rm np_config.enable_numpy_behavior()

[calad0i]

Description

Fixes #1092, while trying to keep the original printing behavior.

[bo3z]

I just ran this code - removing the NumPy behaviour doesn't break prints, it breaks the line above avg_loss = round(epoch_loss_avg.result(), 3) in the file dsp_aware_pruning/keras/__init__.py (line 234). So this has to do with querying EagerTensors (which are not always evaluated until necessary). The error I get is:

AttributeError: EagerTensor object has no attribute 'astype'. 
        If you are looking for numpy-related methods, please run the following:
        from tensorflow.python.ops.numpy_ops import np_config
        np_config.enable_numpy_behavior()

So if we remove this import we need to find a way to print the loss during training.

[calad0i]

I just ran this code - removing the NumPy behaviour doesn't break prints, it breaks the line above avg_loss = round(epoch_loss_avg.result(), 3) in the file dsp_aware_pruning/keras/__init__.py (line 234). So this has to do with querying EagerTensors (which are not always evaluated until necessary). The error I get is:

AttributeError: EagerTensor object has no attribute 'astype'. 
        If you are looking for numpy-related methods, please run the following:
        from tensorflow.python.ops.numpy_ops import np_config
        np_config.enable_numpy_behavior()

So if we remove this import we need to find a way to print the loss during training.

Use tf.cast(x) instead of x.astype would solve that.

[bo3z]

I just ran this code - removing the NumPy behaviour doesn't break prints, it breaks the line above avg_loss = round(epoch_loss_avg.result(), 3) in the file dsp_aware_pruning/keras/__init__.py (line 234). So this has to do with querying EagerTensors (which are not always evaluated until necessary). The error I get is:

AttributeError: EagerTensor object has no attribute 'astype'. 
        If you are looking for numpy-related methods, please run the following:
        from tensorflow.python.ops.numpy_ops import np_config
        np_config.enable_numpy_behavior()

So if we remove this import we need to find a way to print the loss during training.

Use tf.cast(x) instead of x.astype would solve that.

That's a likely solution, but in the code we don't actually call x.astype. This is called somewhere internally in TF when trying to get the value of the loss tensor (epoch_loss_avg.result()), so we need to see what the alternative is to get the loss value to this function. Also do we now if this buggy behaviour is introduced by an updated in NumPy / TF / QKeras?

[bo3z]

Turns out the changes in this PR have already been merged into main and DSP-aware pruning is actually broken in the main branch. My guess is that one of @calad0i previous PR accidentally included the change to remove enable_numpy_behaviour.

The proposed solution with tf_prints is on the right track, but one also needs to remove the round. I also found a minor inconsistency with the function docstring, and to avoid opening a PR to a PR, the two fixes are now in PR #1396.

I suggest we close this PR, as the core change (removing enable_numpy_behaviour) is already in main and merge #1396.

[JanFSchulte]

Closed in favor of #1396