Skip to content

Latest commit

 

History

History
72 lines (66 loc) · 2.73 KB

overview.md

File metadata and controls

72 lines (66 loc) · 2.73 KB
Raw

Usage overview

There is no setup.py script for installing the APS package in the repository and actually I don't suggest readers doing that. The following shows the recommended way to use the APS.

A typical working directory looks like:

cmd conf data exp scripts utils

It tracks several experiments and could be initialized using scripts/init_workspace.sh, e.g.,

export APS_ROOT=/path/to/aps
$APS_ROOT/scripts/init_workspace.sh wsj0_2mix
$APS_ROOT/scripts/init_workspace.sh aishell_v1

will make directory current workspace like (APS_ROOT=../aps):

.
├── cmd -> ../aps/cmd
├── conf
│   ├── wsj0_2mix
│   └── aishell_v1
├── data
│   ├── wsj0_2mix
│   └── aishell_v1
├── scripts -> ../aps/scripts
└── utils -> ../aps/utils

Assuming that we've prepared data and experiment configurations under directory data and conf (see Instruction for details), the model training can be easily kicked off by running the provided scripts under scripts:

  • scripts/train.sh: Single-GPU training for acoustic model (AM), language model (LM) and speech separation/enhancement (SSE) model, respectively.
  • scripts/distributed_train.sh: Distributed training (currently single-node multi-GPU) for AM & SSE models.

E.g., running

./scripts/train_am.sh --batch-size 32 --gpu 0 aishell_v1 1a

will load .yaml configuration from conf/aishell_v1/1a.yaml and create checkpoint directory in exp/aishell_v1/1a (The feature extraction is performed on GPU so there is no need to prepare features for all the tasks). After one epoch is done, the directory looks like:

.
├── cmd -> ../aps/cmd
├── conf
│   ├── aishell_v1
│   │   └── 1a.yaml
│   └── wsj0_2mix
├── data
│   ├── aishell_v1
│   │   ├── dev
│   │   │   ├── text
│   │   │   ├── utt2dur
│   │   │   └── wav.scp
│   │   ├── train
│   │   │   ├── text
│   │   │   ├── utt2dur
│   │   │   └── wav.scp
│   │   └── tst
│   │       ├── text
│   │       └── wav.scp
│   └── wsj0_2mix
├── exp
│   └── aishell_v1
│       └── 1a
│           ├── best.pt.tar
│           ├── last.pt.tar
│           ├── train.yaml
│           └── trainer.log
├── scripts -> ../aps/scripts
└── utils -> ../aps/utils

(the data directory data/asr1/{train,dev,tst} is a typical setup for AM training)

After the end of the model training, we can start task dependent evaluation using the assets under checkpoint directory. Some recipes are available under aps/examples.