SLM Adversarial Training did not start when finetuning

[godspirit00]

I tried to do finetuning on a small dataset with 2 speakers. I set epochs=25, diff_epoch=8, joint_epoch=15. The Style Diffusion training started as expected, but SLM Adversarial Training never started throughout the entire finetuning process.

My config is

log_dir: "Models/xxx"
save_freq: 1
log_interval: 10
device: "cuda"
epochs: 25 # number of finetuning epoch (1 hour of data)
batch_size: 2
max_len: 400 # maximum number of frames
pretrained_model: "Models/LibriTTS/epochs_2nd_00020.pth"
second_stage_load_pretrained: true # set to true if the pre-trained model is for 2nd stage
load_only_params: true # set to true if do not want to load epoch numbers and optimizer parameters

F0_path: "Utils/JDC/bst.t7"
ASR_config: "Utils/ASR/config.yml"
ASR_path: "Utils/ASR/epoch_00080.pth"
PLBERT_dir: 'Utils/PLBERT/'

data_params:
  train_data: "Data/xxx__train.txt"
  val_data: "Data/xxx__val.txt"
  root_path: ""
  OOD_data: "Data/OOD_texts.txt"
  min_length: 50 # sample until texts with this size are obtained for OOD texts

preprocess_params:
  sr: 24000
  spect_params:
    n_fft: 2048
    win_length: 1200
    hop_length: 300

model_params:
  multispeaker: true

  dim_in: 64 
  hidden_dim: 512
  max_conv_dim: 512
  n_layer: 3
  n_mels: 80

  n_token: 178 # number of phoneme tokens
  max_dur: 50 # maximum duration of a single phoneme
  style_dim: 128 # style vector size
  
  dropout: 0.2

  # config for decoder
  decoder: 
      type: 'hifigan' # either hifigan or istftnet
      resblock_kernel_sizes: [3,7,11]
      upsample_rates :  [10,5,3,2]
      upsample_initial_channel: 512
      resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
      upsample_kernel_sizes: [20,10,6,4]
      
  # speech language model config
  slm:
      model: 'microsoft/wavlm-base-plus'
      sr: 16000 # sampling rate of SLM
      hidden: 768 # hidden size of SLM
      nlayers: 13 # number of layers of SLM
      initial_channel: 64 # initial channels of SLM discriminator head
  
  # style diffusion model config
  diffusion:
    embedding_mask_proba: 0.1
    # transformer config
    transformer:
      num_layers: 3
      num_heads: 8
      head_features: 64
      multiplier: 2

    # diffusion distribution config
    dist:
      sigma_data: 0.2 # placeholder for estimate_sigma_data set to false
      estimate_sigma_data: true # estimate sigma_data from the current batch if set to true
      mean: -3.0
      std: 1.0
  
loss_params:
    lambda_mel: 5. # mel reconstruction loss
    lambda_gen: 1. # generator loss
    lambda_slm: 1. # slm feature matching loss
    
    lambda_mono: 1. # monotonic alignment loss (TMA)
    lambda_s2s: 1. # sequence-to-sequence loss (TMA)

    lambda_F0: 1. # F0 reconstruction loss
    lambda_norm: 1. # norm reconstruction loss
    lambda_dur: 1. # duration loss
    lambda_ce: 20. # duration predictor probability output CE loss
    lambda_sty: 1. # style reconstruction loss
    lambda_diff: 1. # score matching loss
    
    diff_epoch: 8 # style diffusion starting epoch
    joint_epoch: 15 # joint training starting epoch

optimizer_params:
  lr: 0.0001 # general learning rate
  bert_lr: 0.00001 # learning rate for PLBERT
  ft_lr: 0.0001 # learning rate for acoustic modules
  
slmadv_params:
  min_len: 400 # minimum length of samples
  max_len: 500 # maximum length of samples
  batch_percentage: 0.5 # to prevent out of memory, only use half of the original batch size
  iter: 10 # update the discriminator every this iterations of generator update
  thresh: 5 # gradient norm above which the gradient is scaled
  scale: 0.01 # gradient scaling factor for predictors from SLM discriminators
  sig: 1.5 # sigma for differentiable duration modeling

What have I missed? Thanks!

