From 57eb1cb68d1c567b25ac256444e5c1a77b8817a7 Mon Sep 17 00:00:00 2001
From: Sam Shleifer <sshleifer@gmail.com>
Date: Mon, 3 Aug 2020 18:22:31 -0400
Subject: [PATCH] [s2s] Document better mbart finetuning command (#6229)

* Document better MT command

* improve multigpu command
---
 examples/seq2seq/README.md                | 10 ++++------
 examples/seq2seq/train_mbart_cc25_enro.sh |  2 +-
 2 files changed, 5 insertions(+), 7 deletions(-)

diff --git a/examples/seq2seq/README.md b/examples/seq2seq/README.md
index 9d12dc33481..dd026784169 100644
--- a/examples/seq2seq/README.md
+++ b/examples/seq2seq/README.md
@@ -113,22 +113,20 @@ Best performing command:
 # optionally
 export ENRO_DIR='wmt_en_ro' # Download instructions above
 # export WANDB_PROJECT="MT" # optional
-export MAX_LEN=200
+export MAX_LEN=128
 export BS=4
-export GAS=8 # gradient accumulation steps
 ./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --label_smoothing 0.1 --fp16_opt_level=O1 --logger_name wandb --sortish_sampler
 ```
-This should take < 6h/epoch on a 16GB v100 and achieve val_avg_ BLEU score above 25. (you can see metrics in wandb or metrics.json).
-To get results in line with fairseq, you need to do some postprocessing.
+This should take < 6h/epoch on a 16GB v100 and achieve test BLEU above 26
+To get results in line with fairseq, you need to do some postprocessing. (see `romanian_postprocessing.md`)
 
 MultiGPU command
 (using 8 GPUS as an example)
 ```bash
 export ENRO_DIR='wmt_en_ro' # Download instructions above
  # export WANDB_PROJECT="MT" # optional
-export MAX_LEN=200
+export MAX_LEN=128
 export BS=4
-export GAS=1  # gradient accumulation steps
 ./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --gpus 8 --logger_name wandb
 ```
 ### Finetuning Outputs
diff --git a/examples/seq2seq/train_mbart_cc25_enro.sh b/examples/seq2seq/train_mbart_cc25_enro.sh
index b8122aee3f4..90bfce3e94d 100755
--- a/examples/seq2seq/train_mbart_cc25_enro.sh
+++ b/examples/seq2seq/train_mbart_cc25_enro.sh
@@ -10,7 +10,7 @@ python finetune.py \
     --num_train_epochs 6 --src_lang en_XX --tgt_lang ro_RO \
     --data_dir $ENRO_DIR \
     --max_source_length $MAX_LEN --max_target_length $MAX_LEN --val_max_target_length $MAX_LEN --test_max_target_length $MAX_LEN \
-    --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS \
+    --train_batch_size=$BS --eval_batch_size=$BS \
     --task translation \
     --warmup_steps 500 \
     --freeze_embeds \