update eend_ola

嘉渊

2023-07-06 c83c406b72623deb973d391635475c5dfd9f8b93

update eend_ola

 egs/callhome/eend_ola/conf/train_diar_eend_ola_callhome_chunk2000.yaml

New file
@@ -0,0 +1,45 @@
# network architecture
# encoder related
encoder: eend_ola_transformer
encoder_conf:
    idim: 345
    n_layers: 4
    n_units: 256

# encoder-decoder attractor related
encoder_decoder_attractor: eda
encoder_decoder_attractor_conf:
    n_units: 256

# model related
model: eend_ola_similar_eend
model_conf:
    attractor_loss_weight:  0.01
    max_n_speaker: 8

# optimization related
accum_grad: 1
grad_clip: 5
max_epoch: 100
val_scheduler_criterion:
    - valid
    - loss
best_model_criterion:
-   - valid
    - loss
    - min
keep_nbest_models: 100

optim: adam
optim_conf:
    lr: 0.00001

dataset_conf:
    data_names: speech_speaker_labels
    data_types: kaldi_ark
    batch_conf:
        batch_type: unsorted
        batch_size: 8
    num_workers: 8

log_interval: 50

 egs/callhome/eend_ola/conf/train_diar_eend_ola_simu_2spkr.yaml

New file
@@ -0,0 +1,52 @@
# network architecture
# encoder related
encoder: eend_ola_transformer
encoder_conf:
    idim: 345
    n_layers: 4
    n_units: 256

# encoder-decoder attractor related
encoder_decoder_attractor: eda
encoder_decoder_attractor_conf:
    n_units: 256

# model related
model: eend_ola_similar_eend
model_conf:
    max_n_speaker: 8

# optimization related
accum_grad: 1
grad_clip: 5
max_epoch: 100
val_scheduler_criterion:
    - valid
    - loss
best_model_criterion:
-   - valid
    - loss
    - min
keep_nbest_models: 100

optim: adam
optim_conf:
    lr: 1.0
    betas:
      - 0.9
      - 0.98
    eps: 1.0e-9
scheduler: noamlr
scheduler_conf:
    model_size: 256
    warmup_steps: 100000

dataset_conf:
    data_names: speech_speaker_labels
    data_types: kaldi_ark
    batch_conf:
        batch_type: unsorted
        batch_size: 64
    num_workers: 8

log_interval: 50

 egs/callhome/eend_ola/conf/train_diar_eend_ola_simu_allspkr.yaml

New file
@@ -0,0 +1,52 @@
# network architecture
# encoder related
encoder: eend_ola_transformer
encoder_conf:
    idim: 345
    n_layers: 4
    n_units: 256

# encoder-decoder attractor related
encoder_decoder_attractor: eda
encoder_decoder_attractor_conf:
    n_units: 256

# model related
model: eend_ola_similar_eend
model_conf:
    max_n_speaker: 8

# optimization related
accum_grad: 1
grad_clip: 5
max_epoch: 25
val_scheduler_criterion:
    - valid
    - loss
best_model_criterion:
-   - valid
    - loss
    - min
keep_nbest_models: 100

optim: adam
optim_conf:
    lr: 1.0
    betas:
      - 0.9
      - 0.98
    eps: 1.0e-9
scheduler: noamlr
scheduler_conf:
    model_size: 256
    warmup_steps: 100000

dataset_conf:
    data_names: speech_speaker_labels
    data_types: kaldi_ark
    batch_conf:
        batch_type: unsorted
        batch_size: 64
    num_workers: 8

log_interval: 50

 egs/callhome/eend_ola/conf/train_diar_eend_ola_simu_allspkr_chunk2000.yaml

New file
@@ -0,0 +1,44 @@
# network architecture
# encoder related
encoder: eend_ola_transformer
encoder_conf:
    idim: 345
    n_layers: 4
    n_units: 256