[DogeLord081]

Same issue

[meng2468]

You seem to be missing the configuration options for when the second training starts

See lines 6 and 7 in the LibriTTS Config File

epochs_1st: 50 # number of epochs for first stage training (pre-training)
epochs_2nd: 30 # number of peochs for second stage training (joint training)

Should be able to kick-off second stage training by loading your current model checkpoint and setting epochs_1st to 0

[DogeLord081]

You seem to be missing the configuration options for when the second training starts

See lines 6 and 7 in the LibriTTS Config File

epochs_1st: 50 # number of epochs for first stage training (pre-training)
epochs_2nd: 30 # number of peochs for second stage training (joint training)

Should be able to kick-off second stage training by loading your current model checkpoint and setting epochs_1st to 0

Thanks, do I also need the first_stage_path:?

[DogeLord081]

You seem to be missing the configuration options for when the second training starts

See lines 6 and 7 in the LibriTTS Config File

epochs_1st: 50 # number of epochs for first stage training (pre-training)
epochs_2nd: 30 # number of peochs for second stage training (joint training)

Should be able to kick-off second stage training by loading your current model checkpoint and setting epochs_1st to 0

This did not fix the issue unfortunately

[78Alpha]

I've managed to get to:

        for bib in range(len(output_lengths)):
            mel_length_pred = output_lengths[bib]
            mel_length_gt = int(mel_input_length[bib].item() / 2)
            if mel_length_gt <= mel_len or mel_length_pred <= mel_len:
                continue
            sp.append(s_preds[bib])

            random_start = np.random.randint(0, mel_length_pred - mel_len)
            en.append(asr_pred[bib, :, random_start:random_start+mel_len])
            p_en.append(p_pred[bib, :, random_start:random_start+mel_len])

            # get ground truth clips
            random_start = np.random.randint(0, mel_length_gt - mel_len)
            y = waves[bib][(random_start * 2) * 300:((random_start+mel_len) * 2) * 300]
            wav.append(torch.from_numpy(y).to(ref_text.device))
            
            if len(wav) >= self.batch_percentage * len(waves): # prevent OOM due to longer lengths
                break
        if len(sp) <= 1:
            return None

Aa where things go sour. If you have a batch size of 2, then it will always be 1, meaning SLMADV never starts. You need to change batch_percentage: 0.5 to at least batch_percentage: 1. However, if you were running at the edge of memory before (Example, 20 GB/24 GB) then you will not be able to comfortably use this unless you have shared memory (28.4 GB/24 GB). It will also take 8 times longer unless you really crank down the min/max slmadv length. So instead of a max of 500, it would be a max of 220 or so.

[godspirit00]

@78Alpha I tried with batch size 2, batch_percentage: 1, max_len: 600; batch size 6, batch_percentage: 0.8, max_len: 200; batch size 4, batch_percentage: 0.5, max_len: 400, but the results were the same, joint_epoch was reached, and occasionally slm_out at L502 was not None, but still SLMADV training didn't seem to start, as the related losses were still zero.

[78Alpha]

They're going to be zero for a while unless the conditions it's looking for are met. On about 1 epoch of training, my tensorboard only showed 60 steps worth of SLM training when i set batch percentage to 1. I don't know what exactly is looking for.

[GUUser91]

@78Alpha I'm not sure if I'm doing this correctly, but does this image from my tensorboard folder mean that I was able to start SLM Adversarial Training?

[78Alpha]

Yeah, that's what it should look like. All graphs filled is the sign of all parts working.

[GUUser91]

I tinkered around with the config_ft.yml file. I set Max_Len to 120. I set batch_percentage to 1. I set slmadv_params min_len to 100 and slmadv_params max_len to 120. Batch size is set to 2. Now the DiscLM and GenLM Loss stats are no longer at 0. I'm using a rx 7900 xtx. Note I'm training a model with style diffusion in one fine-tuning session and adversarial training in another session.

Here a pic from my tensorboard folder.

