52793fdbee | ||
---|---|---|
.. | ||
README.md | ||
synthesize_logic_corpus.py | ||
synthesize_math_corpus.py |
README.md
POET
This is the official repo for the paper Reasoning Like Program Executors.
Pre-training Corpus
You can find the pre-training SQL corpus from here, the pre-training Math corpus from here.
The pre-training SQL corpus can be synthesized following the same procedure as done in TAPEX with the expand_numbers_in_text
function below:
def expand_numbers_in_text(text, delim=" ", ignore_chars=[","], reverse_num=False):
number_pattern = r"[-+]?[.]?[\d]+(,\d+)*[\.]?\d*(?:[eE][-+]?\d+)?%?"
num_char_spans = [(m.start(0), m.end(0)) for m in re.finditer(number_pattern, text)]
if len(num_char_spans) == 0: return text
out_text = ""
last_e = -1
for i, (s, e) in enumerate(num_char_spans):
out_text += text[:s] if i == 0 else text[last_e:s]
num_str = delim.join([c for c in list(text[s:e]) if c not in ignore_chars])
out_text += num_str if not reverse_num else num_str[::-1]
last_e = e
out_text += text[last_e:] # append rest
return out_text
The pre-training Math corpus can be synthesized by the script synthesize_math_corpus.py. The pre-training Logic corpus can be synthesized by the script synthesize_logic_corpus.py.
For all BART-based experiments, we use the fairseq implementation, which means that you can prepare the dataset as the following format:
|- dataset
|- train.src
|- train.tgt
|- valid.src
|- valid.tgt
After necessary preprocessing (you can follow the official guide in fairseq machin translation task), you can use the following script to train the model:
fairseq-train dataset/bin/ \
--save-dir models \
--tensorboard-logdir tensorboard_logs \
--restore-file BART-large/model.pt \
--arch bart_large \
--task translation \
--maximize-best-checkpoint-metric \
--criterion label_smoothed_cross_entropy \
--source-lang src \
--target-lang tgt \
--label-smoothing 0.1 \
--max-tokens 1536 \
--validate-interval 50 \
--save-interval 50 \
--save-interval-updates 3001 \
--validate-interval-updates 3001 \
--keep-interval-updates 5 \
--update-freq 16 \
--warmup-updates 500 \
--max-update 20000 \
--total-num-update 20000 \
--required-batch-size-multiple 1 \
--dropout 0.1 \
--attention-dropout 0.1 \
--relu-dropout 0.0 \
--weight-decay 0.01 \
--optimizer adam \
--adam-betas "(0.9, 0.999)" \
--adam-eps 1e-08 \
--clip-norm 0.1 \
--lr-scheduler polynomial_decay \
--lr 3e-5 \
--ddp-backend no_c10d \
--num-workers 1 \
--reset-meters \
--reset-optimizer \
--reset-dataloader \
--share-all-embeddings \
--layernorm-embedding \
--share-decoder-input-output-embed \
--skip-invalid-size-inputs-valid-test \
--log-format json \
--log-interval 10 \
--patience 10 \
--keep-best-checkpoints 1 \
--report-accuracy \
--no-epoch-checkpoints \
--no-last-checkpoints \
--no-save-optimizer-state
Pre-trained Model Weights
You can find all the available POET model weights at Huggingface Hub.
For all these models, you can try to fine-tune them as the vanilla models. And these models are pre-trained on the following format of natural context
and sentence
:
[sentence] col : [natural context]
where [sentence]
is usually the question in the task, and [natural context]
is usually the passage in the task. Please refer to our paper for more details.