Fine tuning training errors.

[rmcpantoja]

Hello,
I'm trying to fine tune a model using a dataset of an hour and a half, but during the final part of the training I have problems, but first I will explain:
The model I'm trying to work with is: tts_models/es/css10/vits, which was trained with 🐸TTS version 0.8 and therefore:

1. I've reverted the TTS repository back to the commit that corresponds to the version 0.8.
2. I've taken the file config.json of this pre-trained model already downloaded, adjusting it to the paths of my dataset and some training settings.
3. When training with python train TTS, I get the following error: 'VITS' has no attribute 'disc'. All of these steps can be reproduced in a colab notebook (with the machines offered by Google).

Is there something wrong in the configuration? since I have searched and I can't find anything related to "disc", not even the pretrained model had such a configuration. The only thing I found is rl_disc.
Is there any solution for this? Thank you very much in advance.

[erogol]

The released model has no discriminator to keep its memory footprint smaller. You need to set init_discriminator false and try loading the model.

[rmcpantoja]

The released model has no discriminator to keep its memory footprint smaller. You need to set init_discriminator false and try loading the model.

Hi @erogol init_discriminator was always set to False in my config.json, since I took the configs from the pretrained model just addjusting it to my dataset. Here is the config, taking into account that I moved the repository back to version 0.8 also due to changes in the values for the configs:

config:

