transformers/examples/seq2seq

Sequence to Sequence

This directory contains examples for finetuning and evaluating transformers on summarization and translation tasks. Please tag @sshleifer with any issues/unexpected behaviors, or send a PR! For bertabs instructions, see bertabs/README.md.

Supported Architectures

• BartForConditionalGeneration (and anything that inherits from it)
• MarianMTModel
• PegasusForConditionalGeneration
• MBartForConditionalGeneration
• FSMTForConditionalGeneration
• T5ForConditionalGeneration

Datasets

XSUM:

cd examples/seq2seq
wget https://cdn-datasets.huggingface.co/summarization/xsum.tar.gz
tar -xzvf xsum.tar.gz
export XSUM_DIR=${PWD}/xsum

this should make a directory called xsum/ with files like test.source. To use your own data, copy that files format. Each article to be summarized is on its own line.

CNN/DailyMail

cd examples/seq2seq
wget https://cdn-datasets.huggingface.co/summarization/cnn_dm_v2.tgz
tar -xzvf cnn_dm_v2.tgz  # empty lines removed
mv cnn_cln cnn_dm
export CNN_DIR=${PWD}/cnn_dm

this should make a directory called cnn_dm/ with 6 files.

WMT16 English-Romanian Translation Data:

download with this command:

wget https://cdn-datasets.huggingface.co/translation/wmt_en_ro.tar.gz
tar -xzvf wmt_en_ro.tar.gz
export ENRO_DIR=${PWD}/wmt_en_ro

this should make a directory called wmt_en_ro/ with 6 files.

WMT English-German:

wget https://cdn-datasets.huggingface.co/translation/wmt_en_de.tgz
tar -xzvf wmt_en_de.tgz
export DATA_DIR=${PWD}/wmt_en_de

Private Data

If you are using your own data, it must be formatted as one directory with 6 files:

train.source
train.target
val.source
val.target
test.source
test.target

The .source files are the input, the .target files are the desired output.

Tips and Tricks

General Tips:

• since you need to run from examples/seq2seq, and likely need to modify code, the easiest workflow is fork transformers, clone your fork, and run pip install -e . before you get started.
• try --freeze_encoder or --freeze_embeds for faster training/larger batch size. (3hr per epoch with bs=8, see the "xsum_shared_task" command below)
• fp16_opt_level=O1 (the default works best).
• In addition to the pytorch-lightning .ckpt checkpoint, a transformers checkpoint will be saved. Load it with BartForConditionalGeneration.from_pretrained(f'{output_dir}/best_tfmr).
• At the moment, --do_predict does not work in a multi-gpu setting. You need to use evaluate_checkpoint or the run_eval.py code.
• This warning can be safely ignored:

  "Some weights of BartForConditionalGeneration were not initialized from the model checkpoint at facebook/bart-large-xsum and are newly initialized: ['final_logits_bias']"

• Both finetuning and eval are 30% faster with --fp16. For that you need to install apex.
• Read scripts before you run them!

Summarization Tips:

• (summ) 1 epoch at batch size 1 for bart-large takes 24 hours and requires 13GB GPU RAM with fp16 on an NVIDIA-V100.
• If you want to run experiments on improving the summarization finetuning process, try the XSUM Shared Task (below). It's faster to train than CNNDM because the summaries are shorter.
• For CNN/DailyMail, the default val_max_target_length and test_max_target_length will truncate the ground truth labels, resulting in slightly higher rouge scores. To get accurate rouge scores, you should rerun calculate_rouge on the {output_dir}/test_generations.txt file saved by trainer.test()
• --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 is a reasonable setting for XSUM.
• wandb can be used by specifying --logger_name wandb. It is useful for reproducibility. Specify the environment variable WANDB_PROJECT='hf_xsum' to do the XSUM shared task.
• If you are finetuning on your own dataset, start from distilbart-cnn-12-6 if you want long summaries and distilbart-xsum-12-6 if you want short summaries. (It rarely makes sense to start from bart-large unless you are a researching finetuning methods).

Update 2018-07-18 Datasets: LegacySeq2SeqDataset will be used for all tokenizers without a prepare_seq2seq_batch method. Otherwise, Seq2SeqDataset will be used. Future work/help wanted: A new dataset to support multilingual tasks.

Finetuning Scripts

All finetuning bash scripts call finetune.py (or distillation.py) with reasonable command line arguments. They usually require extra command line arguments to work.

