Change the configuration for training

Show usage

There are two ways to show the command line options: --help and --print_config

# Show the command line option
python -m espnet2.bin.asr_train --help
# Show the all configuration in yaml format
python -m espnet2.bin.asr_train --print_config

In this section, we use espnet2.bin.asr_train for an example, but the other training tools based on Task class have the same interface, so you can replace it to another command.

Note that ESPnet2 always selects_ instead of - for the separation for the option name to avoid confusion.

# Bad
--batch-size
# Good
--batch_size

A notable feature of --print_config is that it shows the configuration parsing with the given arguments dynamically: You can look up the parameters for a changeable class.

% # Show parameters of Adam optimizer
% python -m espnet2.bin.asr_train --optim adam --print_config
...
optim: adam
optim_conf:
    lr: 0.001
    betas:
    - 0.9
    - 0.999
    eps: 1.0e-08
    weight_decay: 0
    amsgrad: false
...
% # Show parameters of ReduceLROnPlateau scheduler
% python -m espnet2.bin.asr_train --scheduler ReduceLROnPlateau --print_config
...
scheduler: reducelronplateau
scheduler_conf:
    mode: min
    factor: 0.1
    patience: 10
    verbose: false
    threshold: 0.0001
    threshold_mode: rel
    cooldown: 0
    min_lr: 0
    eps: 1.0e-08
...

Configuration file

You can find the configuration files for DNN training in conf/train_*.yaml.

ls conf/

We adopt ConfigArgParse for this configuration system. The configuration in YAML format has an equivalent effect to the command line argument. e.g. The following two are equivalent:

# config.yaml
foo: 3
bar: 4
python -m espnet2.bin.asr_train --config conf/config.yaml
python -m espnet2.bin.asr_train --foo 3 --bar 4

Change the configuration for dict type value

Some parameters are named as *_conf, e.g. optim_conf, decoder_conf and they has the dict type value. We also provide a way to configure the nested value in such a dict object.

# e.g. Change parameters one by one
python -m espnet2.bin.asr_train --optim_conf lr=0.1 --optim_conf rho=0.8
# e.g. Give the parameters in yaml format
python -m espnet2.bin.asr_train --optim_conf "{lr: 0.1, rho: 0.8}"

Resume training process

python -m espnet2.bin.asr_train --resume true

The state of the training process is saved at the end of every epoch as checkpoint.pth and the training process can be resumed from the start of the next epoch. Checkpoint includes the following states.

  • Model state

  • Optimizer states

  • Scheduler states

  • Reporter state

  • torch.cuda.amp state (from torch=1.6)

Transfer learning / Fine tuning using pretrained model

Use --init_param <file_path>:<src_key>:<dst_key>:<exclude_keys>

# Load all parameters
python -m espnet2.bin.asr_train --init_param model.pth
# Load only the parameters starting with "decoder"
python -m espnet2.bin.asr_train --init_param model.pth:decoder
# Load only the parameters starting with "decoder" and set it to model.decoder
python -m espnet2.bin.asr_train --init_param model.pth:decoder:decoder
# Set parameters to model.decoder
python -m espnet2.bin.asr_train --init_param decoder.pth::decoder
# Load all parameters excluding "decoder.embed"
python -m espnet2.bin.asr_train --init_param model.pth:::decoder.embed
# Load all parameters excluding "encoder" and "decoder.embed"
python -m espnet2.bin.asr_train --init_param model.pth:::encoder,decoder.embed

Freeze parameters

python -m espnet2.bin.asr_train --freeze_param encoder.enc encoder.decoder

Change logging interval

The result in the middle state of the training will be shown by the specified number:

python -m espnet2.bin.asr_train --log_interval 100

Change the number of iterations in each epoch

By default, an epoch indicates using up whole data in the training corpus and the following steps will also run after training for every epoch:

  • Validation

  • Saving model and checkpoint

  • Show result in the epoch

Sometimes the validation after training with a whole corpus is too coarse if using large corpus. For that case, --num_iters_per_epoch can restrict the number of iteration of each epoch.

python -m espnet2.bin.asr_train --num_iters_per_epoch 1000

Note that the training process can’t be resumed at the middle of an epoch because data iterators are stateless, but don’t worry it! Our iterator is built at the start of each epoch and the random seed is fixed by the epoch number, just like:

epoch_iter_factory = Task.build_epoch_iter_factory()
for epoch in range(max_epoch):
  iterator = epoch_iter_factory.build_iter(epoch)

Therefore, the training can be resumed at the start of the epoch.

Weights & Biases integration

About Weights & Biases: https://docs.wandb.com/

  1. Installation and setup

    See: https://docs.wandb.com/quickstart

    wandb login
    
  2. Enable wandb

    python -m espnet2.bin.asr_train --use_wandb true
    

    and go to the shown URL.

