|
#!/usr/bin/env bash |
|
|
|
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 |
|
export LOCAL_RANK=0 |
|
export WORLD_SIZE=8 |
|
|
|
export TOKENIZERS_PARALLELISM=false |
|
export NUM_WORKER=4 |
|
export MAX_STEPS=100000 |
|
|
|
export CUDA_LAUNCH_BLOCKING=1 |
|
export NCCL_DEBUG=INFO |
|
|
|
export EXP_NAME=cc.T03b_absum50.moco-2e14.contriever256-special50.bert-base-uncased.avg.dot.q128d256.step100k.bs1024.lr5e5 |
|
export PROJECT_DIR=output_dir/$EXP_NAME |
|
mkdir -p $PROJECT_DIR |
|
cp "$0" $PROJECT_DIR |
|
echo $PROJECT_DIR |
|
|
|
export WANDB_NAME=$EXP_NAME |
|
export WANDB_PROJECT=unsup_retrieval_cc |
|
export WANDB_DIR=$PROJECT_DIR |
|
mkdir -p $WANDB_DIR/wandb |
|
|
|
nohup python -m torch.distributed.launch --nproc_per_node=8 --master_port=31133 --max_restarts=0 train.py --model_name_or_path bert-base-uncased --arch_type moco --train_file /export/home/data/search/upr/cc/T03B_PileCC_absum.json --dev_file /export/home/data/pretrain/wiki2021_structure/wiki_psgs_w100.dev.tail2e13.tsv --data_type hf --data_pipeline_name contriever256-special50% --remove_unused_columns False --sim_type dot --queue_size 16384 --momentum 0.9995 --output_dir $PROJECT_DIR --cache_dir /export/home/data/pretrain/.cache --max_steps $MAX_STEPS --warmup_steps 10000 --logging_steps 100 --eval_steps 10000 --save_steps 1000000 --per_device_train_batch_size 256 --per_device_eval_batch_size 256 --dataloader_num_workers $NUM_WORKER --learning_rate 5e-5 --max_q_tokens 128 --max_d_tokens 256 --evaluation_strategy steps --load_best_model_at_end --overwrite_output_dir --do_train --do_eval --run_name $EXP_NAME --fp16 --seed 42 --report_to wandb --wiki_passage_path /export/home/data/search/nq/psgs_w100.tsv --qa_datasets_path /export/home/data/search/nq/qas/*-test.csv,/export/home/data/search/nq/qas/entityqs/test/P*.test.json > $PROJECT_DIR/nohup.log 2>&1 & echo $! > run.pid |
|
|
|
|
|
|
|
|
|
|
