
    y_shape = K.int_shape(y)[1]
    print(x_shape,y_shape)
    factor = x_shape/y_shape
    loss = tf.sqrt(factor) * tf.norm(x-y)
    return tf.math.abs(loss)

This is the model architecture is:

model = Sequential()
model.add(Dense(32,input_dim=4))
model.add(Dense(64,activation='relu'))
model.add(Dense(128,activation='relu'))
model.add(Dense(64,activation='relu'))
model.add(Dense(2,activation='relu'))
opt = Adam(lr = 0.001)
model.compile(optimizer = opt, loss=distance_loss,metrics=['accuracy'])

When I ran the model.compile line. The custom loss prints

None 2

and throws an error

TypeError: unsupported operand type(s) for /: 'NoneType' and 'int'

I read that the input shape of the training data is only known during the training phase. Is there any way to bypass this issue?

Answer 1

Use K.shape instead:

def distance_loss(x,y):
    x_shape = K.shape(x)[1]
    y_shape = K.shape(y)[1]
    factor = K.cast(x_shape, x.dtype) / K.cast(y_shape, y.dtype)
    loss = tf.sqrt(factor) * tf.norm(x-y)
    return tf.math.abs(loss)