Official repository of OFA (ICML 2022). Paper: OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework
Hi @Tangshengku,
I've been trying to obtain the MuE weights following the steps provided, but I seem to be encountering some issues. I would greatly appreciate it if you could take a look at my process and point out any potential problems or share your script, which might be helpful to me.
train_caption_stage1_base_MuE.sh
#!/usr/bin/env
# The port for communication. Note that if you want to run multiple tasks on the same machine,
# you need to specify different port numbers.
export MASTER_PORT=1061
log_dir=./stg1_8_3e-6_MuE
save_dir=./stg1_8_3e-6_chk_MuE
mkdir -p $log_dir $save_dir
bpe_dir=../../utils/BPE
user_dir=../../ofa_module
data_dir=../../dataset/caption_data
data=${data_dir}/caption_stage1_train.tsv,${data_dir}/caption_val.tsv
restore_file=../../checkpoints/ofa_base.pt
selected_cols=0,4,2
task=caption
arch=ofa_base
criterion=MuE_Task_Loss
label_smoothing=0.1
lr=3e-5
max_epoch=6
warmup_ratio=0.06
batch_size=8
update_freq=4
resnet_drop_path_rate=0.0
encoder_drop_path_rate=0.1
decoder_drop_path_rate=0.1
dropout=0.1
attention_dropout=0.0
max_src_length=80
max_tgt_length=20
num_bins=1000
patch_image_size=480
eval_cider_cached=${data_dir}/cider_cached_tokens/coco-valid-words.p
drop_worst_ratio=0.2
for max_epoch in {6,}; do
echo "max_epoch "${max_epoch}
for warmup_ratio in {0.06,}; do
echo "warmup_ratio "${warmup_ratio}
for drop_worst_after in {6000,}; do
echo "drop_worst_after "${drop_worst_after}
log_file=${log_dir}/${max_epoch}"_"${warmup_ratio}"_"${drop_worst_after}".log"
save_path=${save_dir}/${max_epoch}"_"${warmup_ratio}"_"${drop_worst_after}
mkdir -p $save_path
CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch --nproc_per_node=2 --master_port=${MASTER_PORT} ../../train.py \
$data \
--selected-cols=${selected_cols} \
--bpe-dir=${bpe_dir} \
--user-dir=${user_dir} \
--restore-file=${restore_file} \
--reset-optimizer --reset-dataloader --reset-meters \
--save-dir=${save_path} \
--task=${task} \
--arch=${arch} \
--criterion=${criterion} \
--label-smoothing=${label_smoothing} \
--batch-size=${batch_size} \
--update-freq=${update_freq} \
--encoder-normalize-before \
--decoder-normalize-before \
--share-decoder-input-output-embed \
--share-all-embeddings \
--layernorm-embedding \
--patch-layernorm-embedding \
--code-layernorm-embedding \
--resnet-drop-path-rate=${resnet_drop_path_rate} \
--encoder-drop-path-rate=${encoder_drop_path_rate} \
--decoder-drop-path-rate=${decoder_drop_path_rate} \
--dropout=${dropout} \
--attention-dropout=${attention_dropout} \
--weight-decay=0.01 --optimizer=adam --adam-betas="(0.9,0.999)" --adam-eps=1e-08 --clip-norm=1.0 \
