Tacotron2: Model can not converge

[HiiamCong]

Hello, I have a question.

I'm using a dataset with > 17k sentences (about 30 hours audio), 90% for training and 10% for validating.
It's been training for 3 days (using batch_size 8) and reaching Epoch 56.
Plz see training info below
[Grad norm]

[Training loss]

[Validation loss]

I thought it looks good.
But when I tested it, the output audio was wrong and Attention looks awful.

And the loss seems can not decrease any more.
Do I have to train for more Epoch or there was something wrong with my dataset, or something else?
Plz help me, thank u guys so much.

[wizardk]

1. Attention with n_frame_per_step = 1 is hard to converge
2. Convergence of attention needs more time
3. Adding EOS will help and accelerate convergence of attention

[terryyizhong]

1. n_frame_per_step

does this repo support n_frame_per_step larger than 1 now?

[terryyizhong]

same problem, I suppose it's because the batch_size.

[HiiamCong]

@wizardk thank u for your advice, I'll wait for some more Epochs to see what happen.

[HiiamCong]

@terryyizhong yeah, I do think so, actually, batch_size 8 is pretty small. Unfortunately, My RTX 2080Ti can not run with batch_size more than 10, I'm not sure why. Maybe the input audio durations are too long. (all of my audios < 20s)
Btw, when I trained the model with about 4k sentences, the Attention converges perfectly.

[terryyizhong]

@HiiamCong
Did you mean your Attention can converges perfectly with 4k sentences though the batch_size is 8 ?

[HiiamCong]

@terryyizhong
Yes, and the output audio is not perfect but understandable. To get smoother output audio, I try to increase training data to 17k sentences and then getting this Attention problem.

[hadaev8]

Batch size very important, i was able to coverage on 16 only then used transfer learning from english model.

[HiiamCong]

@hadaev8 I wanna use tacotron2 for another language so I have to train the model from scratch. Do you have any idea to increase batch size?

[hadaev8]

English checkpoint for russian dataset works pretty well.
You may drop too long audios from dataset (ljspeech max length is 10 secs) or use google colab with 15gb t4 gpu.

[terryyizhong]

thx for the info. btw, I failed learn alignment using another language dataset at first. And success after change the decoder dropout from 0.1 to 0.5 (used by another taco2 repo), with batch size 48

[HiiamCong]

@terryyizhong
p_attention_dropout=0.1, p_decoder_dropout=0.1,
how about attention dropout ? Did you also increase it?

[wizardk]

1. n_frame_per_step

does this repo support n_frame_per_step larger than 1 now?

Check this: https://github.com/BogiHsu/Tacotron2-PyTorch

[wizardk]

@wizardk thank u for your advice, I'll wait for some more Epochs to see what happen.

Finally, you should get alignment like this:

[terryyizhong]

Check this: https://github.com/BogiHsu/Tacotron2-PyTorch
@wizardk
thx for the link, have you tried this repo? how is the performance compare to this repo?

[terryyizhong]

@terryyizhong
p_attention_dropout=0.1, p_decoder_dropout=0.1,
how about attention dropout ? Did you also increase it?

@HiiamCong No, I'm still doing experiment about the dropout. And I found the format of punctuation is very important to learn attention. After clean the spaces before some punctuations, the attention alignment starts looks better though still not converge after 100k.

[HiiamCong]

@terryyizhong
Have you tried to decrease the learning rate?

[terryyizhong]

@terryyizhong
Have you tried to decrease the learning rate?

@HiiamCong yeah, I am using exponential learning rate decay. But I am now wondering should I decrease the learning rate before the attention converge. Or should I "increase" the initial learning rate for the attention learning

[terryyizhong]

here are my attention plots and loss curve at step 130k. I am using private english data, 5 hours(totally 3000 sents, including 800 sentences from lj recorded in other voice) batch 32, other parameters are the same as default.