# encoder-decoder attractor related
encoder_decoder_attractor: eda
encoder_decoder_attractor_conf:
    n_units: 256

# model related
model: eend_ola_similar_eend
model_conf:
    max_n_speaker: 8

# optimization related
accum_grad: 1
grad_clip: 5
max_epoch: 1
val_scheduler_criterion:
    - valid
    - loss
best_model_criterion:
-   - valid
    - loss
    - min
keep_nbest_models: 100

optim: adam
optim_conf:
    lr: 0.00001

dataset_conf:
    data_names: speech_speaker_labels
    data_types: kaldi_ark
    batch_conf:
        batch_type: unsorted
        batch_size: 8
    num_workers: 8

log_interval: 50

 egs/callhome/eend_ola/local/model_averaging.py

New file
@@ -0,0 +1,28 @@
#!/usr/bin/env python3

import argparse

import torch


def average_model(input_files, output_file):
    output_model = {}
    for ckpt_path in input_files:
        model_params = torch.load(ckpt_path, map_location="cpu")
        for key, value in model_params.items():
            if key not in output_model:
                output_model[key] = value
            else:
                output_model[key] += value
    for key in output_model.keys():
        output_model[key] /= len(input_files)
    torch.save(output_model, output_file)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("output_file")
    parser.add_argument("input_files", nargs='+')
    args = parser.parse_args()

    average_model(args.input_files, args.output_file)

 egs/callhome/eend_ola/path.sh

New file
@@ -0,0 +1,6 @@
export FUNASR_DIR=$PWD/../../..

# NOTE(kan-bayashi): Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=../../../:$PYTHONPATH
export PATH=$FUNASR_DIR/funasr/bin:$PATH

 egs/callhome/eend_ola/run.sh

New file
@@ -0,0 +1,242 @@
#!/usr/bin/env bash

. ./path.sh || exit 1;

# machines configuration
CUDA_VISIBLE_DEVICES="7"
gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
count=1

# general configuration
simu_feats_dir="/nfs/wangjiaming.wjm/EEND_ARK_DATA/dump/simu_data/data"
simu_feats_dir_chunk2000="/nfs/wangjiaming.wjm/EEND_ARK_DATA/dump/simu_data_chunk2000/data"
callhome_feats_dir_chunk2000="/nfs/wangjiaming.wjm/EEND_ARK_DATA/dump/callhome_chunk2000/data"
simu_train_dataset=train
simu_valid_dataset=dev
callhome_train_dataset=callhome1_allspk
callhome_valid_dataset=callhome2_allspk
callhome2_wav_scp_file=wav.scp

# model average
simu_average_2spkr_start=91
simu_average_2spkr_end=100
simu_average_allspkr_start=16
simu_average_allspkr_end=25
callhome_average_start=91
callhome_average_end=100

exp_dir="."
input_size=345
stage=1
stop_stage=4

# exp tag
tag="exp_fix"

. utils/parse_options.sh || exit 1;

# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
set -e
set -u
set -o pipefail

simu_2spkr_diar_config=conf/train_diar_eend_ola_simu_2spkr.yaml
simu_allspkr_diar_config=conf/train_diar_eend_ola_simu_allspkr.yaml
simu_allspkr_chunk2000_diar_config=conf/train_diar_eend_ola_simu_allspkr_chunk2000.yaml
callhome_diar_config=conf/train_diar_eend_ola_callhome_chunk2000.yaml
simu_2spkr_model_dir="baseline_$(basename "${simu_2spkr_diar_config}" .yaml)_${tag}"
simu_allspkr_model_dir="baseline_$(basename "${simu_allspkr_diar_config}" .yaml)_${tag}"
simu_allspkr_chunk2000_model_dir="baseline_$(basename "${simu_allspkr_chunk2000_diar_config}" .yaml)_${tag}"
callhome_model_dir="baseline_$(basename "${callhome_diar_config}" .yaml)_${tag}"

# Prepare data for training and inference
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
    echo "stage 0: Prepare data for training and inference"
fi

