TAPAS¶
Note
This is a recently introduced model so the API hasn’t been tested extensively. There may be some bugs or slight breaking changes to fix them in the future.
Overview¶
The TAPAS model was proposed in TAPAS: Weakly Supervised Table Parsing via Pre-training by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos. It’s a BERT-based model specifically designed (and pre-trained) for answering questions about tabular data. Compared to BERT, TAPAS uses relative position embeddings and has 7 token types that encode tabular structure. TAPAS is pre-trained on the masked language modeling (MLM) objective on a large dataset comprising millions of tables from English Wikipedia and corresponding texts. For question answering, TAPAS has 2 heads on top: a cell selection head and an aggregation head, for (optionally) performing aggregations (such as counting or summing) among selected cells. TAPAS has been fine-tuned on several datasets: SQA (Sequential Question Answering by Microsoft), WTQ (Wiki Table Questions by Stanford University) and WikiSQL (by Salesforce). It achieves state-of-the-art on both SQA and WTQ, while having comparable performance to SOTA on WikiSQL, with a much simpler architecture.
The abstract from the paper is the following:
Answering natural language questions over tables is usually seen as a semantic parsing task. To alleviate the collection cost of full logical forms, one popular approach focuses on weak supervision consisting of denotations instead of logical forms. However, training semantic parsers from weak supervision poses difficulties, and in addition, the generated logical forms are only used as an intermediate step prior to retrieving the denotation. In this paper, we present TAPAS, an approach to question answering over tables without generating logical forms. TAPAS trains from weak supervision, and predicts the denotation by selecting table cells and optionally applying a corresponding aggregation operator to such selection. TAPAS extends BERT’s architecture to encode tables as input, initializes from an effective joint pre-training of text segments and tables crawled from Wikipedia, and is trained end-to-end. We experiment with three different semantic parsing datasets, and find that TAPAS outperforms or rivals semantic parsing models by improving state-of-the-art accuracy on SQA from 55.1 to 67.2 and performing on par with the state-of-the-art on WIKISQL and WIKITQ, but with a simpler model architecture. We additionally find that transfer learning, which is trivial in our setting, from WIKISQL to WIKITQ, yields 48.7 accuracy, 4.2 points above the state-of-the-art.
In addition, the authors have further pre-trained TAPAS to recognize table entailment, by creating a balanced dataset of millions of automatically created training examples which are learned in an intermediate step prior to fine-tuning. The authors of TAPAS call this further pre-training intermediate pre-training (since TAPAS is first pre-trained on MLM, and then on another dataset). They found that intermediate pre-training further improves performance on SQA, achieving a new state-of-the-art as well as state-of-the-art on TabFact, a large-scale dataset with 16k Wikipedia tables for table entailment (a binary classification task). For more details, see their follow-up paper: Understanding tables with intermediate pre-training by Julian Martin Eisenschlos, Syrine Krichene and Thomas MĂĽller.
The original code can be found here.
Tips:
TAPAS is a model that uses relative position embeddings by default (restarting the position embeddings at every cell of the table). Note that this is something that was added after the publication of the original TAPAS paper. According to the authors, this usually results in a slightly better performance, and allows you to encode longer sequences without running out of embeddings. This is reflected in the
reset_position_index_per_cellparameter ofTapasConfig, which is set toTrueby default. The default versions of the models available in the model hub all use relative position embeddings. You can still use the ones with absolute position embeddings by passing in an additional argumentrevision="no_reset"when calling the.from_pretrained()method. Note that it’s usually advised to pad the inputs on the right rather than the left.TAPAS is based on BERT, so
TAPAS-basefor example corresponds to aBERT-basearchitecture. Of course, TAPAS-large will result in the best performance (the results reported in the paper are from TAPAS-large). Results of the various sized models are shown on the original Github repository.TAPAS has checkpoints fine-tuned on SQA, which are capable of answering questions related to a table in a conversational set-up. This means that you can ask follow-up questions such as “what is his age?” related to the previous question. Note that the forward pass of TAPAS is a bit different in case of a conversational set-up: in that case, you have to feed every table-question pair one by one to the model, such that the prev_labels token type ids can be overwritten by the predicted labels of the model to the previous question. See “Usage” section for more info.