To see all the possible command line options, run:

./finetune.py --help

Finetuning Training Params

To override the pretrained model's training params, you can pass them to ./finetune.sh:

./finetune.sh \
    [...]
    --encoder_layerdrop 0.1 \
    --decoder_layerdrop 0.1 \
    --dropout 0.1 \
    --attention_dropout 0.1 \

Summarization Finetuning

Run/modify finetune.sh

The following command should work on a 16GB GPU:

./finetune.sh \
    --data_dir $XSUM_DIR \
    --train_batch_size=1 \
    --eval_batch_size=1 \
    --output_dir=xsum_results \
    --num_train_epochs 6 \
    --model_name_or_path facebook/bart-large

There is a starter finetuning script for pegasus at finetune_pegasus_xsum.sh.

Translation Finetuning

First, follow the wmt_en_ro download instructions. Then you can finetune mbart_cc25 on english-romanian with the following command. Recommendation: Read and potentially modify the fairly opinionated defaults in train_mbart_cc25_enro.sh script before running it.

Best performing command:

# optionally
export ENRO_DIR='wmt_en_ro' # Download instructions above
# export WANDB_PROJECT="MT" # optional
export MAX_LEN=128
export BS=4
./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 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)

export ENRO_DIR='wmt_en_ro' # Download instructions above
 # export WANDB_PROJECT="MT" # optional
export MAX_LEN=128
export BS=4
./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --gpus 8 --logger_name wandb

Finetuning Outputs

As you train, output_dir will be filled with files, that look kind of like this (comments are mine). Some of them are metrics, some of them are checkpoints, some of them are metadata. Here is a quick tour:

output_dir
├── best_tfmr  # this is a huggingface checkpoint generated by save_pretrained. It is the same model as the PL .ckpt file below
│   ├── config.json
│   ├── merges.txt
│   ├── pytorch_model.bin
│   ├── special_tokens_map.json
│   ├── tokenizer_config.json
│   └── vocab.json
├── git_log.json   # repo, branch, and commit hash
├── val_avg_rouge2=0.1984-step_count=11.ckpt  # this is a pytorch lightning checkpoint associated with the best val score. (it will be called BLEU for MT)
├── metrics.json  # new validation metrics will continually be appended to this
├── student  # this is a huggingface checkpoint generated by SummarizationDistiller. It is the student before it gets finetuned.
│   ├── config.json
│   └── pytorch_model.bin
├── test_generations.txt
# ^^ are the summaries or translations produced by your best checkpoint on the test data. Populated when training is done
├── test_results.txt  # a convenience file with the test set metrics. This data is also in metrics.json['test']
├── hparams.pkl  # the command line args passed after some light preprocessing. Should be saved fairly quickly.

After training, you can recover the best checkpoint by running

from transformers import AutoModelForSeq2SeqLM
model = AutoModelForSeq2SeqLM.from_pretrained(f'{output_dir}/best_tfmr')

Fine-tuning using Seq2SeqTrainer

To use Seq2SeqTrainer for fine-tuning you should use the finetune_trainer.py script. It subclasses Trainer to extend it for seq2seq training. Except the Trainer releated TrainingArguments, it shares the same argument names as that of finetune.py file. One notable difference is that, calculating generative metrics (BLEU, ROUGE) is optional and is controlled using the --predict_with_generate argument, set this argument to calculate BLEU and ROUGE metrics.

With PyTorch 1.6+ it'll automatically use native AMP when --fp16 is set.

To see all the possible command line options, run:

./builtin_trainer/finetune.sh --help # This calls python finetune_trainer.py --help

At the moment, Seq2SeqTrainer does not support with teacher distillation.

All Seq2SeqTrainer based fine-tuning scripts are included in the builtin_trainer directory.

TPU Training

Seq2SeqTrainer supports TPU training with few caveats

1. As generate method does not work on TPU at the moment, predict_with_generate can not be used. You should use --prediction_loss_only to only calculate loss, and do not set --do_predict and --predict_with_generate.
2. All sequences should be padded to be of equal length otherwise it leads to extremely slow training. (finetune_trainer.py does this automatically when running on TPU.)

We provide a very simple launcher script named xla_spawn.py that lets you run our example scripts on multiple TPU cores without any boilerplate. Just pass a --num_cores flag to this script, then your regular training script with its arguments (this is similar to the torch.distributed.launch helper for torch.distributed).

builtin_trainer/finetune_tpu.sh script provides minimal arguments needed for TPU training.