# Training on simulated two-speaker data
world_size=$gpu_num
simu_2spkr_ave_id=avg${simu_average_2spkr_start}-${simu_average_2spkr_end}
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
    echo "stage 1: Training on simulated two-speaker data"
    mkdir -p ${exp_dir}/exp/${simu_2spkr_model_dir}
    mkdir -p ${exp_dir}/exp/${simu_2spkr_model_dir}/log
    INIT_FILE=${exp_dir}/exp/${simu_2spkr_model_dir}/ddp_init
    if [ -f $INIT_FILE ];then
        rm -f $INIT_FILE
    fi
    init_method=file://$(readlink -f $INIT_FILE)
    echo "$0: init method is $init_method"
    for ((i = 0; i < $gpu_num; ++i)); do
        {
            rank=$i
            local_rank=$i
            gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
            train.py \
                --task_name diar \
                --gpu_id $gpu_id \
                --use_preprocessor false \
                --input_size $input_size \
                --data_dir ${simu_feats_dir} \
                --train_set ${simu_train_dataset} \
                --valid_set ${simu_valid_dataset} \
                --data_file_names "feats_2spkr.scp" \
                --resume true \
                --output_dir ${exp_dir}/exp/${simu_2spkr_model_dir} \
                --config $simu_2spkr_diar_config \
                --ngpu $gpu_num \
                --num_worker_count $count \
                --dist_init_method $init_method \
                --dist_world_size $world_size \
                --dist_rank $rank \
                --local_rank $local_rank 1> ${exp_dir}/exp/${simu_2spkr_model_dir}/log/train.log.$i 2>&1
        } &
        done
        wait
    echo "averaging model parameters into ${exp_dir}/exp/$simu_2spkr_model_dir/$simu_2spkr_ave_id.pb"
    models=`eval echo ${exp_dir}/exp/${simu_2spkr_model_dir}/{$simu_average_2spkr_start..$simu_average_2spkr_end}epoch.pb`
    python local/model_averaging.py ${exp_dir}/exp/${simu_2spkr_model_dir}/$simu_2spkr_ave_id.pb $models
fi

# Training on simulated all-speaker data
world_size=$gpu_num
simu_allspkr_ave_id=avg${simu_average_allspkr_start}-${simu_average_allspkr_end}
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
    echo "stage 2: Training on simulated all-speaker data"
    mkdir -p ${exp_dir}/exp/${simu_allspkr_model_dir}
    mkdir -p ${exp_dir}/exp/${simu_allspkr_model_dir}/log
    INIT_FILE=${exp_dir}/exp/${simu_allspkr_model_dir}/ddp_init
    if [ -f $INIT_FILE ];then
        rm -f $INIT_FILE
    fi
    init_method=file://$(readlink -f $INIT_FILE)
    echo "$0: init method is $init_method"
    for ((i = 0; i < $gpu_num; ++i)); do
        {
            rank=$i
            local_rank=$i
            gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
            train.py \
                --task_name diar \
                --gpu_id $gpu_id \
                --use_preprocessor false \
                --input_size $input_size \
                --data_dir ${simu_feats_dir} \
                --train_set ${simu_train_dataset} \
                --valid_set ${simu_valid_dataset} \
                --data_file_names "feats.scp" \
                --resume true \
                --init_param ${exp_dir}/exp/${simu_2spkr_model_dir}/$simu_2spkr_ave_id.pb \
                --output_dir ${exp_dir}/exp/${simu_allspkr_model_dir} \
                --config $simu_allspkr_diar_config \
                --ngpu $gpu_num \
                --num_worker_count $count \
                --dist_init_method $init_method \
                --dist_world_size $world_size \
                --dist_rank $rank \
                --local_rank $local_rank 1> ${exp_dir}/exp/${simu_allspkr_model_dir}/log/train.log.$i 2>&1
        } &
        done
        wait
    echo "averaging model parameters into ${exp_dir}/exp/$simu_allspkr_model_dir/$simu_allspkr_ave_id.pb"
    models=`eval echo ${exp_dir}/exp/${simu_allspkr_model_dir}/{$simu_average_allspkr_start..$simu_average_allspkr_end}epoch.pb`
    python local/model_averaging.py ${exp_dir}/exp/${simu_allspkr_model_dir}/$simu_allspkr_ave_id.pb $models
fi