TAPAS is similar to BERT and therefore relies on the masked language modeling (MLM) objective. It is therefore efficient at predicting masked tokens and at NLU in general, but is not optimal for text generation. Models trained with a causal language modeling (CLM) objective are better in that regard.
Usage: fine-tuning¶
Here we explain how you can fine-tune TapasForQuestionAnswering on your own dataset.
STEP 1: Choose one of the 3 ways in which you can use TAPAS - or experiment
Basically, there are 3 different ways in which one can fine-tune TapasForQuestionAnswering,
corresponding to the different datasets on which Tapas was fine-tuned:
SQA: if you’re interested in asking follow-up questions related to a table, in a conversational set-up. For example if you first ask “what’s the name of the first actor?” then you can ask a follow-up question such as “how old is he?”. Here, questions do not involve any aggregation (all questions are cell selection questions).
WTQ: if you’re not interested in asking questions in a conversational set-up, but rather just asking questions related to a table, which might involve aggregation, such as counting a number of rows, summing up cell values or averaging cell values. You can then for example ask “what’s the total number of goals Cristiano Ronaldo made in his career?”. This case is also called weak supervision, since the model itself must learn the appropriate aggregation operator (SUM/COUNT/AVERAGE/NONE) given only the answer to the question as supervision.
WikiSQL-supervised: this dataset is based on WikiSQL with the model being given the ground truth aggregation operator during training. This is also called strong supervision. Here, learning the appropriate aggregation operator is much easier.
To summarize:
Task |
Example dataset |
Description |
Conversational |
SQA |
Conversational, only cell selection questions |
Weak supervision for aggregation |
WTQ |
Questions might involve aggregation, and the model must learn this given only the answer as supervision |
Strong supervision for aggregation |
WikiSQL-supervised |
Questions might involve aggregation, and the model must learn this given the gold aggregation operator |
Initializing a model with a pre-trained base and randomly initialized classification heads from the model hub can be done as follows (be sure to have installed the torch-scatter dependency for your environment):
>>> from transformers import TapasConfig, TapasForQuestionAnswering
>>> # for example, the base sized model with default SQA configuration
>>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base')
>>> # or, the base sized model with WTQ configuration
>>> config = TapasConfig.from_pretrained('google/tapas-base-finetuned-wtq')
>>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base', config=config)
>>> # or, the base sized model with WikiSQL configuration
>>> config = TapasConfig('google-base-finetuned-wikisql-supervised')
>>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base', config=config)
Of course, you don’t necessarily have to follow one of these three ways in which TAPAS was fine-tuned. You can also
experiment by defining any hyperparameters you want when initializing TapasConfig, and then
create a TapasForQuestionAnswering based on that configuration. For example, if you have a
dataset that has both conversational questions and questions that might involve aggregation, then you can do it this
way. Here’s an example:
>>> from transformers import TapasConfig, TapasForQuestionAnswering
>>> # you can initialize the classification heads any way you want (see docs of TapasConfig)
>>> config = TapasConfig(num_aggregation_labels=3, average_logits_per_cell=True, select_one_column=False)
>>> # initializing the pre-trained base sized model with our custom classification heads
>>> model = TapasForQuestionAnswering.from_pretrained('google/tapas-base', config=config)
What you can also do is start from an already fine-tuned checkpoint. A note here is that the already fine-tuned checkpoint on WTQ has some issues due to the L2-loss which is somewhat brittle. See here for more info.
For a list of all pre-trained and fine-tuned TAPAS checkpoints available in the HuggingFace model hub, see here.
STEP 2: Prepare your data in the SQA format
Second, no matter what you picked above, you should prepare your dataset in the SQA format. This format is a TSV/CSV file with the following columns:
id: optional, id of the table-question pair, for bookkeeping purposes.annotator: optional, id of the person who annotated the table-question pair, for bookkeeping purposes.position: integer indicating if the question is the first, second, third,… related to the table. Only required in case of conversational setup (SQA). You don’t need this column in case you’re going for WTQ/WikiSQL-supervised.question: stringtable_file: string, name of a csv file containing the tabular dataanswer_coordinates: list of one or more tuples (each tuple being a cell coordinate, i.e. row, column pair that is part of the answer)answer_text: list of one or more strings (each string being a cell value that is part of the answer)aggregation_label: index of the aggregation operator. Only required in case of strong supervision for aggregation (the WikiSQL-supervised case)float_answer: the float answer to the question, if there is one (np.nan if there isn’t). Only required in case of weak supervision for aggregation (such as WTQ and WikiSQL)
The tables themselves should be present in a folder, each table being a separate csv file. Note that the authors of the
TAPAS algorithm used conversion scripts with some automated logic to convert the other datasets (WTQ, WikiSQL) into the
SQA format. The author explains this here. Interestingly, these conversion scripts
are not perfect (the answer_coordinates and float_answer fields are populated based on the answer_text),
meaning that WTQ and WikiSQL results could actually be improved.
