RuntimeError: The expanded size of the tensor (64) must match the existing size (112) at non-singleton dimension 2. Target sizes: [64, 80, 64]. Tensor sizes: [64, 1, 112]

[a-froghyar]

Hey,

I'm trying to run a training with Tacotron 1 using GST. I get the error on the first batch already.

Pytorch version: 1.8 and 1.7.1 (both yielded the same error) Python version: 3.8.0

Traceback (most recent call last): File "TTS/bin/train_tacotron.py", line 721, in <module> main(args) File "TTS/bin/train_tacotron.py", line 619, in main train_avg_loss_dict, global_step = train(train_loader, model, File "TTS/bin/train_tacotron.py", line 168, in train decoder_output, postnet_output, alignments, stop_tokens = model( File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl result = self.forward(*input, **kwargs) File "/home/big-boy/projects/TTS/TTS/tts/models/tacotron.py", line 173, in forward decoder_outputs = decoder_outputs * output_mask.unsqueeze(1).expand_as(decoder_outputs) RuntimeError: The expanded size of the tensor (64) must match the existing size (112) at non-singleton dimension 2. Target sizes: [64, 80, 64]. Tensor sizes: [64, 1, 112]

My hyperparams: // TRAINING "batch_size": 64, "eval_batch_size": 16, "r": 4, "gradual_training": [ [0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 1, 32] ], "mixed_precision": true,

// MULTI-SPEAKER and GST "use_speaker_embedding": false, // use speaker embedding to enable multi-speaker learning. "use_gst": true, "use_external_speaker_embedding_file": false, "external_speaker_embedding_file": "../../speakers-vctk-en.json", "gst": { // gst parameter if gst is enabled "gst_style_input": null, // Condition the style input either on a // -> wave file [path to wave] or // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15} // with the dictionary being len(dict) <= len(gst_style_tokens). "gst_embedding_dim": 512, "gst_num_heads": 4, "gst_style_tokens": 10, "gst_use_speaker_embedding": false },

[erogol]

@WeberJulian has an idea as he is also using GST

[lexkoro]

"r": 4,

set this to "r": 7

I think since you are using gradual training it loads the first batch of data with r=4 but tries to train with r=7

---

Edit: Just checked and got the same error

Traceback (most recent call last):
  File "TTS/bin/train_tacotron.py", line 721, in <module>
    main(args)
  File "TTS/bin/train_tacotron.py", line 623, in main
    scaler_st)
  File "TTS/bin/train_tacotron.py", line 166, in train
    text_input, text_lengths, mel_input, mel_lengths, speaker_ids=speaker_ids, speaker_embeddings=speaker_embeddings)
  File "/media/alexander/LinuxFS/Projects/TTS-dev/.venv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/media/alexander/LinuxFS/Projects/TTS-dev/TTS/tts/models/tacotron.py", line 173, in forward
    decoder_outputs = decoder_outputs * output_mask.unsqueeze(1).expand_as(decoder_outputs)
RuntimeError: The expanded size of the tensor (216) must match the existing size (378) at non-singleton dimension 2.  Target sizes: [64, 80, 216].  Tensor sizes: [64, 1, 378]

The error is not related to GST. Just set r=7 in your config and it should work.

[WeberJulian]

Hum the error I had was with speaker_id initialisation in eval, but here it's concerning gst and it's during training. Here my best advice would be to compare tacotron and tacotron 2 in the debugger.

[a-froghyar]

Hey @WeberJulian thanks for your answer. What should I be looking for when I compare the two?

[WeberJulian]

@a-froghyar I would look for differences in shapes of the different tensors (like encoder_outputs before and after compute_gst)

[WeberJulian]

Have you checked yet if the same config in tacotron 2 works for you ?

[a-froghyar]

@WeberJulian not yet, I'll be on it later today and I'll report back, thanks again!

[a-froghyar]

@WeberJulian @erogol I'm trying to set up my debuging config, could you shed some light on debugging best practices please? :)

[erogol]

@a-froghyar have you checked @SanjaESC answer?

