nielsr HF staff commited on
Commit
dbfa150
1 Parent(s): 1eb4cc5

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +66 -65
README.md CHANGED
@@ -1,66 +1,67 @@
1
- ---
2
- language: en
3
- tags:
4
- - tapex
5
- license: mit
6
- ---
7
-
8
- # TAPEX (large-sized model)
9
-
10
- TAPEX 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. The original repo can be found [here](https://github.com/microsoft/Table-Pretraining).
11
-
12
- ## Model description
13
-
14
- TAPEX (**Ta**ble **P**re-training via **Ex**ecution) 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.
15
-
16
- 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.
17
-
18
- ## Intended Uses
19
-
20
- You can use the raw model for simulating neural SQL execution, i.e., employ TAPEX to execute a SQL query on a given table. However, the model is mostly meant to be fine-tuned on a supervised dataset. Currently TAPEX can be fine-tuned to tackle table question answering tasks and table fact verification tasks. See the [model hub](https://huggingface.co/models?search=tapex) to look for fine-tuned versions on a task that interests you.
21
-
22
- ### How to Use
23
-
24
- Here is how to use this model in transformers:
25
-
26
- ```python
27
- from transformers import TapexTokenizer, BartForConditionalGeneration
28
- import pandas as pd
29
-
30
- tokenizer = TapexTokenizer.from_pretrained("microsoft/tapex-large-sql-execution")
31
- model = BartForConditionalGeneration.from_pretrained("microsoft/tapex-large-sql-execution")
32
-
33
- data = {
34
- "year": [1896, 1900, 1904, 2004, 2008, 2012],
35
- "city": ["athens", "paris", "st. louis", "athens", "beijing", "london"]
36
- }
37
- table = pd.DataFrame.from_dict(data)
38
-
39
- # tapex accepts uncased input since it is pre-trained on the uncased corpus
40
- query = "select year where city = beijing"
41
- encoding = tokenizer(table=table, query=query, return_tensors="pt")
42
-
43
- outputs = model.generate(**encoding)
44
-
45
- print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
46
- # ['2008']
47
- ```
48
-
49
- ### How to Fine-tuning
50
-
51
- ⚠️ This model checkpoint is **ONLY** used for simulating neural SQL execution (i.e., employ TAPEX to execute a SQL query on a given table), and you **CANNOT** use this model for fine-tuning on downstream tasks. The one that can be used for fine-tuning is at [here](https://huggingface.co/microsoft/tapex-large).
52
-
53
- > This separation of two models for two kinds of intention is because of a known issue in BART large, and we recommend readers to see [this comment](https://github.com/huggingface/transformers/issues/15559#issuecomment-1062880564) for more details.
54
-
55
- ### BibTeX entry and citation info
56
-
57
- ```bibtex
58
- @inproceedings{
59
- liu2022tapex,
60
- title={{TAPEX}: Table Pre-training via Learning a Neural {SQL} Executor},
61
- author={Qian Liu and Bei Chen and Jiaqi Guo and Morteza Ziyadi and Zeqi Lin and Weizhu Chen and Jian-Guang Lou},
62
- booktitle={International Conference on Learning Representations},
63
- year={2022},
64
- url={https://openreview.net/forum?id=O50443AsCP}
65
- }
 
66
  ```
 
1
+ ---
2
+ language: en
3
+ tags:
4
+ - tapex
5
+ - table-question-answering
6
+ license: mit
7
+ ---
8
+
9
+ # TAPEX (large-sized model)
10
+
11
+ TAPEX 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. The original repo can be found [here](https://github.com/microsoft/Table-Pretraining).
12
+
13
+ ## Model description
14
+
15
+ TAPEX (**Ta**ble **P**re-training via **Ex**ecution) 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.
16
+
17
+ 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.
18
+
19
+ ## Intended Uses
20
+
21
+ You can use the raw model for simulating neural SQL execution, i.e., employ TAPEX to execute a SQL query on a given table. However, the model is mostly meant to be fine-tuned on a supervised dataset. Currently TAPEX can be fine-tuned to tackle table question answering tasks and table fact verification tasks. See the [model hub](https://huggingface.co/models?search=tapex) to look for fine-tuned versions on a task that interests you.
22
+
23
+ ### How to Use
24
+
25
+ Here is how to use this model in transformers:
26
+
27
+ ```python
28
+ from transformers import TapexTokenizer, BartForConditionalGeneration
29
+ import pandas as pd
30
+
31
+ tokenizer = TapexTokenizer.from_pretrained("microsoft/tapex-large-sql-execution")
32
+ model = BartForConditionalGeneration.from_pretrained("microsoft/tapex-large-sql-execution")
33
+
34
+ data = {
35
+ "year": [1896, 1900, 1904, 2004, 2008, 2012],
36
+ "city": ["athens", "paris", "st. louis", "athens", "beijing", "london"]
37
+ }
38
+ table = pd.DataFrame.from_dict(data)
39
+
40
+ # tapex accepts uncased input since it is pre-trained on the uncased corpus
41
+ query = "select year where city = beijing"
42
+ encoding = tokenizer(table=table, query=query, return_tensors="pt")
43
+
44
+ outputs = model.generate(**encoding)
45
+
46
+ print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
47
+ # ['2008']
48
+ ```
49
+
50
+ ### How to Fine-tuning
51
+
52
+ ⚠️ This model checkpoint is **ONLY** used for simulating neural SQL execution (i.e., employ TAPEX to execute a SQL query on a given table), and you **CANNOT** use this model for fine-tuning on downstream tasks. The one that can be used for fine-tuning is at [here](https://huggingface.co/microsoft/tapex-large).
53
+
54
+ > This separation of two models for two kinds of intention is because of a known issue in BART large, and we recommend readers to see [this comment](https://github.com/huggingface/transformers/issues/15559#issuecomment-1062880564) for more details.
55
+
56
+ ### BibTeX entry and citation info
57
+
58
+ ```bibtex
59
+ @inproceedings{
60
+ liu2022tapex,
61
+ title={{TAPEX}: Table Pre-training via Learning a Neural {SQL} Executor},
62
+ author={Qian Liu and Bei Chen and Jiaqi Guo and Morteza Ziyadi and Zeqi Lin and Weizhu Chen and Jian-Guang Lou},
63
+ booktitle={International Conference on Learning Representations},
64
+ year={2022},
65
+ url={https://openreview.net/forum?id=O50443AsCP}
66
+ }
67
  ```