当张量不提供 numpy() 时如何获取数据

Question

I've got a tensor which is provided as data parameter by keras to my custom model's train_step . It seems to be handled just fine by the self(data, training=True) call. For evaluation of result I want to look at the data though.

When I print it out, I get:

Tensor("IteratorGetNext:0", shape=(120, 1080), dtype=float32)

But I can't seem to get the data itself. data.numpy() raises exception AttributeError: 'Tensor' object has no attribute 'numpy' .

I don't have a session available at that point for eval() either. How can I get the data out?

Answer 1

You cannot access tensor values as NumPy arrays in graph mode. This data is "leaving" Python and therefore inaccessible ( read more ). In order to access the tensors, you can run eagerly ( read more ) like this:

model.compile(... run_eagerly=True)