Supervised learning

ptls.frames.supervised.SeqToTargetDataset and ptls.frames.supervised.SequenceToTarget for supervised learning

SeqToTargetDataset

Works similar as other datasets described in common patterns

Source data should have a scalar field with target value.

Example:

    dataset = SeqToTargetDataset([{
        'mcc_code': torch.randint(1, 10, (seq_len,)),
        'amount': torch.randn(seq_len),
        'event_time': torch.arange(seq_len),  # shows order between transactions
        'target': target,
    } for seq_len, target in zip(
        torch.randint(100, 200, (4,)),
        [0, 0, 1, 1],
    )], target_col_name='target')
    dl = torch.utils.data.DataLoader(dataset, batch_size=10, collate_fn=dataset.collate_fn)
    x, y = next(iter(dl))
    torch.testing.assert_close(y, torch.LongTensor([0, 0, 1, 1]))

SequenceToTarget in supervised mode

SequenceToTarget is a lightning module for supervised training. This module assumes a target for sequence.

There can be a some types of supervised task, like classification of regression. SequenceToTarget parameters allows to fit this module to your task.

SequenceToTarget requires seq_encoder, head, loss and metrics.

Se en examples of usage in SequenceToTarget docstring.

Layers from seq_encoder, head can be randomly initialized or pretrained.

SequenceToTarget in inference mode

You may just provide pretrained seq_encoder to SequenceToTarget and use trainer.predict to get embeddings from pretrained seq_encoder.

Classes

See docstrings for classes.

• ptls.frames.supervised.SeqToTargetDataset
• ptls.frames.supervised.SequenceToTarget