From 169c0cec8e9922d7ec0940cb2cc80008c48be342 Mon Sep 17 00:00:00 2001 From: dohe0342 Date: Wed, 24 May 2023 13:03:51 +0900 Subject: [PATCH] from local --- egs/librispeech/ASR/lora.sh | 119 ++++++++++++++++++ .../.train_lora.py.swp | Bin 86016 -> 86016 bytes 2 files changed, 119 insertions(+) create mode 100755 egs/librispeech/ASR/lora.sh diff --git a/egs/librispeech/ASR/lora.sh b/egs/librispeech/ASR/lora.sh new file mode 100755 index 000000000..7f4c40bfd --- /dev/null +++ b/egs/librispeech/ASR/lora.sh @@ -0,0 +1,119 @@ +workstation=$3 + +if [ $workstation = "whale" ]; then + #export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + export CUDA_VISIBLE_DEVICES=0 + if [ ! -e ./pruned_transducer_stateless_d2v_v2/$1/.train.done ]; then + ./pruned_transducer_stateless_d2v_v2/prompt_tuning.py \ + --num-buckets 2 \ + --add-adapter True \ + --adapter-lr 0.1 \ + --gender male \ + --wandb False \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim False \ + --world-size 1 \ + --num-epochs 10000 \ + --num-updates 101 \ + --save-every-n 50 \ + --full-libri 1 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 200 \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --use-fp16 0 \ + --accum-grads 8 \ + --encoder-type d2v \ + --additional-block True \ + --prune-range 10 \ + --prompt True \ + --spk-id $2 + #touch ./pruned_transducer_stateless_d2v_v2/$1/.train.done + fi + +# ./pruned_transducer_stateless_d2v_v2/train_adapter.py \ +# --add-adapter True \ +# --adapter-lr 0.001 \ +# --gender female +# --wandb False \ +# --input-strategy AudioSamples \ +# --enable-spec-aug False \ +# --multi-optim False \ +# --world-size 8 \ +# --num-epochs 10 \ +# --full-libri 1 \ +# --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ +# --max-duration 200 \ +# --use-fp16 0 \ +# --encoder-type d2v \ +# --additional-block True \ +# --encoder-dim 768 \ +# --decoder-dim 768 \ +# --joiner-dim 768 \ +# --prune-range 10 +else + export CUDA_VISIBLE_DEVICES="0,1,2,3" + #rm ./pruned_transducer_stateless_d2v_v2/$1/.train.done + if [ ! -e ./pruned_transducer_stateless_d2v_v2/$1/.train.done ]; then + ./pruned_transducer_stateless_d2v_v2/bitfit.py \ + --num-buckets 2 \ + --add-adapter True \ + --adapter-lr 0.01 \ + --gender male \ + --wandb False \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim False \ + --world-size 4 \ + --num-epochs 10000 \ + --num-updates 101 \ + --save-every-n 50 \ + --full-libri 1 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 150 \ + --accum-grads 2 \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --use-fp16 0 \ + --accum-grads 2 \ + --encoder-type d2v \ + --additional-block True \ + --prune-range 10 \ + --spk-id $2 \ + --prompt False \ + --prefix vox + touch ./pruned_transducer_stateless_d2v_v2/$1/.train.done + fi +fi + +#else +# export CUDA_VISIBLE_DEVICES="0,1,2,3" +# ./pruned_transducer_stateless_d2v_v2/train_adapter.py \ +# --num-buckets 2 \ +# --add-adapter True \ +# --adapter-lr 0.02472 \ +# --wandb False \ +# --input-strategy AudioSamples \ +# --enable-spec-aug False \ +# --multi-optim False \ +# --world-size 4 \ +# --num-epochs 31 \ +# --full-libri 1 \ +# --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ +# --max-duration 150 \ +# --encoder-dim 768 \ +# --decoder-dim 768 \ +# --joiner-dim 768 \ +# --use-fp16 0 \ +# --accum-grads 4 \ +# --encoder-type d2v \ +# --additional-block True \ +# --prune-range 10 \ +# --ctc-loss-scale 0.1924 \ +# --lm-scale 0.1254 \ +# --simple-loss-scale 0.2869 \ +# --spk-id $2 +#fi diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train_lora.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_d2v_v2/.train_lora.py.swp index a5e538907fb3a1f5c01f855e419bce57fa90f08e..d6eb1ceea1655453fb63925660e6e84cffc5f606 100644 GIT binary patch delta 35 pcmZozz}m2YRW!*U%+puFQqO<^2m}}yO1pAX8X85niZZ_T2LPrr36KB) delta 35 pcmZozz}m2YRW!*U%+puFQqO<^2m}}y3c7Mr!Wu=liZZ_T2LPm|31I*L