roberta_base

This commit is contained in:
Karol Kaczmarek 2020-06-13 20:24:24 +02:00
parent ddce23e0d4
commit 049966f426
8 changed files with 428657 additions and 0 deletions

8
0-get-models.sh Executable file
View File

@ -0,0 +1,8 @@
#!/usr/bin/env bash
set -e
set -x
wget https://github.com/sdadas/polish-roberta/releases/download/models/roberta_base_fairseq.zip
unzip roberta_base_fairseq.zip -d roberta_base_fairseq

10
1-create-data.sh Executable file
View File

@ -0,0 +1,10 @@
#!/usr/bin/env bash
set -e
set -x
spm_encode --model=roberta_base_fairseq/sentencepiece.bpe.model < data/train/in.tsv > data/train.input0.spm
spm_encode --model=roberta_base_fairseq/sentencepiece.bpe.model < data/dev-0/in.tsv > data/dev.input.spm
cp data/dev-0/expected.tsv data/dev.label
cp data/train/expected.tsv data/train.label

18
2-preproc.sh Executable file
View File

@ -0,0 +1,18 @@
#!/usr/bin/env bash
set -e
set -x
fairseq-preprocess \
--only-source \
--trainpref "data/train.input0.spm" \
--validpref "data/dev.input0.spm" \
--destdir "data-bin/input0" \
--workers 4 --srcdict roberta_base_fairseq/dict.txt
fairseq-preprocess \
--only-source \
--trainpref "data/train.label" \
--validpref "data/dev.label" \
--destdir "data-bin/label" \
--workers 4

31
3-train.py Executable file
View File

@ -0,0 +1,31 @@
#!/usr/bin/env bash
TOTAL_NUM_UPDATES=1000000000000000 # 10 epochs through IMDB for bsz 32
WARMUP_UPDATES=216085 # 6 percent of the number of updates
LR=1e-05 # Peak LR for polynomial LR scheduler.
HEAD_NAME=hesaid # Custom name for the classification head.
NUM_CLASSES=2 # Number of classes for the classification task.
MAX_SENTENCES=35 # Batch size.
ROBERTA_PATH="roberta_base_fairseq/model.pt"
fairseq-train data-bin/ \
--restore-file $ROBERTA_PATH \
--max-positions 512 \
--max-sentences $MAX_SENTENCES \
--max-tokens 8192 \
--task sentence_prediction \
--reset-optimizer --reset-dataloader --reset-meters \
--required-batch-size-multiple 2 \
--init-token 0 --separator-token 2 \
--arch roberta_base \
--criterion sentence_prediction \
--classification-head-name $HEAD_NAME \
--num-classes $NUM_CLASSES \
--dropout 0.1 --attention-dropout 0.1 \
--weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \
--clip-norm 0.0 \
--lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
--max-epoch 5 --log-format tqdm --log-interval 1 --save-interval-updates 15000 --keep-interval-updates 5 --skip-invalid-size-inputs-valid-test \
--best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \
--find-unused-parameters \
--update-freq 1

52
4-eval.py Executable file
View File

@ -0,0 +1,52 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import torch
from fairseq.models.roberta import RobertaModel
from tqdm import tqdm
if __name__ == '__main__':
roberta = RobertaModel.from_pretrained(
model_name_or_path='checkpoints',
data_name_or_path='data-bin',
sentencepiece_vocab='roberta_base_fairseq/sentencepiece.bpe.model',
checkpoint_file='checkpoint_best.pt',
bpe='sentencepiece',
)
roberta.cuda()
roberta.eval()
max_seq = 512
batch_size = 5
pad_index = roberta.task.source_dictionary.pad()
for dir_test in ['dev-0', 'dev-1', 'test-A']:
lines = []
with open(f'data/{dir_test}/in.tsv', 'rt') as f:
for line in tqdm(f, desc=f'Reading {dir_test}'):
line = roberta.encode(line.rstrip('\n'))[:max_seq]
lines.append(line)
predictions = []
for i in tqdm(range(0, len(lines), batch_size), desc='Processing'):
batch_text = lines[i: i + batch_size]
# Get max length of batch
max_len = max([tokens.size(0) for tokens in batch_text])
# Create empty tensor with padding index
input_tensor = torch.LongTensor(len(batch_text), max_len).fill_(pad_index)
# Fill tensor with tokens
for i, tokens in enumerate(batch_text):
input_tensor[i][:tokens.size(0)] = tokens
with torch.no_grad():
raw_prediction = roberta.predict('hesaid', input_tensor)
# Get probability for second class (M class)
out_tensor = torch.exp(raw_prediction[:, 1])
for line_prediction in out_tensor:
# Get probability for first class
predictions.append(line_prediction.item())
with open(f'data/{dir_test}/out.tsv', 'wt') as fw:
fw.write('\n'.join([f'{p:.8f}' for p in predictions]))

137314
dev-0/out.tsv Normal file

File diff suppressed because it is too large Load Diff

156606
dev-1/out.tsv Normal file

File diff suppressed because it is too large Load Diff

134618
test-A/out.tsv Normal file

File diff suppressed because it is too large Load Diff