Edit: I discovered I can do style diffusion and SLM adversarial training together in one session. I set max_len to 252, epoch set to 100, batch_size set to 2, batch_percentage set to 1, slmadv_params min_len set to 180, slmadv_params max_len set to 190, diff_epoch to 10, joint_epoch to 50, I'm using the vokan model as the base model.

I also rented out a h100 from runpod and slmadv_params and slmadv_params max_len are at default settings ( min_len: 400 and max_len: 500), batch size at 2 and batch_percentage to 1 and SLM adversarial training never started.

I tinkered around with the config_ft.yml file again. I set Max_Len to 252. I set batch_percentage to 1. I set slmadv_params min_len to 100 and slmadv_params max_len to 500. Batch size is set to 2. Now the DiscLM and GenLM Loss stats are only occasionally at 0. This is done in one session.

Second edit: I rented out a h100 from runpod again. I edit the config_ft.yml file with micro. I set batch_size to 4 and max_len to 500. I set slmadv_params min_len to 100 and slmadv_params max_len to 500 and batch_percentage to 1. And now SLM adversarial training stats is no longer at zero.

Edit: Turns this is bad. I had set slmadv_params min_len even higher to get better quality.

Here's a screenshot of the vram usage.

This is what I did in runpod. I update the repo.

apt update

I install these.

apt install aria2 p7zip-full curl jq micro

I use the pwd command to find directory / filepath infomation.

pwd

I put the training dataset in a zip file and then I upload it to either https://catbox.moe/ or https://litterbox.catbox.moe/ (Which lets you upload a 1GB file)

I download the vokan base model and zip file.

I download the Vokan model and then upload to gofile.io and then download it with gofile-downloader

aria2c -x 16 -s 16 -k 1M https://files.catbox.moe/XXXXXXXX.zip

I unzip the file

7z x XXXXXXXX.zip

I download the gofile upload script file.

aria2c https://raw.githubusercontent.com/Sushrut1101/GoFile-Upload/master/upload.sh

I give the script permissions

chmod +rx upload.sh

I upload the pth file to https://gofile.io/

./upload.sh model.pth

Third edit: I tinkered around with config_ft.yml file and I trained with these settings on my 4090. I installed the latest version of bitsandbytes. Batch_size is set to 2. batch_percentage is set to 1. This was trained with 7 minutes of audio. The prompt is: Huh. Maybe she got off work later than I thought.

Max_Len is set to 252. slmadv_params min_len is set to 180. slmadv_params max_len is set to 190. https://vocaroo.com/1naWqrrK3Via https://vocaroo.com/1gJfNFoAkErL

Max_Len is set to 252. slmadv_params min_len is set to 252. slmadv_params max_len is set to 252. https://vocaroo.com/1mc7AQst5t4u https://vocaroo.com/1hdoRs1MaLQG

Max_Len is set to 260. slmadv_params min_len is set to 260. slmadv_params max_len is set to 260. https://vocaroo.com/1dcr7vhTDXUP https://vocaroo.com/18S9JyR905XH

Max_Len is set to 280. slmadv_params min_len is set to 280. slmadv_params max_len is set to 280. https://vocaroo.com/1fQaJFghUi01 https://vocaroo.com/1fiq3ubdGEYv

[GUUser91]

Another observation. I trained a model with a dataset of 6 minutes. My previous datasets contained a lot of 1-3 seconds audio files. This one contained a lot of files that are more than 3 seconds. I decided to look at the tensorboard file since I noticed the DiscLM and GenLM Loss stats were constantly above a 0. The tensorboard file showed that the d_loss_slm and gen_loss_slm value were above 0. So I decided to look at another tensorboard file that had a dataset containing audio files that were more than 3 seconds.

And another.

And another.

On my config_ft.yml file, Max_Len is set to 280. slmadv_params min_len is set to 280. slmadv_params max_len is set to 280. Batch_size is set to 2. batch_percentage is 1. diff_epoch and joint_epoch starting epoch are set at default. This was done on my 4090.

[PriyamJha0124]

Which one is best source to train Stts2 model?

[GUUser91]

@PriyamJha0124 You mean fine tuning? I used the vokan model for that. https://huggingface.co/ShoukanLabs/Vokan