  3. [Option] To use HTTPS PROXY

    export HTTPS_PROXY=...your proxy
    export CURL_CA_BUNDLE=your.pem
    export CURL_CA_BUNDLE=   # Disable SSL certificate verification
    

Multi GPUs

python -m espnet2.bin.asr_train --ngpu 2

Just using CUDA_VISIBLE_DEVICES to specify the device number:

CUDA_VISIBLE_DEVICES=2,3 python -m espnet2.bin.asr_train --ngpu 2

About distributed training, see Distributed training.

The relation between mini-batch size and number of GPUs

The batch-size can be changed as follows:

# Change both of the batch_size for training and validation
python -m espnet2.bin.asr_train --batch_size 20
# Change the batch_size for validation
python -m espnet2.bin.asr_train --valid_batch_size 200

The behavior for batch-size during multi-GPU training is different from that of ESPNet1.

  • ESPNet1: The batch-size will be multiplied by the number of GPUs.

    python -m espnet.bin.asr_train --batch_size 10 --ngpu 2 # Actual batch_size is 20 and each GPU devices are assigned to 10
    
  • ESPnet2: The batch-size is not changed regardless of the number of GPUs.

    • Therefore, you should set a more number of batch-size than that of GPUs.

    python -m espnet.bin.asr_train --batch_size 10 --ngpu 2 # Actual batch_size is 10 and each GPU devices are assigned to 5
    

Change mini-batch type

We adopt variable mini-batch size with considering the dimension of the input features to make the best use of the GPU memory.

There are 5 types:

batch_type Option to change batch-size Variable batch-size Requirement
unsorted --batch_size No -
sorted --batch_size No Length information of features
folded --batch_size Yes Length information of features
length --batch_bins Yes Length information of features
numel --batch_bins Yes Shape information of features

Note that –batch_size is ignored if –batch_type=length or –batch_type=numel.

--batch_type unsorted

This mode has nothing special feature and just creates constant-size mini-batches without any sorting by the length order. If you intend to use ESPnet as not Seq2Seq task, this type may be suitable.

Unlike the other mode, this mode doesn’t require the information of the feature dimension. In other words, it’s not mandatory to prepare shape_file:

python -m espnet.bin.asr_train \
  --batch_size 10 --batch_type unsorted \
  --train_data_path_and_name_and_type "train.scp,feats,npy" \
  --valid_data_path_and_name_and_type "valid.scp,feats,npy" \
  --train_shape_file "train.scp" \
  --valid_shape_file "valid.scp"

This system might seem strange and you might also feel --*_shape_file is verbose because the training corpus can be described totally only using --*_data_path_and_name_and_type.

From the viewpoint of the implementation, we separate the data source for the Dataset and BatchSampler in the term of PyTorch and --*_data_path_and_name_and_type and --*_shape_file correspond to them respectively. From the viewpoint of the training strategy, because variable batch-size is supported according to the length/dimension of each feature, thus we need to prepare the shape information before training.

--batch_type sorted

This mode creates constant-size mini-batches with sorting by the length order. This mode requires the information of the length.

python -m espnet.bin.asr_train \
  --batch_size 10 --batch_type sorted \
  --train_data_path_and_name_and_type "train.scp,feats,npy" \
  --train_data_path_and_name_and_type "train2.scp,feats2,npy" \
  --valid_data_path_and_name_and_type "valid.scp,feats,npy" \
  --valid_data_path_and_name_and_type "valid2.scp,feats2,npy" \
  --train_shape_file "train_length.txt" \
  --valid_shape_file "valid_length.txt"

e.g. length.txt

sample_id1 1230
sample_id2 156
sample_id3 890
...

Where the fist column indicates the sample id and the second is the length of the corresponding feature. You can see that shape file is input instead in our recipes.

e.g. shape.txt

sample_id1 1230,80
sample_id2 156,80
sample_id3 890,80
...

This file describes the full information of the feature shape; The first number is the length of the sequence and the second or later are the dimension of feature: Length,Dim1,Dim2,....

Only the first number is referred for --batch_type sorted, --batch_type folded and --batch_type length, and the shape information is required only when --batch_type numel.

--batch_type folded

In ESPnet1, this mode is referred as seq.