[a-froghyar]

@erogol yes I did, same error with T1, T2 threw a librosa error, my Dataset seems to be stereo apparently, will check back. librosa.util.exceptions.ParameterError: Invalid shape for monophonic audio: ndim=2, shape=(19680, 2) - I've checked and weirdly SOME of the wav files in my corpus are stereo, going to convert to mono now and try again.

[erogol]

for debugging you can use the small dataset set under tests folder. There are also some sample configs for model testing that you can copy and paste for debugging.

[a-froghyar]

Thanks, what I meant was actually how to launch the debugger in the IDE using a specific config.json for the training run? If I launch the debugger on train_tacotron.py, how do I feed it the config that I'm using?

[a-froghyar]

Solved my problem, sorry for the noob question, just needed to add the argument into my launch.json file like so: "args": ["--config_path", "TTS/tts/configs/gst_blizzard.json"],

where gst_blizzard.json is my custom config file

[a-froghyar]

@WeberJulian I've realised that I had a higher numpy version before, I downgraded to the one in the requirements and T2 is running, T1 now throws this error: loss = functional.l1_loss(x * mask, target * mask, reduction='sum') RuntimeError: The size of tensor a (80) must match the size of tensor b (513) at non-singleton dimension 2

After a few iterations, target - [64,111,80] actually has a different size to mask - [64,111,513]

[erogol]

t1 needs linear spectrograms. I guess you set the dataloader to return melspectrograms or there is something wrong in your settings. Pls pot tyour config here. @a-froghyar

[erogol]

For debugging I don't use the IDE. I just use breakpoint() or import pdb; pdb.set_trace() depending on th pythonversion and co through the code on termina l

[a-froghyar]

Thanks, doesn't t1 need Mel and Linear? Here's my config:

{ "model": "Tacotron", "run_name": "blizzard-gts", "run_description": "tacotron with DDC and differential spectral loss.",

// AUDIO PARAMETERS
"audio": {
    // stft parameters
    "fft_size": 1024, // number of stft frequency levels. Size of the linear spectogram frame.
    "win_length": 1024, // stft window length in ms.
    "hop_length": 256, // stft window hop-lengh in ms.
    "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
    "frame_shift_ms": null, // stft window hop-lengh in ms. If null, 'hop_length' is used.

    // Audio processing parameters
    "sample_rate": 24000, // DATASET-RELATED: wav sample-rate.
    "preemphasis": 0.0, // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
    "ref_level_db": 20, // reference level db, theoretically 20db is the sound of air.

    // Silence trimming
    "do_trim_silence": true, // enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
    "trim_db": 60, // threshold for timming silence. Set this according to your dataset.

    // Griffin-Lim
    "power": 1.5, // value to sharpen wav signals after GL algorithm.
    "griffin_lim_iters": 60, // #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.

    // MelSpectrogram parameters
    "num_mels": 80, // size of the mel spec frame.
    "mel_fmin": 95.0, // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
    "mel_fmax": 12000.0, // maximum freq level for mel-spec. Tune for dataset!!
    "spec_gain": 1,

    // Normalization parameters
    "signal_norm": true, // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
    "min_level_db": -100, // lower bound for normalization
    "symmetric_norm": true, // move normalization to range [-1, 1]
    "max_norm": 4.0, // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
    "clip_norm": true, // clip normalized values into the range.
    "stats_path": null // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
},

// VOCABULARY PARAMETERS
// if custom character set is not defined,
// default set in symbols.py is used
// "characters":{
//     "pad": "_",
//     "eos": "~",
//     "bos": "^",
//     "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
//     "punctuations":"!'(),-.:;? ",
//     "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
// },

// DISTRIBUTED TRAINING
"distributed": {
    "backend": "nccl",
    "url": "tcp:\/\/localhost:54321"
},

"reinit_layers": [], // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.

// TRAINING
"batch_size": 64, // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
"eval_batch_size": 16,
"r": 7, // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
"gradual_training": [
    [0, 7, 64],
    [1, 5, 64],
    [50000, 3, 32],
    [130000, 2, 32],
    [290000, 1, 32]
], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
"mixed_precision": true, // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.