Following command fine-tunes sshleifer/student_marian_en_ro_6_3 on TPU V3-8 and should complete one epoch in ~5-6 mins.

./builtin_trainer/train_distil_marian_enro_tpu.sh

Evaluation Commands

To create summaries for each article in dataset, we use run_eval.py, here are a few commands that run eval for different tasks and models. If 'translation' is in your task name, the computed metric will be BLEU. Otherwise, ROUGE will be used.

For t5, you need to specify --task translation_{src}to{tgt} as follows:

export DATA_DIR=wmt_en_ro
./run_eval.py t5-base \
    $DATA_DIR/val.source t5_val_generations.txt \
    --reference_path $DATA_DIR/val.target \
    --score_path enro_bleu.json \
    --task translation_en_to_ro \
    --n_obs 100 \
    --device cuda \
    --fp16 \
    --bs 32

This command works for MBART, although the BLEU score is suspiciously low.

export DATA_DIR=wmt_en_ro
./run_eval.py facebook/mbart-large-en-ro $DATA_DIR/val.source mbart_val_generations.txt \
    --reference_path $DATA_DIR/val.target \
    --score_path enro_bleu.json \
    --task translation \
    --n_obs 100 \
    --device cuda \
    --fp16 \
    --bs 32

Summarization (xsum will be very similar):

export DATA_DIR=cnn_dm
./run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_generations.txt \
    --reference_path $DATA_DIR/val.target \
    --score_path cnn_rouge.json \
    --task summarization \
    --n_obs 100 \
    --device cuda \
    --max_source_length 1024 \
    --max_target_length 56 \
    --fp16 \
    --bs 32

Multi-GPU Evaluation

here is a command to run xsum evaluation on 8 GPUS. It is more than linearly faster than run_eval.py in some cases because it uses SortishSampler to minimize padding. You can also use it on 1 GPU. data_dir must have {type_path}.source and {type_path}.target. Run ./run_distributed_eval.py --help for all clargs.

python -m torch.distributed.launch --nproc_per_node=8  run_distributed_eval.py \
    --model_name sshleifer/distilbart-large-xsum-12-3  \
    --save_dir xsum_generations \
    --data_dir xsum \
    --fp16  # you can pass generate kwargs like num_beams here, just like run_eval.py

Contributions that implement this command for other distributed hardware setups are welcome!

Single-GPU Eval: Tips and Tricks

When using run_eval.py, the following features can be useful:

• if you running the script multiple times and want to make it easier to track what arguments produced that output, use --dump-args. Along with the results it will also dump any custom params that were passed to the script. For example if you used: --num_beams 8 --early_stopping true, the output will be:

  {'bleu': 26.887, 'n_obs': 10, 'runtime': 1, 'seconds_per_sample': 0.1, 'num_beams': 8, 'early_stopping': True}
  

  --info is an additional argument available for the same purpose of tracking the conditions of the experiment. It's useful to pass things that weren't in the argument list, e.g. a language pair --info "lang:en-ru". But also if you pass --info without a value it will fallback to the current date/time string, e.g. 2020-09-13 18:44:43.

  If using --dump-args --info, the output will be:

  {'bleu': 26.887, 'n_obs': 10, 'runtime': 1, 'seconds_per_sample': 0.1, 'num_beams': 8, 'early_stopping': True, 'info': '2020-09-13 18:44:43'}
  

  If using --dump-args --info "pair:en-ru chkpt=best, the output will be:

  {'bleu': 26.887, 'n_obs': 10, 'runtime': 1, 'seconds_per_sample': 0.1, 'num_beams': 8, 'early_stopping': True, 'info': 'pair=en-ru chkpt=best'}
  

• if you need to perform a parametric search in order to find the best ones that lead to the highest BLEU score, let run_eval_search.py to do the searching for you.

  The script accepts the exact same arguments as run_eval.py, plus an additional argument --search. The value of --search is parsed, reformatted and fed to run_eval.py as additional args.

