FunASR/docs/get_started.md

# Get Started
This is an easy example which introduces how to train a paraformer model on AISHELL-1 data from scratch. According to this example, you can train other models (conformer, paraformer, etc.) on other datasets (AISHELL-1, AISHELL-2, etc.) similarly.

## Overall Introduction
We provide a recipe `egs/aishell/paraformer/run.sh` for training a paraformer model on AISHELL-1 data  . This recipe consists of five stages and support training on multiple GPUs and decoding by CPU or GPU. Before introduce each stage in detail, we first explain several variables which should be set by users.
- `CUDA_VISIBLE_DEVICES`: visible gpu list
- `gpu_num`: the number of GPUs used for training
- `gpu_inference`: whether to use GPUs for decoding
- `njob`: for CPU decoding, indicating the total number of CPU jobs; for GPU decoding, indicating the number of jobs on each GPU.
- `feats_dir`: the path to save processed data
- `exp_dir`: the path to save experimental results
- `data_aishell`: the path of raw AISHELL-1 data
- `tag`: the suffix of experimental result directory
- `nj`: the number of jobs for data preparation
- `speed_perturb`: the range of speech perturbed

## Stage 0: Data preparation
This stage processes raw AISHELL-1 data `$data_aishell` and generates the corresponding `wav.scp` and `text` in `$feats_dir/data/xxx` and `xxx` means `train/dev/test`. Here we assume you have already downloaded AISHELL-1 data. If not, you can download data [here](https://www.openslr.org/33/) and set the path for `$data_aishell`. Here we show examples for `wav.scp` and `text`, separately.
* `wav.scp`
```
BAC009S0002W0122 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0122.wav
BAC009S0002W0123 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0123.wav
BAC009S0002W0124 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0124.wav
...
```
* `text`
```
BAC009S0002W0122 而 对 楼 市 成 交 抑 制 作 用 最 大 的 限 购
BAC009S0002W0123 也 成 为 地 方 政 府 的 眼 中 钉
BAC009S0002W0124 自 六 月 底 呼 和 浩 特 市 率 先 宣 布 取 消 限 购 后
...
```
We can see that these two files both have two columns while the first column is the wav-id and the second column is the corresponding wav-path/label tokens.

## Stage 1: Feature Generation
This stage extracts FBank feature from raw wav `wav.scp` and apply speed perturbation as data augmentation according to `speed_perturb`. You can set `nj` to control the number of jobs for feature generation. The output features are saved in `$feats_dir/dump/xxx/ark` and the corresponding `feats.scp` files are saved as `$feats_dir/dump/xxx/feats.scp`. An example of `feats.scp` can be seen as follows:
* `feats.scp`
```
...
BAC009S0002W0122_sp0.9 /nfs/haoneng.lhn/funasr_data/aishell-1/dump/fbank/train/ark/feats.16.ark:592751055
...
```
Note that samples in this file have already been shuffled. This file contains two columns. The first column is the wav-id while the second column is the kaldi-ark feature path. Besides, `speech_shape` and `text_shape` are also generated in this stage, denoting the speech feature shape and text length of each sample. The examples are shown as follows:
* `speech_shape`
```
...
BAC009S0002W0122_sp0.9 665,80
...
```
* `text_shape`
```
...
BAC009S0002W0122_sp0.9 15
...
```
These two files have two columns. The first column is the wav-id and the second column is the corresponding speech feature shape and text length.

## Stage 2: Dictionary Preparation
This stage prepares a dictionary, which is used as a mapping between label characters and integer indices during ASR training. The output dictionary file is saved as `$feats_dir/data/$lang_toekn_list/$token_type/tokens.txt`. Here we show an example of `tokens.txt` as follows:
* `tokens.txt`
```
<blank>
<s>
</s>
一
丁
...
龚
龟
<unk>
```
* `<blank>`: indicates the blank token for CTC
* `<s>`: indicates the start-of-sentence token
* `</s>`: indicates the end-of-sentence token
* `<unk>`: indicates the out-of-vocabulary token

## Stage 3: Training
This stage achieves the training of the specified model. To start training, you should manually set `exp_dir`, `CUDA_VISIBLE_DEVICES` and `gpu_num`, which have already been explained above. By default, the best `$keep_nbest_models` checkpoints on validation dataset will be averaged to generate a better model and adopted for decoding.

* DDP Training

We support the DistributedDataParallel (DDP) training and the detail can be found [here](https://pytorch.org/tutorials/intermediate/ddp_tutorial.html). To enable DDP training, please set `gpu_num` greater than 1. For example, if you set `CUDA_VISIBLE_DEVICES=0,1,5,6,7` and `gpu_num=3`, then the gpus with ids 0, 1 and 5 will be used for training.

* DataLoader

[comment]: <> (We support two types of DataLoaders for small and large datasets, respectively. By default, the small DataLoader is used and you can set `dataset_type=large` to enable large DataLoader. For small DataLoader, )
We support an optional iterable-style DataLoader based on [Pytorch Iterable-style DataPipes](https://pytorch.org/data/beta/torchdata.datapipes.iter.html) for large dataset and you can set `dataset_type=large` to enable it. 

* Configuration

The parameters of the training, including model, optimization, dataset, etc., are specified by a YAML file in `conf` directory. Also, you can directly specify the parameters in `run.sh` recipe. Please avoid to specify the same parameters in both the YAML file and the recipe.

* Training Steps

We support two parameters to specify the training steps, namely `max_epoch` and `max_update`. `max_epoch` indicates the total training epochs while `max_update` indicates the total training steps. If these two parameters are specified at the same time, once the training reaches any one of the two parameters, the training will be stopped.

* Tensorboard

You can use tensorboard to observe the loss, learning rate, etc. Please run the following command:
```
tensorboard --logdir ${exp_dir}/exp/${model_dir}/tensorboard/train
```

## Stage 4: Decoding
This stage generates the recognition results with acoustic features as input and calculate the `CER` to verify the performance of the trained model. 

* Mode Selection

As we support conformer, paraformer and uniasr in FunASR and they have different inference interfaces, a `mode` param is specified as `asr/paraformer/uniase` according to the trained model.

* Configuration

We support CTC decoding, attention decoding and hybrid CTC-attention decoding in FunASR, which can be specified by `ctc_weight` in a YAML file in `conf` directory. Specifically, `ctc_weight=1.0` indicates attention decoding, `ctc_weight=0.0` indicates CTC decoding, `0.0<ctc_weight<1.0` indicates hybrid CTC-attention decoding.

* CPU/GPU Decoding

We support CPU and GPU decoding in FunASR. For CPU decoding, you should set `gpu_inference=False` and set `njob` to specify the total number of CPU decoding jobs. For GPU decoding, you should set `gpu_inference=True`. You should also set `gpuid_list` to indicate which GPUs are used for decoding and `njobs` to indicate the number of decoding jobs on each GPU.

* Performance

We adopt `CER` to verify the performance. The results are in `$exp_dir/exp/$model_dir/$decoding_yaml_name/$average_model_name/$dset`, namely `text.cer` and `text.cer.txt`. `text.cer` saves the comparison between the recognized text and the reference text while `text.cer.txt` saves the final `CER` result. The following is an example of `text.cer`:
* `text.cer`
```
...
BAC009S0764W0213(nwords=11,cor=11,ins=0,del=0,sub=0) corr=100.00%,cer=0.00%
ref:    构 建 良 好 的 旅 游 市 场 环 境
res:    构 建 良 好 的 旅 游 市 场 环 境
...
```