search

kan-bayashi / ParallelWaveGAN

Unofficial Parallel WaveGAN (+ MelGAN & Multi-band MelGAN & HiFi-GAN & StyleMelGAN) with Pytorch
https://kan-bayashi.github.io/ParallelWaveGAN/
MIT License
1.55k stars 340 forks source link

Loss value became NaN in the middle of the training #184

Closed xxoospring closed 4 years ago

xxoospring commented 4 years ago

when it came to 80K steps, all the prediction generated audios were blank audio: image

kan-bayashi commented 4 years ago

Please describe more info.

xxoospring commented 4 years ago

Please describe more info.

  • What kind of recipe (or dataset) did you use?
  • Which model did you train? (Attach your config)
  • How was the training curve?
kan-bayashi commented 4 years ago

It seems loss value became NaN around 80k steps. I'm not sure which loss became NaN. I've never met loss = NaN so I'm wondering your data or config is something wrong. Please paste your whole config here.

xxoospring commented 4 years ago

It seems loss value became NaN around 80k steps. I'm not sure which loss became NaN. I've never met loss = NaN so I'm wondering your data or config is something wrong. Please paste your whole config here.


# This is the hyperparameter configuration file for MelGAN.
# Please make sure this is adjusted for the LJSpeech dataset. If you want to
# apply to the other dataset, you might need to carefully change some parameters.
# This configuration requires ~ 8GB memory and will finish within 5 days on Titan V.

This configuration is based on multi-band MelGAN but the hop size and sampling

rate is different from the paper (16kHz vs 22.05kHz). The number of iteraions

is now shown in the paper so currently we train 1M iterations (not sure enough

to converge). Compared to multi_band_melgan.v1 config, Adam optimizer without

gradient norm is used, which is based on @dathudeptrai advice.

https://github.com/kan-bayashi/ParallelWaveGAN/issues/143#issuecomment-632539906

We found that the use of small batch_max_steps (e.g. 8192) has no bad effect for

the quality, so if you want to accelerate the training, please reduce it.

###########################################################

FEATURE EXTRACTION SETTING

########################################################### sampling_rate: 22050 # Sampling rate. fft_size: 1024 # FFT size. hop_size: 256 # Hop size. win_length: null # Window length.

If set to null, it will be the same as fft_size.

window: "hann" # Window function. num_mels: 80 # Number of mel basis. fmin: 80 # Minimum freq in mel basis calculation.

modify by yuanxi corresponding with NVIDIA-tacotron2

fmax: 8000 # Maximum frequency in mel basis calculation.

fmax: 7800 # Maximum frequency in mel basis calculation.

global_gain_scale: 1.0 # Will be multiplied to all of waveform. trim_silence: true # Whether to trim the start and end of silence. trim_threshold_in_db: 60 # Need to tune carefully if the recording is not good. trim_frame_size: 2048 # Frame size in trimming. trim_hop_size: 512 # Hop size in trimming. format: "hdf5" # Feature file format. "npy" or "hdf5" is supported.

###########################################################

GENERATOR NETWORK ARCHITECTURE SETTING

########################################################### generator_type: "MelGANGenerator" # Generator type. generator_params: in_channels: 80 # Number of input channels. out_channels: 4 # Number of output channels. kernel_size: 7 # Kernel size of initial and final conv layers. channels: 384 # Initial number of channels for conv layers. upsample_scales: [8, 4, 2] # List of Upsampling scales. stack_kernel_size: 3 # Kernel size of dilated conv layers in residual stack. stacks: 4 # Number of stacks in a single residual stack module. use_weight_norm: True # Whether to use weight normalization. use_causal_conv: False # Whether to use causal convolution.

###########################################################

DISCRIMINATOR NETWORK ARCHITECTURE SETTING

########################################################### discriminator_type: "MelGANMultiScaleDiscriminator" # Discriminator type. discriminator_params: in_channels: 1 # Number of input channels. out_channels: 1 # Number of output channels. scales: 3 # Number of multi-scales. downsample_pooling: "AvgPool1d" # Pooling type for the input downsampling. downsample_pooling_params: # Parameters of the above pooling function. kernel_size: 4 stride: 2 padding: 1 count_include_pad: False kernel_sizes: [5, 3] # List of kernel size. channels: 16 # Number of channels of the initial conv layer. max_downsample_channels: 512 # Maximum number of channels of downsampling layers. downsample_scales: [4, 4, 4] # List of downsampling scales. nonlinear_activation: "LeakyReLU" # Nonlinear activation function. nonlinear_activation_params: # Parameters of nonlinear activation function. negative_slope: 0.2 use_weight_norm: True # Whether to use weight norm.