// LOSS SETTINGS
"loss_masking": false, // enable / disable loss masking against the sequence padding.
"decoder_loss_alpha": 0.5, // original decoder loss weight. If > 0, it is enabled
"postnet_loss_alpha": 0.25, // original postnet loss weight. If > 0, it is enabled
"postnet_diff_spec_alpha": 0.25, // differential spectral loss weight. If > 0, it is enabled
"decoder_diff_spec_alpha": 0.25, // differential spectral loss weight. If > 0, it is enabled
"decoder_ssim_alpha": 0.5, // decoder ssim loss weight. If > 0, it is enabled
"postnet_ssim_alpha": 0.25, // postnet ssim loss weight. If > 0, it is enabled
"ga_alpha": 5.0, // weight for guided attention loss. If > 0, guided attention is enabled.
"stopnet_pos_weight": 15.0, // pos class weight for stopnet loss since there are way more negative samples than positive samples.

// VALIDATION
"run_eval": true,
"test_delay_epochs": 10, //Until attention is aligned, testing only wastes computation time.
"test_sentences_file": null, // set a file to load sentences to be used for testing. If it is null then we use default english sentences.

// OPTIMIZER
"noam_schedule": false, // use noam warmup and lr schedule.
"grad_clip": 1.0, // upper limit for gradients for clipping.
"epochs": 300000, // total number of epochs to train.
"lr": 0.0001, // Initial learning rate. If Noam decay is active, maximum learning rate.
"wd": 0.000001, // Weight decay weight.
"warmup_steps": 4000, // Noam decay steps to increase the learning rate from 0 to "lr"
"seq_len_norm": false, // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.

// TACOTRON PRENET
"memory_size": -1, // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
"prenet_type": "original", // "original" or "bn".
"prenet_dropout": true, // enable/disable dropout at prenet.

// TACOTRON ATTENTION
"attention_type": "graves", // 'original' , 'graves', 'dynamic_convolution'
"attention_heads": 4, // number of attention heads (only for 'graves')
"attention_norm": "sigmoid", // softmax or sigmoid.
"windowing": false, // Enables attention windowing. Used only in eval mode.
"use_forward_attn": false, // if it uses forward attention. In general, it aligns faster.
"forward_attn_mask": false, // Additional masking forcing monotonicity only in eval mode.
"transition_agent": false, // enable/disable transition agent of forward attention.
"location_attn": true, // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
"bidirectional_decoder": false, // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
"double_decoder_consistency": false, // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
"ddc_r": 7, // reduction rate for coarse decoder.

// STOPNET
"stopnet": true, // Train stopnet predicting the end of synthesis.
"separate_stopnet": true, // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.

// TENSORBOARD and LOGGING
"print_step": 25, // Number of steps to log training on console.
"tb_plot_step": 100, // Number of steps to plot TB training figures.
"print_eval": false, // If True, it prints intermediate loss values in evalulation.
"save_step": 5000, // Number of training steps expected to save traninpg stats and checkpoints.
"checkpoint": true, // If true, it saves checkpoints per "save_step"
"tb_model_param_stats": false, // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.

// DATA LOADING
"text_cleaner": "phoneme_cleaners",
"enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
"num_loader_workers": 8, // number of training data loader processes. Don't set it too big. 4-8 are good values.
"num_val_loader_workers": 8, // number of evaluation data loader processes.
"batch_group_size": 4, //Number of batches to shuffle after bucketing.
"min_seq_len": 6, // DATASET-RELATED: minimum text length to use in training
"max_seq_len": 153, // DATASET-RELATED: maximum text length
"compute_input_seq_cache": false, // if true, text sequences are computed before starting training. If phonemes are enabled, they are also computed at this stage.
"use_noise_augment": true,