--lr-scheduler=polynomial_decay --lr=${lr} \
--max-epoch=${max_epoch} --warmup-ratio=${warmup_ratio} \
--log-format=simple --log-interval=10 \
--fixed-validation-seed=7 \
--no-epoch-checkpoints --keep-best-checkpoints=1 \
--save-interval=1 --validate-interval=1 \
--save-interval-updates=500 --validate-interval-updates=500 \
--eval-cider \
--eval-cider-cached-tokens=${eval_cider_cached} \
--eval-args='{"beam":5,"max_len_b":16,"no_repeat_ngram_size":3}' \
--best-checkpoint-metric=cider --maximize-best-checkpoint-metric \
--max-src-length=${max_src_length} \
--max-tgt-length=${max_tgt_length} \
--find-unused-parameters \
--freeze-encoder-embedding \
--freeze-decoder-embedding \
--add-type-embedding \
--scale-attn \
--scale-fc \
--scale-heads \
--disable-entangle \
--num-bins=${num_bins} \
--patch-image-size=${patch_image_size} \
--drop-worst-ratio=${drop_worst_ratio} \
--drop-worst-after=6000 \
--fp16 \
--fp16-scale-window=512 \
--train_mue\
--num-workers=0 > ${log_file} 2>&1
done
done
done
train_caption_stage2_base.sh
#!/usr/bin/env
# The port for communication. Note that if you want to run multiple tasks on the same machine,
# you need to specify different port numbers.
export MASTER_PORT=1062
log_dir=./stg2_bs8_lr1e-5_uf4_ngpu2
save_dir=./stg2_bs8_lr1e-5_uf4_ngpu2_check
mkdir -p $log_dir $save_dir
bpe_dir=../../utils/BPE
user_dir=../../ofa_module
data_dir=../../dataset/caption_data
data=${data_dir}/caption_stage2_train.tsv,${data_dir}/caption_val.tsv
restore_file=stage1/stg1_bs8_lr1e-5_uf4_nmgpu2/stg1_10_1e-5_chk_MuE/best3/checkpoint_best.pt
selected_cols=1,4,2
task=caption
arch=ofa_base
criterion=scst_reward_criterion
label_smoothing=0.1
lr=1e-5
max_epoch=5
warmup_ratio=0.06
batch_size=8
update_freq=4
resnet_drop_path_rate=0.0
encoder_drop_path_rate=0.0
decoder_drop_path_rate=0.0
dropout=0.0
attention_dropout=0.0
max_src_length=80
max_tgt_length=20
num_bins=1000
patch_image_size=480
eval_cider_cached=${data_dir}/cider_cached_tokens/coco-valid-words.p
scst_cider_cached=${data_dir}/cider_cached_tokens/coco-train-words.p
for lr in {1e-5,}; do
echo "lr "${lr}
for max_epoch in {5,}; do
echo "max_epoch "${max_epoch}
log_file=${log_dir}/${lr}"_"${max_epoch}".log"
save_path=${save_dir}/${lr}"_"${max_epoch}
mkdir -p $save_path
CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch --nproc_per_node=2 --master_port=${MASTER_PORT} ../../train.py \
$data \
--selected-cols=${selected_cols} \
--bpe-dir=${bpe_dir} \
--user-dir=${user_dir} \
--restore-file=${restore_file} \
--reset-optimizer --reset-dataloader --reset-meters \
--save-dir=${save_path} \
--task=${task} \
--arch=${arch} \
--criterion=${criterion} \
--batch-size=${batch_size} \
--update-freq=${update_freq} \
--encoder-normalize-before \
--decoder-normalize-before \
--share-decoder-input-output-embed \
--share-all-embeddings \
--layernorm-embedding \
--patch-layernorm-embedding \
--code-layernorm-embedding \
--resnet-drop-path-rate=${resnet_drop_path_rate} \
--encoder-drop-path-rate=${encoder_drop_path_rate} \
--decoder-drop-path-rate=${decoder_drop_path_rate} \
--dropout=${dropout} \
--attention-dropout=${attention_dropout} \
--weight-decay=0.01 --optimizer=adam --adam-betas="(0.9,0.999)" --adam-eps=1e-08 --clip-norm=1.0 \
