tf.keras 两个损失，中间层作为其中一个的输入错误：急切执行的输入 function 不能是 Keras 符号张量

Question

I want to have two losses in my tensorflow keras model and one of them takes an intermediate layer as input. This code works when I use keras but when it comes to tensorflow.keras I face the following error.

def loss_VAE(input_shape, z_mean, z_var, weight_L2=0.1, weight_KL=0.1):

  def loss_VAE_(y_true, y_pred):
      c, H, W, D = input_shape
      n = c * H * W * D

      loss_L2 = K.mean(K.square(y_true - y_pred), axis=(1, 2, 3, 4)) # original axis value is (1,2,3,4).

      loss_KL = (1 / n) * K.sum(
          K.exp(z_var) + K.square(z_mean) - 1. - z_var,
          axis=-1
      )

      return weight_L2 * loss_L2 + weight_KL * loss_KL

  return loss_VAE_

def loss_gt(e=1e-8):

  def loss_gt_(y_true, y_pred):
      intersection = K.sum(K.abs(y_true * y_pred), axis=[-3,-2,-1])
      dn = K.sum(K.square(y_true) + K.square(y_pred), axis=[-3,-2,-1]) + e

      return - K.mean(2 * intersection / dn, axis=[0,1])

  return loss_gt_


model.compile(
    adam(lr=1e-4),
    [loss_gt(dice_e), loss_VAE(input_shape, z_mean, z_var, weight_L2=weight_L2, weight_KL=weight_KL)],
    # metrics=[dice_coefficient]
)

Error:

_SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'Dec_VAE_VDraw_Var/Identity:0' shape=(None, 128) dtype=float32>, <tf.Tensor 'Dec_VAE_VDraw_Mean/Identity:0' shape=(None, 128) dtype=float32>]

Is it bug? please find the complete code in this NOTEBOOK .

and THIS is the link to the data.

Answer 1

If running under the eager mode,tensorflow op will check if the inputs are of type "tensorflow.python.framework.ops.EagerTensor" and Keras ops are implemented as DAGs.So the inputs to the eager mode will be of tensorflow.python.framework.ops.Tensor and this throws the error

You can change the input type to EagerTensor by explicitly telling tensorflow to run in the eager mode for Keras.

tf.config.experimental_run_functions_eagerly(True)

Adding this statement should solve your issue. Although note that there will be significant performance hits since you are running now in eager mode and recommended only for debugging, profiling etc.

Answer 2

Replacing K.mean to tf.reduce_mean and accordingly all the keras backend functions to tensorflow functions solved the problem.