UnboundLocalError in run.py during training about neural-audio-fp HOT 3 CLOSED

@Rodrigo29Almeida

Try setting SAVE_IMG to False in the config file.
But I don't quite understand it, because the training should return the sim_mtx when calculating the loss at the first iteration.

from neural-audio-fp.

Hello, thanks for your help before. There was a problem with a missing "/" in my path to the database. I apologize for the error.

I have installed an anaconda environement following your steps.

As I attempted to run the code in my environment, I encountered an issue in the trainer.py file, specifically in the train_step function. The problematic line is:
feat = m_specaug(m_pre(X)) # (nA+nP, F, T, 1)
The error message I received is as follows:
OperatorNotAllowedInGraphError: using a tf.Tensor as a Python bool is not allowed: AutoGraph did convert this function. This might indicate you are trying to use an unsupported feature.

I am reaching out to seek your guidance on resolving this issue. Additionally, I want to inform you that I previously addressed a problem with len() in the same file. The error message was:
"len is not well defined for symbolic Tensors. (X:0) Please call x.shape rather than len(x) for shape information."
To resolve this, I modified the code from:
n_anchors = len(X[0]) to n_anchors = X[0].shape[0]
I would greatly appreciate your assistance in understanding and resolving the current issue. If you have any suggestions or insights, it would be immensely helpful.

Thank you for your time.

from neural-audio-fp.

@Rodrigo29Almeida
As for that first error (though I'm not entirely sure until I try it myself), it could come from https://github.com/mimbres/neural-audio-fp/blob/main/model/fp/specaug_chain/layers/ncutout_tarray.py
While dealing with binary masks in graph-mode, we should avoid using python operators like (==, >, < , & |). an example. We are utilizing the @tf.function decorator, which performs auto-graph conversion. Its behavior varies with different TensorFlow versions.

It's likely the simplest solution would be to downgrade if you're using a later version of TensorFlow than the 2.1-2.5 I recommended. Such issues are common during the TF upgrade process.

Also, the modifications you've made to the second error case look good. This also seems to be related to the TF version.

from neural-audio-fp.