// PATHS
"output_path": "/home/big-boy/Models/Blizzard/",

// PHONEMES
"phoneme_cache_path": "/home/big-boy/Models/phoneme_cache/", // phoneme computation is slow, therefore, it caches results in the given folder.
"use_phonemes": true, // use phonemes instead of raw characters. It is suggested for better pronounciation.
"phoneme_language": "en-us", // depending on your target language, pick one from  https://github.com/bootphon/phonemizer#languages

// MULTI-SPEAKER and GST
"use_speaker_embedding": false, // use speaker embedding to enable multi-speaker learning.
"use_gst": true, // use global style tokens
"use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
"external_speaker_embedding_file": "../../speakers-vctk-en.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
"gst": { // gst parameter if gst is enabled
    "gst_style_input": null, // Condition the style input either on a
    // -> wave file [path to wave] or
    // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
    // with the dictionary being len(dict) <= len(gst_style_tokens).
    "gst_embedding_dim": 512,
    "gst_num_heads": 4,
    "gst_style_tokens": 10,
    "gst_use_speaker_embedding": false
},

// DATASETS
"datasets": // List of datasets. They all merged and they get different speaker_ids.
    [{
        "name": "ljspeech",
        "path": "/home/big-boy/Data/blizzard2013/segmented/",
        "meta_file_train": "metadata.csv", // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
        "meta_file_val": null
    }]

}

[erogol]

can you post the full error here? I need the line number/

[a-froghyar]

/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/modules/loss.py:94: UserWarning: Using a target size (torch.Size([64, 90, 80])) that is different to the input size (torch.Size([64, 90, 513])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size. return F.l1_loss(input, target, reduction=self.reduction) ! Run is removed from /home/big-boy/Models/Blizzard/blizzard-gts-March-11-2021_05+38PM-45068a9 Traceback (most recent call last): File "/home/big-boy/projects/TTS/TTS/bin/train_tacotron.py", line 721, in main(args) File "/home/big-boy/projects/TTS/TTS/bin/train_tacotron.py", line 619, in main train_avg_loss_dict, global_step = train(train_loader, model, File "/home/big-boy/projects/TTS/TTS/bin/train_tacotron.py", line 180, in train loss_dict = criterion(postnet_output, decoder_output, mel_input, File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, kwargs) File "/home/big-boy/projects/TTS/TTS/tts/layers/losses.py", line 377, in forward postnet_diff_spec_loss = self.criterion_diff_spec(postnet_output, mel_input, output_lens) File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(*input, *kwargs) File "/home/big-boy/projects/TTS/TTS/tts/layers/losses.py", line 203, in forward return self.loss_func(x_diff, target_diff) File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl result = self.forward(input, kwargs) File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/modules/loss.py", line 94, in forward return F.l1_loss(input, target, reduction=self.reduction) File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/nn/functional.py", line 2633, in l1_loss expanded_input, expanded_target = torch.broadcast_tensors(input, target) File "/home/big-boy/anaconda3/envs/PyCapacitron/lib/python3.8/site-packages/torch/functional.py", line 71, in broadcast_tensors return _VF.broadcast_tensors(tensors) # type: ignore RuntimeError: The size of tensor a (513) must match the size of tensor b (80) at non-singleton dimension 2

[erogol]

Ok I figured. It's a :bug: :)

[a-froghyar]

If I understand correctly it's the linear spectrogram that gets passed down instead of the mel? The linear is only needed for the Griffin Lin conversion, right?

[erogol]

There is no "need" but it just works better with GL.

[erogol]

@a-froghyar dev branch should work now

[a-froghyar]

Thanks so much!

[a-froghyar]

@erogol I was actually working in the main branch, I got the following error in the dev branch, not sure how it got to try to phonemize chinese:

Traceback (most recent call last): File "TTS/bin/train_tacotron.py", line 13, in from TTS.utils.arguments import parse_arguments, process_args File "/home/big-boy/projects/TTS/TTS/utils/arguments.py", line 11, in from TTS.tts.utils.text.symbols import parse_symbols File "/home/big-boy/projects/TTS/TTS/tts/utils/text/init.py", line 11, in from TTS.tts.utils.chinese_mandarin.phonemizer import chinese_text_to_phonemes File "/home/big-boy/projects/TTS/TTS/tts/utils/chinese_mandarin/phonemizer.py", line 3, in import pypinyin ModuleNotFoundError: No module named 'pypinyin'

[WeberJulian]

did you pip install -e . again after the checkout ?

