ptls.data_load.padded_batch.PaddedBatch

Input data is a raw feature formats. You can transform your transaction to correct format with ptls.data module. Common description or sequential data and used data formats are here Input data are covered in ptls.data_load.padded_batch.PaddedBatch class.

We can create PaddedBatch object manually for demo and test purposes.

x = PaddedBatch(
    payload={
        'mcc_code': torch.randint(1, 10, (3, 8)),
        'currency': torch.randint(1, 4, (3, 8)),
        'amount': torch.randn(3, 8) * 4 + 5,
    },
    length=torch.Tensor([2, 8, 5]).long()
)

Here x contains three features. Two are categorical and one is numerical:

• mcc_code is categorical with dictionary_size=10
• currency is categorical with dictionary_size=4
• amount is numerical with mean=5 and std=4

x contains 5 sequences with maximum_length=12. Real lengths of each sequence are [2, 8, 5].

We can access x content via PaddedBatch properties x.payload and x.seq_lens.

Real data have sequences are padded with zeros. We can imitate it with x.seq_len_mask. It returns tensor with 1 if a position inside corresponded seq_len and 0 if position outside. Let's check out example

>>> x.seq_len_mask
Out: 
tensor([[1, 1, 0, 0, 0, 0, 0, 0],
        [1, 1, 1, 1, 1, 1, 1, 1],
        [1, 1, 1, 1, 1, 0, 0, 0]])

There are 2, 8 and 5 valid tokens in lines.

More way of seq_len_mask usage are in PaddedBatch docstring.

We can recreate our x with modified content:

x = PaddedBatch({k: v * x.seq_len_mask for k, v in x.payload.items()}, x.seq_lens)

Now we can check x.payload and see features looks like real padded data:

>>> x.payload['mcc_code']
Out: 
tensor([[8, 1, 0, 0, 0, 0, 0, 0],
        [5, 5, 9, 9, 4, 9, 3, 1],
        [4, 2, 2, 3, 3, 0, 0, 0]])

All invalid tokens are replaced with zeros.

Generally, all layers respect PaddedBatch.seq_lens and no explicit zeroing of padded characters is required.

Classes

See docstrings for classes:

• ptls.data_load.padded_batch.PaddedBatch