STEP 3: Convert your data into PyTorch tensors using TapasTokenizer
Third, given that you’ve prepared your data in this TSV/CSV format (and corresponding CSV files containing the tabular
data), you can then use TapasTokenizer to convert table-question pairs into input_ids,
attention_mask, token_type_ids and so on. Again, based on which of the three cases you picked above,
TapasForQuestionAnswering requires different inputs to be fine-tuned:
Task |
Required inputs |
Conversational |
|
Weak supervision for aggregation |
|
Strong supervision for aggregation |
|
TapasTokenizer creates the labels, numeric_values and numeric_values_scale based on
the answer_coordinates and answer_text columns of the TSV file. The float_answer and aggregation_labels
are already in the TSV file of step 2. Here’s an example:
>>> from transformers import TapasTokenizer
>>> import pandas as pd
>>> model_name = 'google/tapas-base'
>>> tokenizer = TapasTokenizer.from_pretrained(model_name)
>>> data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]}
>>> queries = ["What is the name of the first actor?", "How many movies has George Clooney played in?", "What is the total number of movies?"]
>>> answer_coordinates = [[(0, 0)], [(2, 1)], [(0, 1), (1, 1), (2, 1)]]
>>> answer_text = [["Brad Pitt"], ["69"], ["209"]]
>>> table = pd.DataFrame.from_dict(data)
>>> inputs = tokenizer(table=table, queries=queries, answer_coordinates=answer_coordinates, answer_text=answer_text, padding='max_length', return_tensors='pt')
>>> inputs
{'input_ids': tensor([[ ... ]]), 'attention_mask': tensor([[...]]), 'token_type_ids': tensor([[[...]]]),
'numeric_values': tensor([[ ... ]]), 'numeric_values_scale: tensor([[ ... ]]), labels: tensor([[ ... ]])}
Note that TapasTokenizer expects the data of the table to be text-only. You can use
.astype(str) on a dataframe to turn it into text-only data. Of course, this only shows how to encode a single
training example. It is advised to create a PyTorch dataset and a corresponding dataloader:
>>> import torch
>>> import pandas as pd
>>> tsv_path = "your_path_to_the_tsv_file"
>>> table_csv_path = "your_path_to_a_directory_containing_all_csv_files"
>>> class TableDataset(torch.utils.data.Dataset):
... def __init__(self, data, tokenizer):
... self.data = data
... self.tokenizer = tokenizer
...
... def __getitem__(self, idx):
... item = data.iloc[idx]
... table = pd.read_csv(table_csv_path + item.table_file).astype(str) # be sure to make your table data text only
... encoding = self.tokenizer(table=table,
... queries=item.question,
... answer_coordinates=item.answer_coordinates,
... answer_text=item.answer_text,
... truncation=True,
... padding="max_length",
... return_tensors="pt"
... )
... # remove the batch dimension which the tokenizer adds by default
... encoding = {key: val.squeeze(0) for key, val in encoding.items()}
... # add the float_answer which is also required (weak supervision for aggregation case)
... encoding["float_answer"] = torch.tensor(item.float_answer)
... return encoding
...
... def __len__(self):
... return len(self.data)
>>> data = pd.read_csv(tsv_path, sep='\t')
>>> train_dataset = TableDataset(data, tokenizer)
>>> train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=32)
Note that here, we encode each table-question pair independently. This is fine as long as your dataset is not
conversational. In case your dataset involves conversational questions (such as in SQA), then you should first group
together the queries, answer_coordinates and answer_text per table (in the order of their position
index) and batch encode each table with its questions. This will make sure that the prev_labels token types (see
docs of TapasTokenizer) are set correctly. See this notebook
for more info.
