Update RESULTS.md

This commit is contained in:
zr_jin 2023-08-17 16:52:47 +08:00
parent ce380a5fb3
commit 9bc287ca03

View File

@ -2,7 +2,7 @@
### Aishell training result (Stateless Transducer) ### Aishell training result (Stateless Transducer)
#### Zipformer #### Zipformer (Non-streaming)
[./zipformer](./zipformer) [./zipformer](./zipformer)
@ -10,6 +10,8 @@ It's reworked Zipformer with Pruned RNNT loss, note that results below are produ
**⚠️ If you prefer to have the speed perturbation disabled, please manually set `--perturb-speed` to `False` for `./local/compute_fbank_aishell.py` in the `prepare.sh` script.** **⚠️ If you prefer to have the speed perturbation disabled, please manually set `--perturb-speed` to `False` for `./local/compute_fbank_aishell.py` in the `prepare.sh` script.**
##### normal-scaled model, number of model parameters: 73412551, i.e., 73.41 M
| | test | dev | comment | | | test | dev | comment |
|------------------------|------|------|-----------------------------------------| |------------------------|------|------|-----------------------------------------|
| greedy search | 4.92 | 4.61 | --epoch 90 --avg 40 --max-duration 1200 | | greedy search | 4.92 | 4.61 | --epoch 90 --avg 40 --max-duration 1200 |
@ -47,6 +49,53 @@ for m in greedy_search modified_beam_search fast_beam_search ; do
done done
``` ```
##### small-scaled model, number of model parameters: 30167139, i.e., 30.17 M
| | test | dev | comment |
|------------------------|------|------|-----------------------------------------|
| greedy search | 5.15 | 4.93 | --epoch 90 --avg 40 --max-duration 1200 |
| modified beam search | 4.90 | 4.68 | --epoch 90 --avg 40 --max-duration 1200 |
| fast beam search | 5.08 | 4.85 | --epoch 90 --avg 40 --max-duration 1200 |
Command for training is:
```bash
./prepare.sh # after setting --perturb-speed to False in the prepare.sh
export CUDA_VISIBLE_DEVICES="0,1"
./zipformer/train.py \
--world-size 2 \
--num-epochs 100 \
--start-epoch 1 \
--use-fp16 1 \
--context-size 1 \
--exp-dir zipformer/exp-small \
--max-duration 1200 \
--lr-epochs 18 \
--num-encoder-layers 2,2,2,2,2,2 \
--feedforward-dim 512,768,768,768,768,768 \
--encoder-dim 192,256,256,256,256,256 \
--encoder-unmasked-dim 192,192,192,192,192,192
```
Command for decoding is:
```bash
for m in greedy_search modified_beam_search fast_beam_search ; do
./zipformer/decode.py \
--epoch 90 \
--avg 40 \
--exp-dir ./zipformer/exp-small \
--lang-dir data/lang_char \
--context-size 1 \
--max-duration 1200 \
--decoding-method $m \
--num-encoder-layers 2,2,2,2,2,2 \
--feedforward-dim 512,768,768,768,768,768 \
--encoder-dim 192,256,256,256,256,256 \
--encoder-unmasked-dim 192,192,192,192,192,192
done
```
#### Pruned transducer stateless 7 #### Pruned transducer stateless 7
[./pruned_transducer_stateless7](./pruned_transducer_stateless7) [./pruned_transducer_stateless7](./pruned_transducer_stateless7)