Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[examples] update paraformer results on aishell #2324

Merged
merged 2 commits into from
Jan 25, 2024
Merged

[examples] update paraformer results on aishell #2324

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
18f21b2
[examples] update paraformer results on aishell
xingchensong Jan 25, 2024
aaee132
[examples] update paraformer results on aishell
xingchensong Jan 25, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 18 additions & 13 deletions examples/aishell/paraformer/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,36 +5,41 @@ output_dir=exp/paraformer/large
mkdir -p ${output_dir}
. ./path.sh && python wenet/paraformer/convert_paraformer_to_wenet_config_and_ckpt.py \
--output_dir ${output_dir}
# init ctc and embed(used in sampler)
python local/modify_ckpt.py \
--add_list "{\"ctc.ctc_lo.weight\": \"decoder.embed.0.weight\", \"embed.0.weight\": \"decoder.embed.0.weight\"}" \
--input_ckpt exp/paraformer/large/wenet_paraformer.pt \
--output_ckpt exp/paraformer/large/wenet_paraformer.init-ctc.init-embed.pt
```

# Performance Record

## Paraformer (original) Result

| decoding mode | CER |
|---------------------------|-------|
| paraformer greedy search | 1.95 |
| decoding mode | full | 16 |
|---------------------------|-------|-----|
| paraformer greedy search | 1.95 | N/A |

## Paraformer (full-parameter tuning) Result

* Training info: batch size 28, ctc_weight: 0.3, acc_grad 4, 8 * v100 gpu, 40 epochs
* Training info: torch_ddp fp32, batch size 28, ctc_weight: 0.3, acc_grad 1, 8 * 3090 gpu, 60 epochs (about 8h)
* Decoding info: ctc_weight 0.3, average_num 5
* Git hash: TBD

| decoding mode | CER |
|---------------------------|-------|
| ctc greedy search | 4.00 |
| ctc prefix beam search | 4.00 |
| paraformer greedy search | 2.16 |
| decoding mode | full | 16 |
|---------------------------|-------|-----|
| ctc greedy search | 3.45 % N=104765 C=101244 S=3406 D=115 I=91 | N/A |
| ctc prefix beam search | 3.44 % N=104765 C=101247 S=3407 D=111 I=83 | N/A |
| paraformer greedy search | 2.19 % N=104765 C=102643 S=1959 D=163 I=172 | N/A |

## Paraformer-dynamic training (full-parameter tuning) Result

* Training info: batch size 28, ctc_weight: 0.3, acc_grad 4, 8 * v100 gpu, 43 epochs
* Training info: torch_ddp fp32, batch size 28, ctc_weight: 0.3, acc_grad 1, 8 * 3090 gpu, 60 epochs (about 8h)
* Decoding info: ctc_weight 0.3, average_num 5
* Git hash: TBD

| decoding mode | full | 16 |
|---------------------------|--------|------|
| ctc greedy search | 3.93 | 4.94 |
| ctc prefix beam search | 3.93 | 4.94 |
| paraformer greedy search | 2.08 | 2.41 |
| ctc greedy search | 3.46 % N=104765 C=101235 S=3409 D=121 I=98 | 4.18 % N=104765 C=100495 S=4149 D=121 I=107 |
| ctc prefix beam search | 3.45 % N=104765 C=101239 S=3413 D=113 I=91 | 4.17 % N=104765 C=100500 S=4150 D=115 I=103 |
| paraformer greedy search | 2.15 % N=104765 C=102640 S=1977 D=148 I=132 | 2.40 % N=104765 C=102409 S=2220 D=136 I=161 |
18 changes: 9 additions & 9 deletions examples/aishell/paraformer/conf/train_paraformer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,32 +2,32 @@ input_dim: 560

encoder: sanm_encoder
encoder_conf:
attention_dropout_rate: 0.1
attention_dropout_rate: 0.0
attention_heads: 4
dropout_rate: 0.1
dropout_rate: 0.0
input_layer: paraformer_dummy
kernel_size: 11
linear_units: 2048
normalize_before: true
num_blocks: 50
output_size: 512
pos_enc_layer_type: abs_pos_paraformer
positional_dropout_rate: 0.1
positional_dropout_rate: 0.0
sanm_shfit: 0
gradient_checkpointing: true

decoder: sanm_decoder
decoder_conf:
att_layer_num: 16
attention_heads: 4
dropout_rate: 0.1
dropout_rate: 0.0
kernel_size: 11
linear_units: 2048
num_blocks: 16
positional_dropout_rate: 0.1
positional_dropout_rate: 0.0
sanm_shfit: 0
self_attention_dropout_rate: 0.1
src_attention_dropout_rate: 0.1
self_attention_dropout_rate: 0.0
src_attention_dropout_rate: 0.0
gradient_checkpointing: true

tokenizer: paraformer
Expand Down Expand Up @@ -102,12 +102,12 @@ dataset_conf:

grad_clip: 5
accum_grad: 1
max_epoch: 45