--lr-scheduler=polynomial_decay --lr=${lr} \
--max-epoch=${max_epoch} --warmup-ratio=${warmup_ratio} \
--log-format=simple --log-interval=10 \
--fixed-validation-seed=7 \
--no-epoch-checkpoints --keep-best-checkpoints=1 \
--save-interval=1 --validate-interval=1 \
--save-interval-updates=500 --validate-interval-updates=500 \
--eval-cider \
--eval-cider-cached-tokens=${eval_cider_cached} \
--eval-args='{"beam":5,"max_len_b":16,"no_repeat_ngram_size":3}' \
--best-checkpoint-metric=cider --maximize-best-checkpoint-metric \
--max-src-length=${max_src_length} \
--max-tgt-length=${max_tgt_length} \
--find-unused-parameters \
--freeze-encoder-embedding \
--freeze-decoder-embedding \
--add-type-embedding \
--scale-attn \
--scale-fc \
--scale-heads \
--disable-entangle \
--num-bins=${num_bins} \
--patch-image-size=${patch_image_size} \
--scst \
--scst-cider-cached-tokens=${scst_cider_cached} \
--scst-args='{"beam":5,"max_len_b":16,"no_repeat_ngram_size":3}' \
--memory-efficient-fp16 \
--fp16-scale-window=512 \
--num-workers=0 > ${log_file} 2>&1
done
done
evaluate_caption_base_MuE.sh
#!/usr/bin/env bash
# The port for communication. Note that if you want to run multiple tasks on the same machine,
# you need to specify different port numbers.
export MASTER_PORT=1091
user_dir=../../ofa_module
bpe_dir=../../utils/BPE
data=../../dataset/caption_data/caption_test.tsv
path=2_batch_64_check/3e-5_5/checkpoint_best.pt
result_path=../../results/caption
selected_cols=1,4,2
split='test'
CUDA_VISIBLE_DEVICES=0 python3 -m torch.distributed.launch --nproc_per_node=1 --master_port=${MASTER_PORT} ../../evaluate.py \
${data} \
--path=${path} \
--user-dir=${user_dir} \
--task=caption \
--batch-size=1 \
--log-format=simple --log-interval=10 \
--seed=7 \
--gen-subset=${split} \
--results-path=${result_path} \
--beam=5 \
--max-len-b=16 \
--no-repeat-ngram-size=3 \
--fp16 \
--num-workers=0 \
--img_thres=0.9 \
--txt_thres=0.95 \
--decoder_thres=0.9 \
--model-overrides="{\"data\":\"${data}\",\"bpe_dir\":\"${bpe_dir}\",\"eval_cider\":False,\"selected_cols\":\"${selected_cols}\"}"
python coco_eval.py ../../results/caption/test_predict.json ../../dataset/caption_data/test_caption_coco_format.json
When employing the provided script, stage2 finetuning employs TransformerEncoder, as indicated in the log I shared, and it does not employ early exit during inference. The performance scores are identical regardless of the different thresholds applied.
train_caption_stage2_base.sh with argument "--train_mue" added
#!/usr/bin/env
# The port for communication. Note that if you want to run multiple tasks on the same machine,
# you need to specify different port numbers.
export MASTER_PORT=1062
log_dir=./stg2_bs8_lr1e-5_uf4_ngpu2
save_dir=./stg2_bs8_lr1e-5_uf4_ngpu2_check
mkdir -p $log_dir $save_dir
bpe_dir=../../utils/BPE
user_dir=../../ofa_module
data_dir=../../dataset/caption_data
data=${data_dir}/caption_stage2_train.tsv,${data_dir}/caption_val.tsv
restore_file=stage1/stg1_bs8_lr1e-5_uf4_nmgpu2/stg1_10_1e-5_chk_MuE/best3/checkpoint_best.pt
selected_cols=1,4,2
task=caption
arch=ofa_base
criterion=scst_reward_criterion
label_smoothing=0.1
lr=1e-5
max_epoch=5
warmup_ratio=0.06
batch_size=8
update_freq=4
resnet_drop_path_rate=0.0
encoder_drop_path_rate=0.0
decoder_drop_path_rate=0.0
dropout=0.0
attention_dropout=0.0
max_src_length=80
max_tgt_length=20
num_bins=1000
patch_image_size=480
eval_cider_cached=${data_dir}/cider_cached_tokens/coco-valid-words.p
scst_cider_cached=${data_dir}/cider_cached_tokens/coco-train-words.p
for lr in {1e-5,}; do
