From aea7951a1c6b95266649a570a46759072dc8c434 Mon Sep 17 00:00:00 2001 From: Jakub Pokrywka Date: Thu, 4 Nov 2021 12:24:59 +0100 Subject: [PATCH] a --- how_to_run | 8 ++++++++ regular_roberta_from_scratch/2_run.sh | 6 +++--- roberta_ft/2_run.sh | 6 +++--- roberta_with_year_from_scratch/2_run.sh | 6 +++--- roberta_with_year_ft/2_run.sh | 6 +++--- 5 files changed, 20 insertions(+), 12 deletions(-) diff --git a/how_to_run b/how_to_run index 2efa665..a824979 100644 --- a/how_to_run +++ b/how_to_run @@ -4,3 +4,11 @@ instal datasets: pip install datasets install tokenizers: pip install tokenizers to run: go do a solution directory and: bash run.sh + + +adjust: +per_device_train_batch_size +per_device_eval_batch_size +gradient_accumulation_steps + +if you want to run scripts on other dataset remeber to adjust max_seq_length diff --git a/regular_roberta_from_scratch/2_run.sh b/regular_roberta_from_scratch/2_run.sh index ecb336c..7776251 100644 --- a/regular_roberta_from_scratch/2_run.sh +++ b/regular_roberta_from_scratch/2_run.sh @@ -8,9 +8,9 @@ python run_mlm.py \ --validation_file ./dev-0_in.csv \ --do_train \ --do_eval \ - --per_device_train_batch_size=1 \ - --per_device_eval_batch_size=1 \ - --gradient_accumulation_steps=4 \ + --per_device_train_batch_size=256 \ + --per_device_eval_batch_size=256 \ + --gradient_accumulation_steps=1 \ --fp16 False \ --save_steps 1000 \ --eval_steps 1000 \ diff --git a/roberta_ft/2_run.sh b/roberta_ft/2_run.sh index b764c89..ca29db7 100644 --- a/roberta_ft/2_run.sh +++ b/roberta_ft/2_run.sh @@ -8,9 +8,9 @@ python run_mlm.py \ --validation_file ./dev-0_in.csv \ --do_train \ --do_eval \ - --per_device_train_batch_size=32 \ - --per_device_eval_batch_size=32 \ - --gradient_accumulation_steps=8 \ + --per_device_train_batch_size=256 \ + --per_device_eval_batch_size=256 \ + --gradient_accumulation_steps=1 \ --fp16 False \ --save_steps 1000 \ --eval_steps 1000 \ diff --git a/roberta_with_year_from_scratch/2_run.sh b/roberta_with_year_from_scratch/2_run.sh index ecb336c..7776251 100644 --- a/roberta_with_year_from_scratch/2_run.sh +++ b/roberta_with_year_from_scratch/2_run.sh @@ -8,9 +8,9 @@ python run_mlm.py \ --validation_file ./dev-0_in.csv \ --do_train \ --do_eval \ - --per_device_train_batch_size=1 \ - --per_device_eval_batch_size=1 \ - --gradient_accumulation_steps=4 \ + --per_device_train_batch_size=256 \ + --per_device_eval_batch_size=256 \ + --gradient_accumulation_steps=1 \ --fp16 False \ --save_steps 1000 \ --eval_steps 1000 \ diff --git a/roberta_with_year_ft/2_run.sh b/roberta_with_year_ft/2_run.sh index b764c89..ca29db7 100644 --- a/roberta_with_year_ft/2_run.sh +++ b/roberta_with_year_ft/2_run.sh @@ -8,9 +8,9 @@ python run_mlm.py \ --validation_file ./dev-0_in.csv \ --do_train \ --do_eval \ - --per_device_train_batch_size=32 \ - --per_device_eval_batch_size=32 \ - --gradient_accumulation_steps=8 \ + --per_device_train_batch_size=256 \ + --per_device_eval_batch_size=256 \ + --gradient_accumulation_steps=1 \ --fp16 False \ --save_steps 1000 \ --eval_steps 1000 \