error loading checkpoint in synthesis time about tacotron-2 HOT 11 CLOSED

I found similar error.

In my case,
model_1/model/ResidualConv1dGLU_0/residual_block_cin_conv/bias_residual_block_cin_conv/ExponentialMovingAverage was saved as model_1/model_1/model/ResidualConv1dGLU_0/residual_block_cin_conv/bias_residual_block_cin_conv/ExponentialMovingAverage.

You can edit your create_shadow_saver to load variables with those names.

You can inspect your .ckpt file using

from tensorflow.python.tools import inspect_checkpoint as chkp

chkp.print_tensors_in_checkpoint_file("logs/wavenet_r9y9/model/model.ckpt-50000", tensor_name='', all_tensors=True, all_tensor_names=True)

from tacotron-2.

@a3626a Thank you very much.
But I ran again to see whats happened, and I found this error during training:
"tensorflow.python.framework.errors_impl.CancelledError: Enqueue operation was cancelled"
exactly after: "Tacotron training complete after 85000 global steps!"

from tacotron-2.

Check your command line argument tacotron_train_steps

from tacotron-2.

Yeah, I got it. Thank you very much. I really appreciate.

from tacotron-2.

Hi again.
@a3626a Thank you, I used your help (editing names of variables from model_1 to model) and it loaded those models without problem, but it errors for models in encoder and decoder parts of wavenet. I've checked variables during training and non of the variables are found in training variables during training. How can I handle this?
Here is an example of the error:

model/model/inference/decoder/Location_Sensitive_Attention/attention_bias/ExponentialMovingAverage

from tacotron-2.

Inside your checkpoint, trainable variables (tf.trainable_variabels()) are not saved, their Exponential Moving Averages are saved.

Can you explain what you are trying to do?

1. Load trained model and synthesis
2. Load trained model and continue training
3. Just to start training

from tacotron-2.

1 and 2 both have been tried. Actually, I want 1.

from tacotron-2.

Hello! I'm getting the same errors as described above.

When loading the trained model and running synthesis using --model='Both' it's looking for "model_1/model/..." rather than "model/model/..."

I updated create_shadow_saver to replace "model_1" with "model" but then run into the following error:
NotFoundError (see above for traceback): Key model/model/inference/decoder/Location_Sensitive_Attention/attention_bias/ExponentialMovingAverage not found in checkpoint

I usde print_tensors_in_checkpoint_file and as far as I can tell this name is not stored in the checkpoint file at all.

Anyone know what is going on here? Thanks!

from tacotron-2.

create_shadow_saver selects some variables for saving. You can comment this part and use tf.saver.save to save whole tensor variables. Then you have to change synthesis part for loading saved tensors.

from tacotron-2.

Check it.
https://github.com/Rayhane-mamah/Tacotron-2/pulls

from tacotron-2.

Hello, I can't quite manage to reproduce with latest commit, I am closing this as I believe it was fixed meanwhile. If the problem persists feel free to reopen!

I also disabled training with "Both" as it seems to be causing problems for many people, so now simply use "Tacotron-2" for both models.

from tacotron-2.