###########################################################

STFT LOSS SETTING

########################################################### stft_loss_params: fft_sizes: [1024, 2048, 512] # List of FFT size for STFT-based loss. hop_sizes: [120, 240, 50] # List of hop size for STFT-based loss win_lengths: [600, 1200, 240] # List of window length for STFT-based loss. window: "hann_window" # Window function for STFT-based loss use_subband_stft_loss: true subband_stft_loss_params: fft_sizes: [384, 683, 171] # List of FFT size for STFT-based loss. hop_sizes: [30, 60, 10] # List of hop size for STFT-based loss win_lengths: [150, 300, 60] # List of window length for STFT-based loss. window: "hann_window" # Window function for STFT-based loss

###########################################################

ADVERSARIAL LOSS SETTING

########################################################### use_feat_match_loss: false # Whether to use feature matching loss. lambda_adv: 2.5 # Loss balancing coefficient for adversarial loss.

###########################################################

DATA LOADER SETTING

########################################################### batch_size: 64 # Batch size. batch_max_steps: 16384 # Length of each audio in batch. Make sure dividable by hop_size. pin_memory: true # Whether to pin memory in Pytorch DataLoader. num_workers: 4 # Number of workers in Pytorch DataLoader. remove_short_samples: true # Whether to remove samples the length of which are less than batch_max_steps. allow_cache: true # Whether to allow cache in dataset. If true, it requires cpu memory.

###########################################################

OPTIMIZER & SCHEDULER SETTING

########################################################### generator_optimizer_type: "Adam" # Generator's optimizer type. generator_optimizer_params: lr: 1.0e-3 # Generator's learning rate. eps: 1.0e-7 # Generator's epsilon. weight_decay: 0.0 # Generator's weight decay coefficient. amsgrad: true generator_grad_norm: -1 # Generator's gradient norm. generator_scheduler_type: "MultiStepLR" # Generator's scheduler type. generator_scheduler_params: gamma: 0.5 # Generator's scheduler gamma. milestones: # At each milestone, lr will be multiplied by gamma.

###########################################################

INTERVAL SETTING

########################################################### discriminator_train_start_steps: 200000 # Number of steps to start to train discriminator. train_max_steps: 1000000 # Number of training steps. save_interval_steps: 50000 # Interval steps to save checkpoint. eval_interval_steps: 1000 # Interval steps to evaluate the network. log_interval_steps: 1000 # Interval steps to record the training log.

###########################################################

OTHER SETTING

########################################################### num_save_intermediate_results: 4 # Number of results to be saved as intermediate results.

kan-bayashi commented 4 years ago

Config seems fine. Maybe your data has some perfect silence part? (Not sure) Following part is doubtful. Why don't you try to use clamp here?

https://github.com/kan-bayashi/ParallelWaveGAN/blob/88a829b5ef923657b4fb590e0614dec5e106a971/parallel_wavegan/losses/stft_loss.py#L52 https://github.com/kan-bayashi/ParallelWaveGAN/blob/88a829b5ef923657b4fb590e0614dec5e106a971/parallel_wavegan/losses/stft_loss.py#L73

xxoospring commented 4 years ago

Config seems fine. Maybe your data has some perfect silence part? (Not sure) Following part is doubtful. Why don't you try to use clamp here?

https://github.com/kan-bayashi/ParallelWaveGAN/blob/88a829b5ef923657b4fb590e0614dec5e106a971/parallel_wavegan/losses/stft_loss.py#L52

https://github.com/kan-bayashi/ParallelWaveGAN/blob/88a829b5ef923657b4fb590e0614dec5e106a971/parallel_wavegan/losses/stft_loss.py#L73

I didnt modify these code, I just modify run.sh and conf file

kan-bayashi commented 4 years ago

I mean, please modify these parts to avoid NaN errors by yourself and check how it works.

kan-bayashi commented 4 years ago

I checked the code but the following clamp should avoid NaN errors. https://github.com/kan-bayashi/ParallelWaveGAN/blob/88a829b5ef923657b4fb590e0614dec5e106a971/parallel_wavegan/losses/stft_loss.py#L31 I do not have an idea of why NaN has happened in your training.