  The format for the --search value is a simple string with hparams and colon separated values to try, e.g.:

   --search "num_beams=5:10 length_penalty=0.8:1.0:1.2 early_stopping=true:false"
  

  which will generate 12 (2*3*2) searches for a product of each hparam. For example the example that was just used will invoke run_eval.py repeatedly with:

   --num_beams 5 --length_penalty 0.8 --early_stopping true
   --num_beams 5 --length_penalty 0.8 --early_stopping false
   [...]
   --num_beams 10 --length_penalty 1.2 --early_stopping false
  

  On completion, this function prints a markdown table of the results sorted by the best BLEU score and the winning arguments.

bleu  | num_beams | length_penalty | early_stopping
----- | --------- | -------------- | --------------
26.71 |         5 |            1.1 |              1
26.66 |         5 |            0.9 |              1
26.66 |         5 |            0.9 |              0
26.41 |         5 |            1.1 |              0
21.94 |         1 |            0.9 |              1
21.94 |         1 |            0.9 |              0
21.94 |         1 |            1.1 |              1
21.94 |         1 |            1.1 |              0

Best score args:
stas/wmt19-en-ru data/en-ru/val.source data/en-ru/test_translations.txt --reference_path data/en-ru/val.target --score_path data/en-ru/test_bleu.json --bs 8 --task translation --num_beams 5 --length_penalty 1.1 --early_stopping True

If you pass --info "some experiment-specific info" it will get printed before the results table - this is useful for scripting and multiple runs, so one can tell the different sets of results from each other.

DistilBART

For the CNN/DailyMail dataset, (relatively longer, more extractive summaries), we found a simple technique that works: you just copy alternating layers from bart-large-cnn and finetune more on the same data.

For the XSUM dataset, that didn’t work as well so we used that same initialization strategy followed by a combination of Distillbert’s ce_loss and the hidden states MSE loss used in the tinybert paper.

You can see the performance tradeoffs of model sizes here. and more granular timing results here.

No Teacher Distillation

To run the simpler distilbart-cnn style distillation all you need is data, a GPU, and a properly initialized student. You don't even need distillation.py.

Some un-finetuned students are available for replication purposes. They are initialized by copying layers from the associated bart-large-{cnn|xsum} teacher using --init_strategy alternate. (You can read about that in initialization_utils.py) The command that produced sshleifer/distilbart-cnn-12-6 is

./train_distilbart_cnn.sh

runtime: 6H on NVIDIA RTX 24GB GPU

Note: You can get the same simple distillation logic by using ./run_distiller.sh --no_teacher followed by identical arguments as the ones in train_distilbart_cnn.sh. If you are using wandb and comparing the two distillation methods, using this entry point will make your logs consistent, because you will have the same hyperparameters logged in every run.

With a teacher (Intermediate Supervision)

Note only BART variants are supported

In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using BartSummarizationDistiller. This is how sshleifer/distilbart-xsum* checkpoints were produced.

The command that produced sshleifer/distilbart-xsum-12-6 is:

./train_distilbart_xsum.sh --logger_name wandb --gpus 1

runtime: 13H on V-100 16GB GPU.

Contributing

• follow the standard contributing guidelines and code of conduct.
• add tests to test_seq2seq_examples.py
• To run only the seq2seq tests, you must be in the root of the repository and run:

pytest examples/seq2seq/

Converting pytorch-lightning checkpoints

pytorch lightning -do_predict often fails, after you are done training, the best way to evaluate your model is to convert it.

This should be done for you, with a file called {save_dir}/best_tfmr.

If that file doesn't exist but you have a lightning .ckpt file, you can run

python convert_pl_checkpoint_to_hf.py PATH_TO_CKPT  randomly_initialized_hf_model_path save_dir/best_tfmr

Then either run_eval or run_distributed_eval with save_dir/best_tfmr (see previous sections)

Experimental Features

These features are harder to use and not always useful.

Dynamic Batch Size for MT

finetune.py has a command line arg --max_tokens_per_batch that allows batches to be dynamically sized. This feature can only be used:

• with fairseq installed
• on 1 GPU
• without sortish sampler
• after calling ./save_len_file.py $tok $data_dir

For example,

./save_len_file.py Helsinki-NLP/opus-mt-en-ro  wmt_en_ro
./dynamic_bs_example.sh --max_tokens_per_batch=2000 --output_dir benchmark_dynamic_bs

splits wmt_en_ro/train into 11,197 uneven lengthed batches and can finish 1 epoch in 8 minutes on a v100.

For comparison,

./dynamic_bs_example.sh --sortish_sampler --train_batch_size 48

uses 12,723 batches of length 48 and takes slightly more time 9.5 minutes.

The feature is still experimental, because:

• we can make it much more robust if we have memory mapped/preprocessed datasets.
• The speedup over sortish sampler is not that large at the moment.