max_epoch: 60
log_interval: 100

optim: adam
optim_conf:
lr: 0.0005
scheduler: warmuplr
scheduler_conf:
warmup_steps: 25000
warmup_steps: 12000
18 changes: 9 additions & 9 deletions examples/aishell/paraformer/conf/train_paraformer_dynamic.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,17 +2,17 @@ input_dim: 560

encoder: sanm_encoder
encoder_conf:
attention_dropout_rate: 0.1
attention_dropout_rate: 0.0
attention_heads: 4
dropout_rate: 0.1
dropout_rate: 0.0
input_layer: paraformer_dummy
kernel_size: 11
linear_units: 2048
normalize_before: true
num_blocks: 50
output_size: 512
pos_enc_layer_type: abs_pos_paraformer
positional_dropout_rate: 0.1
positional_dropout_rate: 0.0
sanm_shfit: 0
use_dynamic_chunk: true
gradient_checkpointing: true
Expand All @@ -21,14 +21,14 @@ decoder: sanm_decoder
decoder_conf:
att_layer_num: 16
attention_heads: 4
dropout_rate: 0.1
dropout_rate: 0.0
kernel_size: 11
linear_units: 2048
num_blocks: 16
positional_dropout_rate: 0.1
positional_dropout_rate: 0.0
sanm_shfit: 0
self_attention_dropout_rate: 0.1
src_attention_dropout_rate: 0.1
self_attention_dropout_rate: 0.0
src_attention_dropout_rate: 0.0
gradient_checkpointing: true

tokenizer: paraformer
Expand Down Expand Up @@ -103,12 +103,12 @@ dataset_conf:

grad_clip: 5
accum_grad: 1
max_epoch: 45
max_epoch: 60
log_interval: 100

optim: adam
optim_conf:
lr: 0.0005
scheduler: warmuplr
scheduler_conf:
warmup_steps: 25000
warmup_steps: 12000
31 changes: 19 additions & 12 deletions examples/aishell/paraformer/run.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,13 +28,13 @@ job_id=2024
# data_type can be `raw` or `shard`. Typically, raw is used for small dataset,
# `shard` is used for large dataset which is over 1k hours, and `shard` is
# faster on reading data and training.
data_type=raw
data_type=shard

train_set=train

train_config=conf/train_paraformer.yaml
checkpoint=exp/paraformer/large/wenet_paraformer.pt
dir=exp/finetune_paraformer
train_config=conf/train_paraformer_dynamic.yaml
checkpoint=exp/paraformer/large/wenet_paraformer.init-ctc.init-embed.pt
dir=exp/finetune_paraformer_dynamic
tensorboard_dir=tensorboard
num_workers=8
prefetch=500
Expand All @@ -44,6 +44,12 @@ average_checkpoint=true
decode_checkpoint=$dir/final.pt
average_num=5
decode_modes="ctc_greedy_search ctc_prefix_beam_search paraformer_greedy_search"
decode_device=0
decoding_chunk_size=-1
decode_batch=16
ctc_weight=0.3
reverse_weight=0.5
max_epoch=100

train_engine=torch_ddp

Expand Down Expand Up @@ -124,36 +130,37 @@ fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# Test model, please specify the model you want to test by --checkpoint
if [ ${average_checkpoint} == true ]; then
decode_checkpoint=$dir/avg_${average_num}.pt
decode_checkpoint=$dir/avg_${average_num}_maxepoch_${max_epoch}.pt
echo "do model average and final checkpoint is $decode_checkpoint"
python wenet/bin/average_model.py \
--dst_model $decode_checkpoint \
--src_path $dir \
--num ${average_num} \
--max_epoch ${max_epoch} \
--val_best
fi
# Please specify decoding_chunk_size for unified streaming and
# non-streaming model. The default value is -1, which is full chunk
# for non-streaming inference.
decoding_chunk_size=
ctc_weight=0.3
reverse_weight=0.5
python wenet/bin/recognize.py --gpu 0 \
base=$(basename $decode_checkpoint)
result_dir=$dir/${base}_chunk${decoding_chunk_size}_ctc${ctc_weight}_reverse${reverse_weight}
mkdir -p ${result_dir}
python wenet/bin/recognize.py --gpu ${decode_device} \
--modes $decode_modes \
--config $dir/train.yaml \
--data_type $data_type \
--test_data data/test/data.list \
--checkpoint $decode_checkpoint \
--beam_size 10 \
--batch_size 16 \
--batch_size ${decode_batch} \
--penalty 0.0 \
--ctc_weight $ctc_weight \
--reverse_weight $reverse_weight \
--result_dir $dir \
--result_dir $result_dir \
${decoding_chunk_size:+--decoding_chunk_size $decoding_chunk_size}
for mode in ${decode_modes}; do
python tools/compute-wer.py --char=1 --v=1 \
data/test/text $dir/$mode/text > $dir/$mode/wer
data/test/text $result_dir/$mode/text > $result_dir/$mode/wer
done
fi

Expand Down
Loading