How many epochs should i train? about avsr-tf1 HOT 2 CLOSED

hi @dandelioner
Thank you for attempting to reproduce our experiments.
The default parameters in experiment_tcd_audio.py are not the ones that we used in the ICMI 2018 work. I will list them below for better reproducibility.

encoder_units_per_layer=((), (256, 256, 256)),
cell_type='lstm',
decoder_units_per_layer=(256, ),
optimiser='Adam', # because we removed the support for AMSGrad due to a bug

On the Speaker Dependent (SD) partition of TCD-TIMIT I normally train the models for 400 epochs at a learning rate of 0.001 and an additional 100 epochs at 0.0001. Decaying the learning rate typically results in a substantial increase of accuracy on the test set (5-8%).

In case there is still a difference between your values and ours, I can think of two other reasons.
-The experiments for ICMI 2018 were run in last year in March, when we used an older version of the OpenFace toolkit to detect and align the faces. With OpenFace 2.0 we now get a 99% detection rate even on the more challenging LRS2 dataset.
-Locally I was able to fix several label errors, there were many problems with the videos of volunteer 42M in particular. Unfortunately I do not have write access to the public repository of TCD-TIMIT to update the files there as well. However I don't expect to see a difference larger than 1% because of this aspect.

Please let me know if this fully answers your question.

from avsr-tf1.

Hi @georgesterpu
Thank you for your reply, it's very helpful for me.
After I change the parameters, I get the result: 20.54%(cer)/44.04%(wer). It's very close to the result in paper.

from avsr-tf1.