This mode creates mini-batch which has the size of base_batch_size // max_i(1 + L_i // f_i). Where L_i is the maximum length in the mini-batch for ith feature and f_i is the --fold length corresponding to the feature. This mode requires the information of length.

python -m espnet.bin.asr_train \
  --batch_size 20 --batch_type folded \
  --train_data_path_and_name_and_type "train.scp,feats,npy" \
  --train_data_path_and_name_and_type "train2.scp,feats2,npy" \
  --valid_data_path_and_name_and_type "valid.scp,feats,npy" \
  --valid_data_path_and_name_and_type "valid2.scp,feats2,npy" \
  --train_shape_file "train_length.scp" \
  --train_shape_file "train_length2.scp" \
  --valid_shape_file "valid_length.scp" \
  --valid_shape_file "valid_length2.scp" \
  --fold_length 5000 \
  --fold_length 300

Note that the repeat number of *_shape_file must equal to the number of --fold_length, but you don’t need to input same number of shape files as the number of data file. i.e. You can give it as follows:

python -m espnet.bin.asr_train \
  --batch_size 20 --batch_type folded \
  --train_data_path_and_name_and_type "train.scp,feats,npy" \
  --train_data_path_and_name_and_type "train2.scp,feats2,npy" \
  --valid_data_path_and_name_and_type "valid.scp,feats,npy" \
  --valid_data_path_and_name_and_type "valid2.scp,feats2,npy" \
  --train_shape_file "train_length.txt" \
  --valid_shape_file "valid_length.txt" \
  --fold_length 5000

In this example, the length of the first feature is considered while the second can be ignored. This technique can be also applied for --batch_type length and --batch_type numel.

--batch_type length

In ESPnet1, this mode is referred as frame.

You need to specify --batch_bins to determine the mini-batch size instead of --batch_size. Each mini-batch has equal number of bins as possible counting by the total length in the mini-batch; i.e. bins = sum(len(feat) for feats in batch for feat in feats). This mode requires the information of length.

python -m espnet.bin.asr_train \
  --batch_bins 10000 --batch_type length \
  --train_data_path_and_name_and_type "train.scp,feats,npy" \
  --train_data_path_and_name_and_type "train2.scp,feats2,npy" \
  --valid_data_path_and_name_and_type "valid.scp,feats,npy" \
  --valid_data_path_and_name_and_type "valid2.scp,feats2,npy" \
  --train_shape_file "train_length.txt" \
  --train_shape_file "train_length2.txt" \
  --valid_shape_file "valid_length.txt" \
  --valid_shape_file "valid_length2.txt" \

--batch_type numel

In ESPnet1, this mode is referred as bins.

You need to specify --batch_bins to determine the mini-batch size instead of --batch_size. Each mini-batches has equal number of bins as possible counting by the total number of elements; i.e. bins = sum(numel(feat) for feats in batch for feat in feats), where numel returns the infinite product of the shape of each feature; shape[0] * shape[1] * ...

python -m espnet.bin.asr_train \
  --batch_bins 200000 --batch_type numel \
  --train_data_path_and_name_and_type "train.scp,feats,npy" \
  --train_data_path_and_name_and_type "train2.scp,feats2,npy" \
  --valid_data_path_and_name_and_type  "valid.scp,feats,npy" \
  --valid_data_path_and_name_and_type  "valid2.scp,feats2,npy" \
  --train_shape_file "train_shape.txt" \
  --train_shape_file "train_shape2.txt" \
  --valid_shape_file "valid_shape.txt" \
  --valid_shape_file "valid_shape2.txt"

Gradient accumulating

There are several ways to deal with larger model architectures than the capacity of your GPU device memory during training.

  • Using a larger number of GPUs

  • Using a half decision tensor

  • Using torch.utils.checkpoint

  • Gradient accumulating

Gradient accumulating is a technique to handle larger mini-batch than available size.

Split a mini-batch into several numbers and forward and backward for each piece and accumulate the gradients ony by one, while optimizer’s updating is invoked every the number of forwarding just like following:

# accum_grad is the number of pieces
for i, batch in enumerate(iterator):
    loss = net(batch)
    (loss / accum_grad).backward() # Gradients are accumulated
    if i % accum_grad:
        optim.update()
        optim.zero_grads()

Give --accum_grad <int> to use this option.

python -m espnet.bin.asr_train --accum_grad 2

The effective batch_size becomes almost same as accum_grad * batch_size except for:

  • The random state

  • Some statistical layers based on mini-batch e.g. BatchNormalization

  • The case that the batch_size is not unified for each iteration.

Automatic Mixed Precision training

python -m espnet.bin.asr_train --use_amp true

Reproducibility and determinization

There are some possibilities to make training non-reproducible.

  • Initialization of parameters that come from PyTorch/ESPnet version difference.

  • Reducing order for float values during multi GPUs training.

    • I don’t know whether NCCL is deterministic or not.

  • Random seed difference

    • We fixed the random seed for each epoch.

  • CuDNN or some non-deterministic operations for CUDA: See https://pytorch.org/docs/stable/notes/randomness.html

By default, CuDNN performs deterministic mode in our training and it can be turned off by:

python -m espnet.bin.asr_train --cudnn_deterministic false