[a-froghyar]

Training is running, thank you guys so much!

I had to downgrade to librosa==0.6.3, though because of this:

librosa.util.exceptions.ParameterError: Audio buffer is not Fortran-contiguous. Use numpy.asfortranarray to ensure Fortran contiguity.

[lpierron]

I have the same problem with v0.0.12:

 CUDA_VISIBLE_DEVICES="0" python ../../TTS/bin/train_tacotron.py --config_path model_config.json
 > Using CUDA:  True
 > Number of GPUs:  1
 > Git Hash: 59ab268
 > Experiment folder: /home/lpierron/Mozilla_TTS/CORPUS_LP/Models/maiLabs-fr-dca/mailabs-fr-ddc-April-23-2021_03+38PM-59ab268
 > Setting up Audio Processor...
 | > sample_rate:22050
 | > resample:False
 | > num_mels:80
 | > min_level_db:-100
 | > frame_shift_ms:None
 | > frame_length_ms:None
 | > ref_level_db:20
 | > fft_size:1024
 | > power:1.5
 | > preemphasis:0.0
 | > griffin_lim_iters:60
 | > signal_norm:True
 | > symmetric_norm:True
 | > mel_fmin:0
 | > mel_fmax:8000.0
 | > spec_gain:1.0
 | > stft_pad_mode:reflect
 | > max_norm:4.0
 | > clip_norm:True
 | > do_trim_silence:True
 | > trim_db:60
 | > do_sound_norm:False
 | > stats_path:./scale_stats.npy
 | > log_func:<ufunc 'log10'>
 | > exp_func:<function AudioProcessor.__init__.<locals>.<lambda> at 0x7f1ef7ac6c10>
 | > hop_length:256
 | > win_length:1024
 | > /tmp/tts/by_book/female/ezwa/monsieur_lecoq/metadata.csv
 | > Found 14211 files in /tmp/tts
 > Using model: Tacotron2

 > Model has 28183506 parameters
 > Starting with inf best loss.

 > DataLoader initialization
 | > Use phonemes: True
   | > phoneme language: fr-fr
 | > Number of instances : 14069
 | > Max length sequence: 281
 | > Min length sequence: 3
 | > Avg length sequence: 105.0826640130784
 | > Num. instances discarded by max-min (max=153, min=6) seq limits: 2420
 | > Batch group size: 128.

 > EPOCH: 0/1000

 > Number of output frames: 7

 > TRAINING (2021-04-23 15:38:18)
 ! Run is removed from /home/lpierron/Mozilla_TTS/CORPUS_LP/Models/maiLabs-fr-dca/mailabs-fr-ddc-April-23-2021_03+38PM-59ab268
Traceback (most recent call last):
  File "../../TTS/bin/train_tacotron.py", line 744, in <module>
    main(args)
  File "../../TTS/bin/train_tacotron.py", line 704, in main
    train_avg_loss_dict, global_step = train(
  File "../../TTS/bin/train_tacotron.py", line 198, in train
    decoder_output, postnet_output, alignments, stop_tokens = model(
  File "/home/lpierron/miniconda3/envs/tts/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/lpierron/Mozilla_TTS/COQUI-TTS/TTS/TTS/tts/models/tacotron2.py", line 226, in forward
    decoder_outputs = decoder_outputs * output_mask.unsqueeze(1).expand_as(decoder_outputs)
RuntimeError: The expanded size of the tensor (12) must match the existing size (84) at non-singleton dimension 2.  Target sizes: [64, 80, 12].  Tensor sizes: [64, 1, 84]

I have downgraded librosa==0.6.3 but it doesn't work.

See my configuration next:

model_config.json.txt