echo "lr "${lr}
for max_epoch in {5,}; do
echo "max_epoch "${max_epoch}
log_file=${log_dir}/${lr}"_"${max_epoch}".log"
save_path=${save_dir}/${lr}"_"${max_epoch}
mkdir -p $save_path
CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch --nproc_per_node=2 --master_port=${MASTER_PORT} ../../train.py \
$data \
--selected-cols=${selected_cols} \
--bpe-dir=${bpe_dir} \
--user-dir=${user_dir} \
--restore-file=${restore_file} \
--reset-optimizer --reset-dataloader --reset-meters \
--save-dir=${save_path} \
--task=${task} \
--arch=${arch} \
--criterion=${criterion} \
--batch-size=${batch_size} \
--update-freq=${update_freq} \
--encoder-normalize-before \
--decoder-normalize-before \
--share-decoder-input-output-embed \
--share-all-embeddings \
--layernorm-embedding \
--patch-layernorm-embedding \
--code-layernorm-embedding \
--resnet-drop-path-rate=${resnet_drop_path_rate} \
--encoder-drop-path-rate=${encoder_drop_path_rate} \
--decoder-drop-path-rate=${decoder_drop_path_rate} \
--dropout=${dropout} \
--attention-dropout=${attention_dropout} \
--weight-decay=0.01 --optimizer=adam --adam-betas="(0.9,0.999)" --adam-eps=1e-08 --clip-norm=1.0 \
--lr-scheduler=polynomial_decay --lr=${lr} \
--max-epoch=${max_epoch} --warmup-ratio=${warmup_ratio} \
--log-format=simple --log-interval=10 \
--fixed-validation-seed=7 \
--no-epoch-checkpoints --keep-best-checkpoints=1 \
--save-interval=1 --validate-interval=1 \
--save-interval-updates=500 --validate-interval-updates=500 \
--eval-cider \
--eval-cider-cached-tokens=${eval_cider_cached} \
--eval-args='{"beam":5,"max_len_b":16,"no_repeat_ngram_size":3}' \
--best-checkpoint-metric=cider --maximize-best-checkpoint-metric \
--max-src-length=${max_src_length} \
--max-tgt-length=${max_tgt_length} \
--find-unused-parameters \
--freeze-encoder-embedding \
--freeze-decoder-embedding \
--add-type-embedding \
--scale-attn \
--scale-fc \
--scale-heads \
--disable-entangle \
--num-bins=${num_bins} \
--patch-image-size=${patch_image_size} \
--scst \
--scst-cider-cached-tokens=${scst_cider_cached} \
--scst-args='{"beam":5,"max_len_b":16,"no_repeat_ngram_size":3}' \
--memory-efficient-fp16 \
--fp16-scale-window=512 \
--train_mue \
--num-workers=0 > ${log_file} 2>&1
done
done
log for stage2 finetuning with argument "--train_mue" added
When I add the "--train_mue" argument during stage2 finetuning, it utilizes the Transformer_MuE and effectively performs an early exit during inference. However, the performance score is notably lower than the one reported in the paper.
Hi @Tangshengku, I've been trying to obtain the MuE weights following the steps provided, but I seem to be encountering some issues. I would greatly appreciate it if you could take a look at my process and point out any potential problems or share your script, which might be helpful to me.
train_caption_stage1_base_MuE.sh
train_caption_stage2_base.sh
evaluate_caption_base_MuE.sh
training log for finetuning stage2
When employing the provided script, stage2 finetuning employs TransformerEncoder, as indicated in the log I shared, and it does not employ early exit during inference. The performance scores are identical regardless of the different thresholds applied.
train_caption_stage2_base.sh with argument "--train_mue" added
log for stage2 finetuning with argument "--train_mue" added
When I add the "--train_mue" argument during stage2 finetuning, it utilizes the Transformer_MuE and effectively performs an early exit during inference. However, the performance score is notably lower than the one reported in the paper.