Update results.

This commit is contained in:
Fangjun Kuang 2022-03-02 11:33:11 +08:00
parent c80dae7b07
commit 131f087dcb
3 changed files with 84 additions and 8 deletions

View File

@ -1,4 +1,6 @@
# Introduction
Please refer to <https://icefall.readthedocs.io/en/latest/recipes/aishell.html>
for how to run models in this recipe.
@ -7,11 +9,11 @@ for how to run models in this recipe.
There are various folders containing the name `transducer` in this folder.
The following table lists the differences among them.
| | Encoder | Decoder | Comment |
|------------------------|-----------|--------------------|-----------------------------------------------------------------------------------|
| `transducer_stateless` | Conformer | Embedding + Conv1d | with `k2.rnnt_loss` |
| `transducer_stateless_modified` | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` |
| `transducer_stateless_modified-2` | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` + extra data |
| | Encoder | Decoder | Comment |
|------------------------------------|-----------|--------------------|-----------------------------------------------------------------------------------|
| `transducer_stateless` | Conformer | Embedding + Conv1d | with `k2.rnnt_loss` |
| `transducer_stateless_modified` | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` |
| `transducer_stateless_modified-2` | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` + extra data |
The decoder in `transducer_stateless` is modified from the paper
[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).

View File

@ -1,12 +1,86 @@
## Results
### Aishell training result(Transducer-stateless)
#### 2022-03-01
[./transducer_stateless_modified-2](./transducer_stateless_modified-2)
Stateless transducer + modified transducer + using [aidatatang_200zh](http://www.openslr.org/62/) as extra training data.
| | test |comment |
|------------------------|------|----------------------------------------------------------------|
| greedy search | 4.94 |--epoch 89, --avg 38, --max-duration 100, --max-sym-per-frame 1 |
| modified beam search | 4.68 |--epoch 89, --avg 38, --max-duration 100 --beam-size 4 |
The training commands are:
```bash
cd egs/aishell/ASR
./prepare.sh --stop-stage 6
./prepare_aidatatang_200zh.sh
export CUDA_VISIBLE_DEVICES="0,1,2"
./transducer_stateless_modified-2/train.py \
--world-size 3 \
--num-epochs 90 \
--start-epoch 0 \
--exp-dir transducer_stateless_modified-2/exp-2 \
--max-duration 250 \
--lr-factor 2.0 \
--context-size 2 \
--modified-transducer-prob 0.25 \
--datatang-prob 0.2
```
The tensorboard log is available at
<https://tensorboard.dev/experiment/oG72ZlWaSGua6fXkcGRRjA/>
The commands for decoding are
```bash
# greedy search
for epoch in 89; do
for avg in 38; do
./transducer_stateless_modified-2/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir transducer_stateless_modified-2/exp-2 \
--max-duration 100 \
--context-size 2 \
--decoding-method greedy_search \
--max-sym-per-frame 1
done
done
# modified beam search
for epoch in 89; do
for avg in 38; do
./transducer_stateless_modified-2/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir transducer_stateless_modified-2/exp-2 \
--max-duration 100 \
--context-size 2 \
--decoding-method modified_beam_search \
--beam-size 4
done
done
```
You can find a pre-trained model and decoding logs and results at
<https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2-2022-03-01>
#### 2022-2-19
(Duo Ma): The tensorboard log for training is available at https://tensorboard.dev/experiment/25PmX3MxSVGTdvIdhOwllw/#scalars
You can find a pretrained model by visiting https://huggingface.co/shuanguanma/icefall_aishell_transducer_stateless_context_size2_epoch60_2022_2_19
| | test |comment |
|---------------------------|------|-----------------------------------------|
| greedy search | 5.4 |--epoch 59, --avg 10, --max-duration 100 |
| beam search | 5.05|--epoch 59, --avg 10, --max-duration 100 |
| greedy search | 5.4 |--epoch 59, --avg 10, --max-duration 100 |
| beam search | 5.05|--epoch 59, --avg 10, --max-duration 100 |
You can use the following commands to reproduce our results:

View File

@ -95,7 +95,7 @@ def get_parser():
parser.add_argument(
"--exp-dir",
type=str,
default="transducer_stateless_modified/exp",
default="transducer_stateless_modified-2/exp",
help="The experiment dir",
)