I think the plot is becoming diagnal. But the plot didn't change much in recent 30k. Any suggestion about learn the alignment? btw, I learn the alignment success using LJspeech dataset in 30k steps using the same params. @rafaelvalle

[HiiamCong]

@terryyizhong
Thank for your information.
Btw how did you implement exponential learning rate decay with this nvidia's tacotron code? I can not find these settings in hparams.

[terryyizhong]

@terryyizhong
Thank for your information.
Btw how did you implement exponential learning rate decay with this nvidia's tacotron code? I can not find these settings in hparams.

I just add code like:
learning_rate = init_lr * (0.01 ** (epoch / 1000.0))
in the main loop of train.py

[Clement-Hui]

batch size 8 is good. model converages perfectly using this batch size
I used a 3-4 hour dataset without any punctuation and pretrained english model and the model converages around 5k-10k steps, without problem. I set the dropout of attention to be 0.4 for both. no exponential learning rate decay. The model started overfitting around 10k steps. The speech is perfectly understandable. The dataset is a cantonese dataset, without any tone label.

[terryyizhong]

batch size 8 is good. model converages perfectly using this batch size
I used a 3-4 hour dataset without any punctuation and pretrained english model and the model converages around 5k-10k steps, without problem. I set the dropout of attention to be 0.4 for both. no exponential learning rate decay. The model started overfitting around 10k steps. The speech is perfectly understandable. The dataset is a cantonese dataset, without any tone label.

How can you train a Cantonese model use a pretrained english model?

[Clement-Hui]

Oh. I use jyutping which simply translates Cantonese unicode characters to alphabets and numbers. That means "你好" which means hello get converted to nei5 hou2. Also, I added Eos and symbols that indicate words in the string of input in a second training

[terryyizhong]

Oh. I use jyutping which simply translates Cantonese unicode characters to alphabets and numbers. That means "你好" which means hello get converted to nei5 hou2. Also, I added Eos and symbols that indicate words in the string of input in a second training

@Clement-Hui Thanks for your reply! I am surprise it works that way. Cause I thought the character embeddings of these two language are different.
Besides, change the dropout rate of attention to 0.4 helps a lot! I am now get a better attention alignment in only 13k, hope it will converge successfully!

[Clement-Hui]

Oh. I use jyutping which simply translates Cantonese unicode characters to alphabets and numbers. That means "你好" which means hello get converted to nei5 hou2. Also, I added Eos and symbols that indicate words in the string of input in a second training

@HiiamCong Thanks for your reply! I am surprise it works that way. Cause I thought the character embeddings of these two language are different.
Besides, change the dropout rate of attention to 0.4 helps a lot! I am now get a better attention alignment in only 13k, hope it will converge successfully!

The model uses the --warm start parameter and remove the embedding layer, so it can be retrained. The symbols.py is mostly the same, but only include a-z and 1-6 and spaces + punctuation. Also, I am training the model by considering the structure in 3 pieces and treating the Cantonese character in three parts, instead of alphabets. It took a longer time to coverage, as it is quite different from English, but it is converging.

3200 steps

5300 steps

14100 steps

21400 steps

42900 steps

[HiiamCong]

batch size 8 is good. model converages perfectly using this batch size
I used a 3-4 hour dataset without any punctuation and pretrained english model and the model converages around 5k-10k steps, without problem. I set the dropout of attention to be 0.4 for both. no exponential learning rate decay. The model started overfitting around 10k steps. The speech is perfectly understandable. The dataset is a cantonese dataset, without any tone label.

Thank for your useful information, btw what was your validation loss when the model converged?

[Clement-Hui]

It was around 0.4 and increased to 4.8 after 10k more steps. The output audio haven't changed in quality

[chazo1994]

1. Attention with n_frame_per_step = 1 is hard to converge
2. Convergence of attention needs more time
3. Adding EOS will help and accelerate convergence of attention

I'm sorrry but can you help me to explain what is EOS and how to add it into source code. Please!

[Clement-Hui]

EOS stands for End of sentence. You can use some symbol such as semicolon to represent EOS. Just append the symbol at the end of each line.

[rafaelvalle]

Closing due to inactivity.