From bee8346c4b0fd9eb4acb8910620be6173f31cf92 Mon Sep 17 00:00:00 2001
From: 志浩 <neo.dzh@alibaba-inc.com>
Date: 星期三, 02 八月 2023 10:59:31 +0800
Subject: [PATCH] TOLD/SOND: update finetune and train recipe
---
egs/callhome/diarization/sond/finetune.sh | 25 ++++++++----
egs/callhome/diarization/sond/run.sh | 78 +++++++--------------------------------
2 files changed, 31 insertions(+), 72 deletions(-)
diff --git a/egs/callhome/diarization/sond/finetune.sh b/egs/callhome/diarization/sond/finetune.sh
index 8e161f9..84ec103 100644
--- a/egs/callhome/diarization/sond/finetune.sh
+++ b/egs/callhome/diarization/sond/finetune.sh
@@ -8,13 +8,18 @@
# [2] Speaker Overlap-aware Neural Diarization for Multi-party Meeting Analysis, EMNLP 2022
# We recommend you run this script stage by stage.
+# This recipe includes:
+# 1. downloading a pretrained model on the simulated data from switchboard and NIST,
+# 2. finetuning the pretrained model on Callhome1.
+# Finally, you will get a slightly better DER result 9.95% on Callhome2 than that in the paper 10.14%.
+
# environment configuration
if [ ! -e utils ]; then
ln -s ../../../aishell/transformer/utils ./utils
fi
# machines configuration
-gpu_devices="0,1,2,3"
+gpu_devices="0,1,2,3" # for V100-16G, need 4 gpus.
gpu_num=4
count=1
@@ -76,10 +81,14 @@
# Download required resources
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
echo "Stage 0: Download required resources."
- wget told_finetune_resources.zip
+ if [ ! -e told_finetune_resources.tar.gz ]; then
+ # MD5SUM: abc7424e4e86ce6f040e9cba4178123b
+ wget --no-check-certificate https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/Speaker_Diar/told_finetune_resources.tar.gz
+ tar zxf told_finetune_resources.tar.gz
+ fi
fi
-# Finetune model on callhome1
+# Finetune model on callhome1, this will take about 1.5 hours.
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "Stage 1: Finetune pretrained model on callhome1."
world_size=$gpu_num # run on one machine
@@ -230,11 +239,11 @@
# Then find the wav files to construct wav.scp and put it at data/callhome2/wav.scp.
# After iteratively perform SOAP, you will get DER results like:
# iters : oracle_vad | system_vad
-# iter_0: 9.68 | 10.51
-# iter_1: 9.26 | 10.14 (reported in the paper)
-# iter_2: 9.18 | 10.08
-# iter_3: 9.24 | 10.15
-# iter_4: 9.27 | 10.17
+# iter_0: 9.63 | 10.43
+# iter_1: 9.17 | 10.03
+# iter_2: 9.11 | 9.98
+# iter_3: 9.08 | 9.96
+# iter_4: 9.07 | 9.95
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
if [ ! -e ${expdir}/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch ]; then
git lfs install
diff --git a/egs/callhome/diarization/sond/run.sh b/egs/callhome/diarization/sond/run.sh
index 3758f0c..c0ecd35 100644
--- a/egs/callhome/diarization/sond/run.sh
+++ b/egs/callhome/diarization/sond/run.sh
@@ -8,6 +8,15 @@
# [2] Speaker Overlap-aware Neural Diarization for Multi-party Meeting Analysis, EMNLP 2022
# We recommend you run this script stage by stage.
+# [developing] This recipe includes:
+# 1. simulating data with switchboard and NIST.
+# 2. training the model from scratch for 3 stages:
+# 2-1. pre-train on simu_swbd_sre
+# 2-2. train on simu_swbd_sre
+# 2-3. finetune on callhome1
+# 3. evaluating model with the results from the first stage EEND-OLA,
+# Finally, you will get a similar DER result claimed in the paper.
+
# environment configuration
kaldi_root=
@@ -26,8 +35,8 @@
fi
# machines configuration
-gpu_devices="6,7"
-gpu_num=2
+gpu_devices="4,5,6,7" # for V100-16G, use 4 GPUs
+gpu_num=4
count=1
# general configuration
@@ -417,7 +426,7 @@
rank=$i
local_rank=$i
gpu_id=$(echo $gpu_devices | cut -d',' -f$[$i+1])
- diar_train.py \
+ python -m funasr.bin.diar_train \
--gpu_id $gpu_id \
--use_preprocessor false \
--token_type char \
@@ -565,7 +574,7 @@
rank=$i
local_rank=$i
gpu_id=$(echo $gpu_devices | cut -d',' -f$[$i+1])
- diar_train.py \
+ python -m funasr.bin.diar_train \
--gpu_id $gpu_id \
--use_preprocessor false \
--token_type char \
@@ -710,7 +719,7 @@
rank=$i
local_rank=$i
gpu_id=$(echo $gpu_devices | cut -d',' -f$[$i+1])
- diar_train.py \
+ python -m funasr.bin.diar_train \
--gpu_id $gpu_id \
--use_preprocessor false \
--token_type char \
@@ -942,62 +951,3 @@
echo "Done."
done
fi
-
-
-if [ ${stage} -le 30 ] && [ ${stop_stage} -ge 30 ]; then
- echo "stage 30: training phase 1, pretraining on simulated data"
- world_size=$gpu_num # run on one machine
- mkdir -p ${expdir}/${model_dir}
- mkdir -p ${expdir}/${model_dir}/log
- mkdir -p /tmp/${model_dir}
- INIT_FILE=/tmp/${model_dir}/ddp_init
- if [ -f $INIT_FILE ];then
- rm -f $INIT_FILE
- fi
- init_opt=""
- if [ ! -z "${init_param}" ]; then
- init_opt="--init_param ${init_param}"
- echo ${init_opt}
- fi
-
- freeze_opt=""
- if [ ! -z "${freeze_param}" ]; then
- freeze_opt="--freeze_param ${freeze_param}"
- echo ${freeze_opt}
- fi
-
- init_method=file://$(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 $gpu_devices | cut -d',' -f$[$i+1])
- diar_train.py \
- --gpu_id $gpu_id \
- --use_preprocessor false \
- --token_type char \
- --token_list $token_list \
- --dataset_type large \
- --train_data_file ${datadir}/${train_set}/dumped_files/data_file.list \
- --valid_data_file ${datadir}/${valid_set}/dumped_files/data_file.list \
- --init_param ${expdir}/speech_xvector_sv-en-us-callhome-8k-spk6135-pytorch/sv.pth:encoder:encoder \
- --freeze_param encoder \
- ${init_opt} \
- ${freeze_opt} \
- --ignore_init_mismatch true \
- --resume true \
- --output_dir ${expdir}/${model_dir} \
- --config $train_config \
- --ngpu $gpu_num \
- --num_worker_count $count \
- --multiprocessing_distributed true \
- --dist_init_method $init_method \
- --dist_world_size $world_size \
- --dist_rank $rank \
- --local_rank $local_rank 1> ${expdir}/${model_dir}/log/train.log.$i 2>&1
- } &
- done
- echo "Training log can be found at ${expdir}/${model_dir}/log/train.log.*"
- wait
-fi
\ No newline at end of file
--
Gitblit v1.9.1