TEXT.build_vocab for two datasets

[antgr]

❓ Questions and Help

Description

Hi, I want to do multitask learning, and to use two datasets.
I want to create common vocabulary from the two datasets. For example from imdb and snli. How should I use TEXT.build_vocab to achieve this?

[zhangguanheng66]

@mttk Can we pass a pretrained vocab object to build a torchtext dataset?

[mttk]

You can build vocab from multiple datasets (the argument for TEXT.build_vocab can be a sequence of Dataset instances).
Illustratively:

TEXT = Field(...)
imdb_train, imdb_test = IMDB.splits(text_field=TEXT, ...)
snli_train, snli_valid, snli_test = SNLI.splits(text_field=TEXT, ...)
TEXT.build_vocab(imdb_train, snli_train)

AFAIK there is no way to pass a pretrained vocab to build a dataset without some hacking, but this should work in this case.

[antgr]

You mean that I could not use for example "Glove" etc. Right?
For example something like:

TEXT.build_vocab(imdb_train, snli_train
                 min_freq = MIN_FREQ,
                 vectors = "glove.6B.300d",
                 unk_init = torch.Tensor.normal_)

[mttk]

That code should work with GloVe. The Field has to be shared between the Datasets (as it is in the example above), and you will get a common vocab. The vector assignment is done after the vocabulary is determined.

Concretely, the word frequencies are computed here from all the Dataset instances: https://github.com/pytorch/text/blob/master/torchtext/data/field.py#L304
And after that point, the Vectors class doesn't know how many source datasets were used to obtain the frequencies.

[antgr]

OK thank you! And if i would like to use a custom dataset instead of, for example, imdb? How could I use there the TEXT? Can I use split (with text_field as argument) in a custom dataset as well?

[mttk]

You have to load that dataset by e.g. using TabularDataset and assigning TEXT as a field in the constructor, then you can use the builtin dataset.split method. Now you have your custom dataset splits loaded.

Then, you can pass any instance of Dataset as an argument to TEXT.build_vocab.