# Training on simulated all-speaker data with chunk_size=2000
world_size=$gpu_num
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
    echo "stage 3: Training on simulated all-speaker data with chunk_size=2000"
    mkdir -p ${exp_dir}/exp/${simu_allspkr_chunk2000_model_dir}
    mkdir -p ${exp_dir}/exp/${simu_allspkr_chunk2000_model_dir}/log
    INIT_FILE=${exp_dir}/exp/${simu_allspkr_chunk2000_model_dir}/ddp_init
    if [ -f $INIT_FILE ];then
        rm -f $INIT_FILE
    fi
    init_method=file://$(readlink -f $INIT_FILE)
    echo "$0: init method is $init_method"
    for ((i = 0; i < $gpu_num; ++i)); do
        {
            rank=$i
            local_rank=$i
            gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
            train.py \
                --task_name diar \
                --gpu_id $gpu_id \
                --use_preprocessor false \
                --input_size $input_size \
                --data_dir ${simu_feats_dir_chunk2000} \
                --train_set ${simu_train_dataset} \
                --valid_set ${simu_valid_dataset} \
                --data_file_names "feats.scp" \
                --resume true \
                --init_param ${exp_dir}/exp/${simu_allspkr_model_dir}/$simu_allspkr_ave_id.pb \
                --output_dir ${exp_dir}/exp/${simu_allspkr_chunk2000_model_dir} \
                --config $simu_allspkr_chunk2000_diar_config \
                --ngpu $gpu_num \
                --num_worker_count $count \
                --dist_init_method $init_method \
                --dist_world_size $world_size \
                --dist_rank $rank \
                --local_rank $local_rank 1> ${exp_dir}/exp/${simu_allspkr_chunk2000_model_dir}/log/train.log.$i 2>&1
        } &
        done
        wait
fi

# Training on callhome all-speaker data with chunk_size=2000
world_size=$gpu_num
callhome_ave_id=avg${callhome_average_start}-${callhome_average_end}
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
    echo "stage 4: Training on callhome all-speaker data with chunk_size=2000"
    mkdir -p ${exp_dir}/exp/${callhome_model_dir}
    mkdir -p ${exp_dir}/exp/${callhome_model_dir}/log
    INIT_FILE=${exp_dir}/exp/${callhome_model_dir}/ddp_init
    if [ -f $INIT_FILE ];then
        rm -f $INIT_FILE
    fi
    init_method=file://$(readlink -f $INIT_FILE)
    echo "$0: init method is $init_method"
    for ((i = 0; i < $gpu_num; ++i)); do
        {
            rank=$i
            local_rank=$i
            gpu_id=$(echo $CUDA_VISIBLE_DEVICES | cut -d',' -f$[$i+1])
            train.py \
                --task_name diar \
                --gpu_id $gpu_id \
                --use_preprocessor false \
                --input_size $input_size \
                --data_dir ${callhome_feats_dir_chunk2000} \
                --train_set ${callhome_train_dataset} \
                --valid_set ${callhome_valid_dataset} \
                --data_file_names "feats.scp" \
                --resume true \
                --init_param ${exp_dir}/exp/${simu_allspkr_chunk2000_model_dir}/1epoch.pb \
                --output_dir ${exp_dir}/exp/${callhome_model_dir} \
                --config $callhome_diar_config \
                --ngpu $gpu_num \
                --num_worker_count $count \
                --dist_init_method $init_method \
                --dist_world_size $world_size \
                --dist_rank $rank \
                --local_rank $local_rank 1> ${exp_dir}/exp/${callhome_model_dir}/log/train.log.$i 2>&1
        } &
        done
        wait
    echo "averaging model parameters into ${exp_dir}/exp/$callhome_model_dir/$callhome_ave_id.pb"
    models=`eval echo ${exp_dir}/exp/${callhome_model_dir}/{$callhome_average_start..$callhome_average_end}epoch.pb`
    python local/model_averaging.py ${exp_dir}/exp/${callhome_model_dir}/$callhome_ave_id.pb $models
fi

## inference
#if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
#    echo "Inference"
#    mkdir -p ${exp_dir}/exp/${callhome_model_dir}/inference/log
#    CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES python local/infer.py \
#        --config_file ${exp_dir}/exp/${callhome_model_dir}/config.yaml \
#        --model_file ${exp_dir}/exp/${callhome_model_dir}/$callhome_ave_id.pb \
#        --output_rttm_file ${exp_dir}/exp/${callhome_model_dir}/inference/rttm \
#        --wav_scp_file ${callhome_feats_dir_chunk2000}/${callhome_valid_dataset}/${callhome2_wav_scp_file} 1> ${exp_dir}/exp/${callhome_model_dir}/inference/log/infer.log 2>&1
#fi

 egs/callhome/eend_ola/utils

New file
@@ -0,0 +1 @@
../../aishell/transformer/utils