Fix SpeechLLM recipe (#1926)

This commit is contained in:
Yifan Yang 2025-04-23 16:18:38 +08:00 committed by GitHub
parent 64c5364085
commit 5ec95e5482
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 13 additions and 9 deletions

View File

@ -42,8 +42,8 @@ huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishel
# For multi-hans fine-tuned whisper model
# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt
# huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct
huggingface-clie download --local-dir models/qwen Qwen/Qwen2-1.5B-Instruct
# huggingface-cli download --local-dir models/qwen Qwen/Qwen2-7B-Instruct
huggingface-cli download --local-dir models/qwen Qwen/Qwen2-1.5B-Instruct
# First, we only train the projector and freeze other modules.
torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \
@ -57,7 +57,7 @@ torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \
--use-flash-attn True \
--use-lora False --unfreeze-llm False
# Then we jointly train the projector and LLM LoRA modules.
# Then, we jointly train the projector and LLM LoRA modules.
torchrun --nproc_per_node 8 ./whisper_llm_zh/train.py \
--max-duration 200 \
--exp-dir ./whisper_llm_zh/exp_test \
@ -81,7 +81,7 @@ huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_aishel
# For multi-hans fine-tuned whisper model
# huggingface-cli download --local-dir models/whisper yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt
huggingface-clie download --local-dir models/qwen Qwen/Qwen2-7B-Instruct
huggingface-cli download --local-dir models/qwen Qwen/Qwen2-7B-Instruct
mkdir -p whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B
ln -s models/checkpoint/epoch-10-avg-5.pt whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B/epoch-999.pt

13
egs/speech_llm/ASR_LLM/prepare.sh Normal file → Executable file
View File

@ -7,6 +7,9 @@ set -eou pipefail
stage=0
stop_stage=0
. shared/parse_options.sh || exit 1
# All files generated by this script are saved in "data".
# You can safely remove "data" and rerun this script to regenerate it.
mkdir -p data
@ -23,7 +26,7 @@ if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
# pip install huggingface_hub['cli']
# for aishell 1
huggingface-cli download --local-dir data yuekai/aishell_whisper_fbank_lhotse
huggingface-cli download --repo-type dataset --local-dir data yuekai/aishell_whisper_fbank_lhotse
fi
@ -31,9 +34,9 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
log "stage 1: Download whisper-large-v2 multi-hans-zh fbank feature from huggingface"
# for multi-hans-zh
huggingface-cli download --local-dir data/fbank yuekai/wenetspeech_whisper_fbank_lhotse
huggingface-cli download --local-dir data/fbank yuekai/multi_hans_zh_whisper_fbank_lhotse
huggingface-cli download --local-dir data/fbank yuekai/alimeeting_aishell4_training_whisper_fbank_lhotse
huggingface-cli download --repo-type dataset --local-dir data/fbank yuekai/wenetspeech_whisper_fbank_lhotse
huggingface-cli download --repo-type dataset --local-dir data/fbank yuekai/multi_hans_zh_whisper_fbank_lhotse
huggingface-cli download --repo-type dataset --local-dir data/fbank yuekai/alimeeting_aishell4_training_whisper_fbank_lhotse
fi
if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
@ -41,6 +44,6 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
# for speechio test sets
mkdir data_speechio
huggingface-cli download --local-dir data_speechio yuekai/icefall_asr_speechio
huggingface-cli download --repo-type model --local-dir data_speechio yuekai/icefall_asr_speechio
mv data_speechio/fbank/* data/fbank
fi

View File

@ -0,0 +1 @@
../../../icefall/shared