{
"output_path": "/content/drive/MyDrive/TTSWork",
"logger_uri": null,
"run_name": "FT",
"project_name": null,
"run_description": "",
"print_step": 25,
"plot_step": 100,
"model_param_stats": false,
"wandb_entity": null,
"dashboard_logger": "tensorboard",
"log_model_step": 1000,
"save_step": 10000,
"save_n_checkpoints": 5,
"save_checkpoints": true,
"save_all_best": false,
"save_best_after": 10000,
"target_loss": null,
"print_eval": false,
"test_delay_epochs": -1,
"run_eval": true,
"distributed_backend": "nccl",
"distributed_url": "tcp://localhost:54321",
"mixed_precision": true,
"epochs": 1000,
"batch_size": 32,
"eval_batch_size": 4,
"grad_clip": [
1000,
1000
],
"scheduler_after_epoch": true,
"lr": 0.0002,
"optimizer": "AdamW",
"optimizer_params": {
"betas": [
0.8,
0.99
],
"eps": 1e-09,
"weight_decay": 0.01
},
"lr_scheduler": "",
"lr_scheduler_params": {},
"use_grad_scaler": false,
"cudnn_enable": true,
"cudnn_deterministic": false,
"cudnn_benchmark": false,
"training_seed": 54321,
"model": "vits",
"num_loader_workers": 8,
"num_eval_loader_workers": 8,
"use_noise_augment": false,
"audio": {
"fft_size": 1024,
"win_length": 1024,
"hop_length": 256,
"frame_shift_ms": null,
"frame_length_ms": null,
"stft_pad_mode": "reflect",
"sample_rate": 22050,
"resample": false,
"preemphasis": 0.0,
"ref_level_db": 20,
"do_sound_norm": false,
"log_func": "np.log",
"do_trim_silence": true,
"trim_db": 45,
"do_rms_norm": false,
"db_level": null,
"power": 1.5,
"griffin_lim_iters": 60,
"num_mels": 80,
"mel_fmin": 0.0,
"mel_fmax": null,
"spec_gain": 1.0,
"do_amp_to_db_linear": false,
"do_amp_to_db_mel": true,
"pitch_fmax": 640.0,
"pitch_fmin": 0.0,
"signal_norm": true,
"min_level_db": -100,
"symmetric_norm": true,
"max_norm": 4.0,
"clip_norm": true,
"stats_path": null
},
"use_phonemes": false,
"phonemizer": null,
"phoneme_language": null,
"compute_input_seq_cache": true,
"text_cleaner": "multilingual_cleaners",
"enable_eos_bos_chars": false,
"test_sentences_file": "",
"phoneme_cache_path": null,
"characters": {
"characters_class": "TTS.tts.utils.text.characters.Graphemes",
"vocab_dict": null,
"pad": "",
"eos": "",
"bos": "",
"blank": "",
"characters": "abcdefghijklmnopqrstuvwxyz\u00af\u00b7\u00df\u00e0\u00e1\u00e2\u00e3\u00e4\u00e6\u00e7\u00e8\u00e9\u00ea\u00eb\u00ec\u00ed\u00ee\u00ef\u00f1\u00f2\u00f3\u00f4\u00f5\u00f6\u00f9\u00fa\u00fb\u00fc\u00ff\u0101\u0105\u0107\u0113\u0119\u011b\u012b\u0131\u0142\u0144\u014d\u0151\u0153\u015b\u016b\u0171\u017a\u017c\u01ce\u01d0\u01d2\u01d4\u0430\u0431\u0432\u0433\u0434\u0435\u0436\u0437\u0438\u0439\u043a\u043b\u043c\u043d\u043e\u043f\u0440\u0441\u0442\u0443\u0444\u0445\u0446\u0447\u0448\u0449\u044a\u044b\u044c\u044d\u044e\u044f\u0451\u0454\u0456\u0457\u0491",
"punctuations": "¡!'(),-.:;¿? ",
"phonemes": null,
"is_unique": true,
"is_sorted": true
},
"add_blank": true,
"batch_group_size": 0,
"loss_masking": null,
"sort_by_audio_len": false,
"min_audio_len": 32768,
"max_audio_len": 132300,
"min_text_len": 1,
"max_text_len": Infinity,
"compute_f0": false,
"compute_linear_spec": true,
"precompute_num_workers": 12,
"start_by_longest": false,
"datasets": [
{
"name": "css10",
"path": "/content/TTS/data",
"meta_file_train": "metadata.csv",
"ignored_speakers": null,
"language": "es",
"meta_file_val": "",
"meta_file_attn_mask": ""
}
],
"test_sentences": [
[
"Un arco\u00edris\u200b o arco iris es un fen\u00f3meno \u00f3ptico y meteorol\u00f3gico que causa la aparici\u00f3n en la atm\u00f3sfera terrestre de un arco multicolor.",
"css10",
null,
"es"
]
],
"eval_split_max_size": null,
"eval_split_size": 0.01,
"use_speaker_weighted_sampler": false,
"speaker_weighted_sampler_alpha": 1.0,
"use_language_weighted_sampler": true,
"language_weighted_sampler_alpha": 1.0,
"model_args": {
"num_chars": 129,
"out_channels": 513,
"spec_segment_size": 32,
"hidden_channels": 192,
"hidden_channels_ffn_text_encoder": 768,
"num_heads_text_encoder": 2,
"num_layers_text_encoder": 6,
"kernel_size_text_encoder": 3,
"dropout_p_text_encoder": 0.1,
"dropout_p_duration_predictor": 0.5,
"kernel_size_posterior_encoder": 5,
"dilation_rate_posterior_encoder": 1,
"num_layers_posterior_encoder": 16,
"kernel_size_flow": 5,
"dilation_rate_flow": 1,
"num_layers_flow": 4,
"resblock_type_decoder": "2",
"resblock_kernel_sizes_decoder": [
3,
5,
7
],
"resblock_dilation_sizes_decoder": [
[
1,
2
],
[
2,
6
],
[
3,
12
]
],
"upsample_rates_decoder": [
8,
8,
4
],
"upsample_initial_channel_decoder": 256,
"upsample_kernel_sizes_decoder": [
16,
16,
8
],
"use_sdp": true,
"noise_scale": 1.0,
"inference_noise_scale": 0.3,
"length_scale": 1,
"noise_scale_dp": 1.0,
"inference_noise_scale_dp": 0.3,
"max_inference_len": null,
"init_discriminator": false,
"use_spectral_norm_disriminator": false,
"use_speaker_embedding": true,
"num_speakers": 0,
"speakers_file": "/content/TTS/tts_models--es--css10--vits/speaker_ids.json",
"d_vector_file": null,
"speaker_embedding_channels": 256,
"use_d_vector_file": false,
"d_vector_dim": 0,
"detach_dp_input": true,
"use_language_embedding": true,
"embedded_language_dim": 4,
"num_languages": 1,
"language_ids_file": "/content/TTS/tts_models--es--css10--vits/language_ids.json",
"use_speaker_encoder_as_loss": false,
"speaker_encoder_config_path": "",
"speaker_encoder_model_path": "",
"condition_dp_on_speaker": true,
"freeze_encoder": false,
"freeze_DP": false,
"freeze_PE": false,
"freeze_flow_decoder": false,
"freeze_waveform_decoder": false
},
"lr_gen": 0.0002,
"lr_disc": 0.0002,
"lr_scheduler_gen": "ExponentialLR",
"lr_scheduler_gen_params": {
"gamma": 0.999875,
"last_epoch": -1
},
"lr_scheduler_disc": "ExponentialLR",
"lr_scheduler_disc_params": {
"gamma": 0.999875,
"last_epoch": -1
},
"kl_loss_alpha": 1.0,
"disc_loss_alpha": 1.0,
"gen_loss_alpha": 1.0,
"feat_loss_alpha": 1.0,
"mel_loss_alpha": 45.0,
"dur_loss_alpha": 1.0,
"speaker_encoder_loss_alpha": 1.0,
"return_wav": true,
"r": 1,
"num_speakers": 1,
"use_speaker_embedding": true,
"speakers_file": "/content/TTS/tts_models--es--css10--vits/speaker_ids.json",
"speaker_embedding_channels": 256,
"language_ids_file": "/content/TTS/tts_models--es--css10--vits/language_ids.json",
"use_language_embedding": true,
"use_d_vector_file": false,
"d_vector_file": null,
"d_vector_dim": 0
}

training log:

Command: !CUDA_VISIBLE_DEVICES="0" python TTS/bin/train_tts.py --config_path "/content/config.json" --restore_path "/content/TTS/tts_models--es--css10--vits/model_file.pth.tar"
| > Found 535 files in /content/TTS/data

Using model: vits
Setting up Audio Processor...
| > sample_rate:22050
| > resample:False
| > num_mels:80
| > log_func:np.log10
| > min_level_db:0
| > frame_shift_ms:None
| > frame_length_ms:None
| > ref_level_db:None
| > fft_size:1024
| > power:None
| > preemphasis:0.0
| > griffin_lim_iters:None
| > signal_norm:None
| > symmetric_norm:None
| > mel_fmin:0
| > mel_fmax:None
| > pitch_fmin:None
| > pitch_fmax:None
| > spec_gain:20.0
| > stft_pad_mode:reflect
| > max_norm:1.0
| > clip_norm:True
| > do_trim_silence:False
| > trim_db:60
| > do_sound_norm:False
| > do_amp_to_db_linear:True
| > do_amp_to_db_mel:True
| > do_rms_norm:False
| > db_level:None
| > stats_path:None
| > base:10
| > hop_length:256
| > win_length:1024
initialization of speaker-embedding layers.
initialization of language-embedding layers.
Training Environment:
| > Current device: 0
| > Num. of GPUs: 1
| > Num. of CPUs: 2
| > Num. of Torch Threads: 1
| > Torch seed: 54321
| > Torch CUDNN: True
| > Torch CUDNN deterministic: False
| > Torch CUDNN benchmark: False
Start Tensorboard: tensorboard --logdir=/content/drive/MyDrive/TTSWork/FT-February-06-2023_03+17PM-2c9f00a8
Traceback (most recent call last):
File "TTS/bin/train_tts.py", line 71, in
main()
File "TTS/bin/train_tts.py", line 58, in main
trainer = Trainer(
File "/usr/local/lib/python3.8/dist-packages/trainer/trainer.py", line 489, in init
self.optimizer = self.get_optimizer(self.model, self.config)
File "/usr/local/lib/python3.8/dist-packages/trainer/trainer.py", line 1767, in get_optimizer
optimizer = model.get_optimizer()
File "/content/TTS/TTS/tts/models/vits.py", line 1643, in get_optimizer
optimizer0 = get_optimizer(self.config.optimizer, self.config.optimizer_params, self.config.lr_disc, self.disc)
File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 1269, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'Vits' object has no attribute 'disc'

[MaxIakovliev]

same issue here. Still persists.
Any recommendations on how to fix is much appreciated!

[rmcpantoja]

same issue here. Still persists. Any recommendations on how to fix is much appreciated!

Hi @MaxIakovliev,
I've fixed it by setting init_discriminator = True in the config.

[MaxIakovliev]

@rmcpantoja - it works! Thanks

[francois-vz]

Could still be due to versioning.

[rmcpantoja]

Could still be due to versioning.

Hi @Fransie-dev Yes, I have rolled back the commits to versions 0.8, 0.7.1 and 0.7 and the problem persists. I could change to a minor version, but the question is that it is not known with which commit that model was trained.
I could train a model from scratch with the current version of 🐸TTS, but I don't have a GPU, only the ones from the free colab where you also have a six hour limit.

[statscol]

the VitsArgs config in which this model was trained has some parameter names different to the current version of TTS.

tts_models/es/css10/vits default config:

VitsArgs(num_chars=129, out_channels=513, spec_segment_size=32, hidden_channels=192, hidden_channels_ffn_text_encoder=768, num_heads_text_encoder=2, num_layers_text_encoder=6, kernel_size_text_encoder=3, dropout_p_text_encoder=0.1, dropout_p_duration_predictor=0.5, kernel_size_posterior_encoder=5, dilation_rate_posterior_encoder=1, num_layers_posterior_encoder=16, kernel_size_flow=5, dilation_rate_flow=1, num_layers_flow=4, resblock_type_decoder='2', resblock_kernel_sizes_decoder=[3, 5, 7], resblock_dilation_sizes_decoder=[[1, 2], [2, 6], [3, 12]], upsample_rates_decoder=[8, 8, 4], upsample_initial_channel_decoder=256, upsample_kernel_sizes_decoder=[16, 16, 8], periods_multi_period_discriminator=[2, 3, 5, 7, 11], use_sdp=True, noise_scale=1.0, inference_noise_scale=0.3, length_scale=1.0, noise_scale_dp=1.0, inference_noise_scale_dp=0.3, max_inference_len=None, init_discriminator=False, use_spectral_norm_disriminator=False, use_speaker_embedding=True, num_speakers=0, speakers_file='/root/.local/share/tts/tts_models--es--css10--vits/speaker_ids.json', d_vector_file=None, speaker_embedding_channels=256, use_d_vector_file=False, d_vector_dim=0, detach_dp_input=True, use_language_embedding=True, embedded_language_dim=4, num_languages=1, language_ids_file=None, use_speaker_encoder_as_loss=False, speaker_encoder_config_path='', speaker_encoder_model_path='', condition_dp_on_speaker=True, freeze_encoder=False, freeze_DP=False, freeze_PE=False, freeze_flow_decoder=False, freeze_waveform_decoder=False, encoder_sample_rate=None, interpolate_z=True, reinit_DP=False, reinit_text_encoder=False)

using the same args with an updated version of TTS

VitsArgs(num_chars=129, out_channels=513, spec_segment_size=32, hidden_channels=192, hidden_channels_ffn_text_encoder=768, num_heads_text_encoder=2, num_layers_text_encoder=6, kernel_size_text_encoder=3, dropout_p_text_encoder=0.1, dropout_p_duration_predictor=0.5, kernel_size_posterior_encoder=5, dilation_rate_posterior_encoder=1, num_layers_posterior_encoder=16, kernel_size_flow=5, dilation_rate_flow=1, num_layers_flow=4, resblock_type_decoder='2', resblock_kernel_sizes_decoder=[3, 5, 7], resblock_dilation_sizes_decoder=[[1, 2], [2, 6], [3, 12]], upsample_rates_decoder=[8, 8, 4], upsample_initial_channel_decoder=256, upsample_kernel_sizes_decoder=[16, 16, 8], periods_multi_period_discriminator=[2, 3, 5, 7, 11], use_sdp=True, noise_scale=1.0, inference_noise_scale=0.3, length_scale=1, noise_scale_dp=1.0, inference_noise_scale_dp=1.0, max_inference_len=None, init_discriminator=True, use_spectral_norm_disriminator=False, use_speaker_embedding=True, num_speakers=0, speakers_file='/root/.local/share/tts/tts_models--es--css10--vits/speaker_ids.json', d_vector_file=None, speaker_embedding_channels=256, use_d_vector_file=False, d_vector_dim=0, detach_dp_input=True, use_language_embedding=True, embedded_language_dim=4, num_languages=1, language_ids_file=None, use_speaker_encoder_as_loss=False, speaker_encoder_config_path='', speaker_encoder_model_path='', condition_dp_on_speaker=True, freeze_encoder=False, freeze_DP=False, freeze_PE=False, freeze_flow_decoder=False, freeze_waveform_decoder=False, encoder_sample_rate=None, interpolate_z=True, reinit_DP=False, reinit_text_encoder=False)

This is how i managed to fine-tune the model.

[rmcpantoja]

Hi @statscol , so that was the problem? So, don't you necessarily need the discriminator of the pretrained model to fine tuning? Anyway, thanks a lot for the comparison. I'll try to train a Spanish ljspeech dataset in a few hours and i will post the results.

[nicemanis]

I found out that there are differences in layer names between versions:

Maybe fixing layer names resolves the issue

[youness-658]

i m trying to fine toning the coqui xtts v2 but i get this ereur : Traceback (most recent call last):
File "C:\Users\zh\Desktop\project ai\TTS\TTS\bin\train_tts.py", line 71, in
main()
File "C:\Users\zh\Desktop\project ai\TTS\TTS\bin\train_tts.py", line 58, in main
trainer = Trainer(
File "C:\Users\zh\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\LocalCache\local-packages\Python310\site-packages\trainer\trainer.py", line 506, in init
self.criterion = self.get_criterion(self.model)
File "C:\Users\zh\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\LocalCache\local-packages\Python310\site-packages\trainer\trainer.py", line 2098, in get_criterion
criterion = model.get_criterion()
File "C:\Users\zh\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\LocalCache\local-packages\Python310\site-packages\torch\nn\modules\module.py", line 1940, in getattr
raise AttributeError(
AttributeError: 'Xtts' object has no attribute 'get_criterion'

anyone can help please

[eginhard]

This should be fixed in our fork (available via pip install coqui-tts). This repository is not maintained anymore.