update eend_ola

egs/callhome/eend_ola/conf/train_diar_eend_ola_callhome_chunk2000.yaml

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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

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