Getting value of a Tensor inside train_step when eager execution is enabled

I’m following the tutorial here to create a custom train_step. I’m trying to save some intermediate results (which are Tensor objects) inside a custom Model class with eager execution enabled.

When I try something like tf.keras.backend.get_value(some_tensor), some_tensor.numpy() inside the train_step, I get an error that Tensor object has no attribute numpy. I’ve also tried some_tensor.eval(session = tf.compat.v1.Session()) and also got an error. I’m wondering if there’s any way to extract the value of tensors computed inside the train_step.

1 Like

Hi @Lu_Bin_Liu, You can define a list in the custom train step and can append the values which you want and after training the model you can access those values. please refer to this gist for code example. Thank You.