TAPEX (base-sized model)

TAPEX was proposed in TAPEX: Table Pre-training via Learning a Neural SQL Executor by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou. The original repo can be found here.

Model description

TAPEX (Table Pre-training via Execution) is a conceptually simple and empirically powerful pre-training approach to empower existing models with table reasoning skills. TAPEX realizes table pre-training by learning a neural SQL executor over a synthetic corpus, which is obtained by automatically synthesizing executable SQL queries.

TAPEX is based on the BART architecture, the transformer encoder-encoder (seq2seq) model with a bidirectional (BERT-like) encoder and an autoregressive (GPT-like) decoder.

This model is the tapex-base model fine-tuned on the WikiSQL dataset.

Intended Uses

You can use the model for table question answering on relatively simple questions. Some solveable questions are shown below (corresponding tables now shown):

Question	Answer
tell me what the notes are for south australia	no slogan on current series
what position does the player who played for butler cc (ks) play?	guard-forward
how many schools did player number 3 play at?	1.0
how many winning drivers in the kraco twin 125 (r2) race were there?	1.0
for the episode(s) aired in the u.s. on 4 april 2008, what were the names?	"bust a move" part one, "bust a move" part two

How to Use

Here is how to use this model in transformers:

from transformers import TapexTokenizer, BartForConditionalGeneration
import pandas as pd

tokenizer = TapexTokenizer.from_pretrained("microsoft/tapex-base-finetuned-wikisql")
model = BartForConditionalGeneration.from_pretrained("microsoft/tapex-base-finetuned-wikisql")

data = {
    "year": [1896, 1900, 1904, 2004, 2008, 2012],
    "city": ["athens", "paris", "st. louis", "athens", "beijing", "london"]
}
table = pd.DataFrame.from_dict(data)

# tapex accepts uncased input since it is pre-trained on the uncased corpus
query = "In which year did beijing host the Olympic Games?"
encoding = tokenizer(table=table, query=query, return_tensors="pt")

outputs = model.generate(**encoding)

print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
# [' 2008.0']