patil-suraj
commented
3 years ago cc @NielsRogge
Closed xiaopi-ouo closed 3 years ago
patil-suraj
commented
3 years ago cc @NielsRogge
xiaopi-ouo
commented
3 years ago It doesn't work because type_vocab_sizes=[3, 256, 256, 2, 256, 256, 10] which means the default max size of a table is 256*256. However, the tokenizer doesn't check it.
Strictly speaking, it's not a bug but I think in that kind of tasks, many tables' size can exceed the default value.
Maybe it's good to add codes to handle this problem.
NielsRogge
commented
3 years ago Ok I've investigated this a bit more. The problem here is that there's a column called "Area (acres, 1872)" in the table whose values cause the column_ranks token types to exceed the vocab size of 256. Deleting this column resolves the issue. Also, the table that you provide is actually (way) too large for TAPAS. Note that it only has a max sequence length of 512 tokens, and that in the original paper, only tables having at most 500 cells were considered.
To correctly truncate the table, you have to initialize TapasTokenizer as follows:
from transformers import TapasTokenizer
tokenizer = TapasTokenizer.from_pretrained("google/tapas-base-finetuned-tabfact", drop_rows_to_fit=True)and then encode the table + query as follows:
inputs = tokenizer(table=table, queries=queries,
padding="max_length",
truncation=True, return_tensors="pt")cc @LysandreJik
It's maybe a bit weird that people have to initialize TapasTokenizer with drop_rows_to_fit to True and set truncation to True when calling it.
xiaopi-ouo
commented
3 years ago Thank you for the reply. It's really helpful!
LysandreJik
commented
3 years ago You're right @NielsRogge; should we switch the drop_rows_to_fit to True by default given that no models can handle it set to False when overflowing?
NielsRogge
commented
3 years ago I guess it should only be set to True when calling the tokenizer with truncation set to True or to drop_rows_to_fit. If the user does not specify truncation, and the table is too large, then an error will be thrown as shown here.
LysandreJik
commented
3 years ago Fixed by #9507
Tanx-123
Environment info
transformersversion: 4.1.0.dev0Who can help
May @LysandreJik help?
Information
Model I am using (Bert, XLNet ...): TAPAS
The problem arises when using:
The tasks I am working on is:
To reproduce
Steps to reproduce the behavior:
data = json.loads(data) model = TapasModel.from_pretrained("google/tapas-base-finetuned-tabfact") tokenizer = TapasTokenizer.from_pretrained("google/tapas-base-finetuned-tabfact") table = pd.DataFrame(data['table_list'][1:], columns=data['table_list'][0]).astype(str) queries = [data['sentence_annotations'][0]['final_sentence']] inputs = tokenizer(table=table, queries=queries, padding="max_length", return_tensors="pt", max_length=512, truncation=True) input_ids = inputs['input_ids'] attention_mask = inputs['attention_mask'] token_type_ids = inputs['token_type_ids'] x = model(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
IndexError Traceback (most recent call last)