|
--- |
|
language: en |
|
tags: |
|
- tapex |
|
license: apache-2.0 |
|
datasets: |
|
- tab_fact |
|
inference: false |
|
--- |
|
|
|
TAPEX-large model fine-tuned on WTQ. This model was proposed in [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou. Original repo can be found [here](https://github.com/microsoft/Table-Pretraining). |
|
|
|
To load it and run inference, you can do the following: |
|
|
|
``` |
|
from transformers import BartTokenizer, BartForSequenceClassification |
|
import pandas as pd |
|
|
|
tokenizer = BartTokenizer.from_pretrained("nielsr/tapex-large-finetuned-tabfact") |
|
model = BartForSequenceClassification.from_pretrained("nielsr/tapex-large-finetuned-tabfact") |
|
|
|
# create table |
|
data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]} |
|
table = pd.DataFrame.from_dict(data) |
|
|
|
# turn into dict |
|
table_dict = {"header": list(table.columns), "rows": [list(row.values) for i,row in table.iterrows()]} |
|
|
|
# turn into format TAPEX expects |
|
# define the linearizer based on this code: https://github.com/microsoft/Table-Pretraining/blob/main/tapex/processor/table_linearize.py |
|
linearizer = IndexedRowTableLinearize() |
|
linear_table = linearizer.process_table(table_dict) |
|
|
|
# add sentence |
|
sentence = "George Clooney has 69 movies" |
|
joint_input = sentence + " " + linear_table |
|
|
|
# encode |
|
encoding = tokenizer(joint_input, return_tensors="pt") |
|
|
|
# forward pass |
|
outputs = model(**encoding) |
|
|
|
# print prediction |
|
logits = outputs.logits |
|
print(logits.argmax(-1)) |
|
``` |
