Nones returned in tape.gradient in Conv-TasNet Implementation

beepbopboop · February 15, 2022, 3:03am

Hi all;

I am attempting to implement Conv-TasNet for a course project. My train_step code is at:

JustinGoheen/tf2-ConvTasNet/blob/87ce38922aa1e80aa7b0714cb52ace3a75a5a9eb/research/algorithm.py#L340

      
        
                    raise KeyError(
                        f"Expected one of (weights, variables, x, pred). Got: {self.apply_loss_to}"
                    )
                # ENDNOTE
            
            
    # Trackers
                self.mean_loss = tf.keras.metrics.Mean(name="mean-loss")
                self.cossim_loss = tf.keras.metrics.CosineSimilarity(name="cos-sim")
                self.rmse_loss = tf.keras.metrics.RootMeanSquaredError(name="rmse-loss")
            
            
def train_step(self, batch):
                """custom training
            
            
    # Note
            
            
        see https://keras.io/examples/audio/melgan_spectrogram_inversion/
                    for a guide on writing a custom train_step.
                """
                x_train, y_true = batch
            
            
    with tf.GradientTape(persistent=True) as tape:

I’ve read the docs concerning possible causes of having all Nones returned by tape.gradient; I’ve made sure to remove any NumPy code from my custom loss function, and am not quite sure what else may be causing the issue.

I don’t think the error is caused by the loss function; I ran a debugging session with a built-in loss function, and the same result (all Nones) occurs.

The trainable_variables are tracked by tape, meaning the loss should be applied to a list of Variables as intended; so this should work, at least according to my knowledge of TensorFlow and Keras – which only spans about 6 weeks at the time of this posting.

I used the MelGAN example as a guide for how to implement a similar model; a main difference I can see in their code and mine is, I am not using (what appears to be) functional API concepts in my ConvBlock. Could that be the issue?

Regarding system/env:

Mac M1 Max
Python 3.9.7
TensorFlow/Keras 2.7

Conda environment is managed with miniforge.