bart-base-japanese

This model is converted from the original Japanese BART Pretrained model released by Kyoto University.

Both the encoder and decoder outputs are identical to the original Fairseq model.

How to use the model

The input text should be tokenized by BartJapaneseTokenizer.

Tokenizer requirements:

• Juman++
• zenhan
• pyknp
• sentencepiece

Simple FillMaskPipeline

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline

model_name = "Formzu/bart-base-japanese"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

masked_text = "天気が<mask>から散歩しましょう。"

fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer)
out = fill_mask(masked_text)
print(out)
# [{'score': 0.19255658984184265, 'token': 1718, 'token_str': 'よく', 'sequence': '天気 が よく から 散歩 し ましょう 。'}, 
#  {'score': 0.14426815509796143, 'token': 5478, 'token_str': '良く', 'sequence': '天気 が 良く から 散歩 し ましょう 。'}, 
#  {'score': 0.05554169788956642, 'token': 6561, 'token_str': '悪い', 'sequence': '天気 が 悪い から 散歩 し ましょう 。'}, 
#  {'score': 0.05524599179625511, 'token': 3553, 'token_str': '良い', 'sequence': '天気 が 良い から 散歩 し ましょう 。'}, 
#  {'score': 0.03720080852508545, 'token': 1370, 'token_str': '良', 'sequence': '天気 が 良 から 散歩 し ましょう 。'}]

Text Generation

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

model_name = "Formzu/bart-base-japanese"
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

masked_text = "天気が<mask>から散歩しましょう。"

inp = tokenizer(masked_text, return_tensors='pt').to(device)

out = model.generate(**inp, num_beams=1, min_length=0, max_length=20, early_stopping=True,  no_repeat_ngram_size=2)
res = "".join(tokenizer.decode(out.squeeze(0).tolist(), skip_special_tokens=True).split(" "))
print(res)
# 天気がよくなってから散歩しましょう。天気のよく合っているところにいる

Framework versions

• Transformers 4.21.2
• Pytorch 1.12.1+cu116
• Tokenizers 0.12.1