mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-09-07 16:14:17 +00:00
this commit should finalize the PR (hopefully)
This commit is contained in:
parent
9bc287ca03
commit
c65f80a11f
@ -7,6 +7,8 @@ set -eou pipefail
|
||||
|
||||
stage=-1
|
||||
stop_stage=100
|
||||
perturb_speed=true
|
||||
|
||||
|
||||
# We assume dl_dir (download dir) contains the following
|
||||
# directories and files. If not, they will be downloaded
|
||||
@ -77,7 +79,7 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
|
||||
log "Stage 4: Compute fbank for aidatatang_200zh"
|
||||
if [ ! -f data/fbank/.aidatatang_200zh.done ]; then
|
||||
mkdir -p data/fbank
|
||||
./local/compute_fbank_aidatatang_200zh.py --perturb-speed True
|
||||
./local/compute_fbank_aidatatang_200zh.py --perturb-speed ${perturb_speed}
|
||||
touch data/fbank/.aidatatang_200zh.done
|
||||
fi
|
||||
fi
|
||||
|
@ -7,6 +7,54 @@
|
||||
[./zipformer](./zipformer)
|
||||
|
||||
It's reworked Zipformer with Pruned RNNT loss, note that results below are produced by model trained on data without speed perturbation applied.
|
||||
**Caution**: It uses `--context-size=1`.
|
||||
|
||||
##### normal-scaled model, number of model parameters: 73412551, i.e., 73.41 M
|
||||
|
||||
| | test | dev | comment |
|
||||
|------------------------|------|------|-----------------------------------------|
|
||||
| greedy search | 4.73 | 4.54 | --epoch 38 --avg 14 |
|
||||
| modified beam search | 4.49 | 4.27 | --epoch 40 --avg 12 |
|
||||
| fast beam search | 4.65 | 4.4 | --epoch 40 --avg 12 |
|
||||
|
||||
Command for training is:
|
||||
```bash
|
||||
./prepare.sh
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="0,1"
|
||||
|
||||
./zipformer/train.py \
|
||||
--world-size 2 \
|
||||
--num-epochs 40 \
|
||||
--start-epoch 1 \
|
||||
--use-fp16 1 \
|
||||
--context-size 1 \
|
||||
--enable-musan 0 \
|
||||
--exp-dir zipformer/exp \
|
||||
--max-duration 1000
|
||||
```
|
||||
|
||||
Command for decoding is:
|
||||
```bash
|
||||
./zipformer/decode.py \
|
||||
--epoch 38 \
|
||||
--avg 14 \
|
||||
--exp-dir ./zipformer/exp \
|
||||
--lang-dir data/lang_char \
|
||||
--context-size 1 \
|
||||
--decoding-method greedy_search
|
||||
|
||||
for m in modified_beam_search fast_beam_search ; do
|
||||
./zipformer/decode.py \
|
||||
--epoch 40 \
|
||||
--avg 12 \
|
||||
--exp-dir ./zipformer/exp \
|
||||
--lang-dir data/lang_char \
|
||||
--context-size 1 \
|
||||
--decoding-method $m
|
||||
done
|
||||
```
|
||||
|
||||
|
||||
**⚠️ 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.**
|
||||
|
||||
@ -20,7 +68,7 @@ It's reworked Zipformer with Pruned RNNT loss, note that results below are produ
|
||||
|
||||
Command for training is:
|
||||
```bash
|
||||
./prepare.sh # after setting --perturb-speed to False in the prepare.sh
|
||||
./prepare.sh --perturb-speed false
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="0,1"
|
||||
|
||||
@ -59,7 +107,7 @@ done
|
||||
|
||||
Command for training is:
|
||||
```bash
|
||||
./prepare.sh # after setting --perturb-speed to False in the prepare.sh
|
||||
./prepare.sh --perturb-speed false
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="0,1"
|
||||
|
||||
|
@ -8,6 +8,7 @@ set -eou pipefail
|
||||
nj=15
|
||||
stage=-1
|
||||
stop_stage=11
|
||||
perturb_speed=true
|
||||
|
||||
# We assume dl_dir (download dir) contains the following
|
||||
# directories and files. If not, they will be downloaded
|
||||
@ -114,7 +115,7 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
|
||||
log "Stage 3: Compute fbank for aishell"
|
||||
if [ ! -f data/fbank/.aishell.done ]; then
|
||||
mkdir -p data/fbank
|
||||
./local/compute_fbank_aishell.py --perturb-speed True
|
||||
./local/compute_fbank_aishell.py --perturb-speed ${perturb_speed}
|
||||
touch data/fbank/.aishell.done
|
||||
fi
|
||||
fi
|
||||
|
@ -21,7 +21,7 @@ It's reworked Zipformer with Pruned RNNT loss, note that results below are produ
|
||||
The training command for reproducing is given below:
|
||||
|
||||
```bash
|
||||
./prepare.sh # after setting --perturb-speed to False in the prepare.sh
|
||||
./prepare.sh --perturb-speed false
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="0,1"
|
||||
|
||||
|
@ -8,6 +8,8 @@ set -eou pipefail
|
||||
nj=30
|
||||
stage=0
|
||||
stop_stage=7
|
||||
perturb_speed=true
|
||||
|
||||
|
||||
# We assume dl_dir (download dir) contains the following
|
||||
# directories and files. If not, you need to apply aishell2 through
|
||||
@ -101,7 +103,7 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
|
||||
log "Stage 3: Compute fbank for aishell2"
|
||||
if [ ! -f data/fbank/.aishell2.done ]; then
|
||||
mkdir -p data/fbank
|
||||
./local/compute_fbank_aishell2.py --perturb-speed True
|
||||
./local/compute_fbank_aishell2.py --perturb-speed ${perturb_speed}
|
||||
touch data/fbank/.aishell2.done
|
||||
fi
|
||||
fi
|
||||
|
@ -20,7 +20,7 @@ It's reworked Zipformer with Pruned RNNT loss, note that results below are produ
|
||||
|
||||
Command for training is:
|
||||
```bash
|
||||
./prepare.sh # after setting --perturb-speed to False in the prepare.sh
|
||||
./prepare.sh --perturb-speed false
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="0,1"
|
||||
|
||||
|
@ -7,6 +7,8 @@ set -eou pipefail
|
||||
|
||||
stage=-1
|
||||
stop_stage=100
|
||||
perturb_speed=true
|
||||
|
||||
|
||||
# We assume dl_dir (download dir) contains the following
|
||||
# directories and files. If not, they will be downloaded
|
||||
@ -107,7 +109,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
|
||||
log "Stage 5: Compute fbank for aishell4"
|
||||
if [ ! -f data/fbank/.aishell4.done ]; then
|
||||
mkdir -p data/fbank
|
||||
./local/compute_fbank_aishell4.py --perturb-speed True
|
||||
./local/compute_fbank_aishell4.py --perturb-speed ${perturb_speed}
|
||||
touch data/fbank/.aishell4.done
|
||||
fi
|
||||
fi
|
||||
|
Loading…
x
Reference in New Issue
Block a user