STEP 4: Train (fine-tune) TapasForQuestionAnswering
You can then fine-tune TapasForQuestionAnswering using native PyTorch as follows (shown here for
the weak supervision for aggregation case):
>>> from transformers import TapasConfig, TapasForQuestionAnswering, AdamW
>>> # this is the default WTQ configuration
>>> config = TapasConfig(
... num_aggregation_labels = 4,
... use_answer_as_supervision = True,
... answer_loss_cutoff = 0.664694,
... cell_selection_preference = 0.207951,
... huber_loss_delta = 0.121194,
... init_cell_selection_weights_to_zero = True,
... select_one_column = True,
... allow_empty_column_selection = False,
... temperature = 0.0352513,
... )
>>> model = TapasForQuestionAnswering.from_pretrained("google/tapas-base", config=config)
>>> optimizer = AdamW(model.parameters(), lr=5e-5)
>>> for epoch in range(2): # loop over the dataset multiple times
... for idx, batch in enumerate(train_dataloader):
... # get the inputs;
... input_ids = batch["input_ids"]
... attention_mask = batch["attention_mask"]
... token_type_ids = batch["token_type_ids"]
... labels = batch["labels"]
... numeric_values = batch["numeric_values"]
... numeric_values_scale = batch["numeric_values_scale"]
... float_answer = batch["float_answer"]
... # zero the parameter gradients
... optimizer.zero_grad()
... # forward + backward + optimize
... outputs = model(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids,
... labels=labels, numeric_values=numeric_values, numeric_values_scale=numeric_values_scale,
... float_answer=float_answer)
... loss = outputs.loss
... loss.backward()
... optimizer.step()
Usage: inference¶
Here we explain how you can use TapasForQuestionAnswering for inference (i.e. making predictions
on new data). For inference, only input_ids, attention_mask and token_type_ids (which you can obtain using
TapasTokenizer) have to be provided to the model to obtain the logits. Next, you can use the
handy convert_logits_to_predictions method of TapasTokenizer to convert these into predicted
coordinates and optional aggregation indices.
However, note that inference is different depending on whether or not the setup is conversational. In a non-conversational set-up, inference can be done in parallel on all table-question pairs of a batch. Here’s an example of that:
>>> from transformers import TapasTokenizer, TapasForQuestionAnswering
>>> import pandas as pd
>>> model_name = 'google/tapas-base-finetuned-wtq'
>>> model = TapasForQuestionAnswering.from_pretrained(model_name)
>>> tokenizer = TapasTokenizer.from_pretrained(model_name)
>>> data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]}
>>> queries = ["What is the name of the first actor?", "How many movies has George Clooney played in?", "What is the total number of movies?"]
>>> table = pd.DataFrame.from_dict(data)
>>> inputs = tokenizer(table=table, queries=queries, padding='max_length', return_tensors="pt")
>>> outputs = model(**inputs)
>>> predicted_answer_coordinates, predicted_aggregation_indices = tokenizer.convert_logits_to_predictions(
... inputs,
... outputs.logits.detach(),
... outputs.logits_aggregation.detach()
... )
>>> # let's print out the results:
>>> id2aggregation = {0: "NONE", 1: "SUM", 2: "AVERAGE", 3:"COUNT"}
>>> aggregation_predictions_string = [id2aggregation[x] for x in predicted_aggregation_indices]
>>> answers = []
>>> for coordinates in predicted_answer_coordinates:
... if len(coordinates) == 1:
... # only a single cell:
... answers.append(table.iat[coordinates[0]])
... else:
... # multiple cells
... cell_values = []
... for coordinate in coordinates:
... cell_values.append(table.iat[coordinate])
... answers.append(", ".join(cell_values))
>>> display(table)
>>> print("")
>>> for query, answer, predicted_agg in zip(queries, answers, aggregation_predictions_string):
... print(query)
... if predicted_agg == "NONE":
... print("Predicted answer: " + answer)
... else:
... print("Predicted answer: " + predicted_agg + " > " + answer)
What is the name of the first actor?
Predicted answer: Brad Pitt
How many movies has George Clooney played in?
Predicted answer: COUNT > 69
What is the total number of movies?
Predicted answer: SUM > 87, 53, 69
In case of a conversational set-up, then each table-question pair must be provided sequentially to the model, such
that the prev_labels token types can be overwritten by the predicted labels of the previous table-question
pair. Again, more info can be found in this notebook.