Introducing BatchBlock #1192
Introducing BatchBlock #1192
Conversation
Documentation preview |
marcromeyn
force-pushed
the
torch/batch-block
branch
2 times, most recently
from
fc1d25f
to
1582d6a
Compare
marcromeyn
force-pushed
the
torch/batch-block
branch
from
1582d6a
to
69d17df
Compare
| A dictionary containing all the flattened features, targets, and sequences. | ||
| """ | ||
| flat_dict: Dict[str, torch.Tensor] = self._flatten() | ||
| dummy_tensor = torch.tensor(0) |
There was a problem hiding this comment.
Can you explain why we need the dummy_tensor variable?
There was a problem hiding this comment.
We never use it, but we need to have it in order to keep the type of the Dict[str, torch.Tensor]. We could store the original value but that might take more memory, so that's why I added the dummy. We only care about the keys of the original inputs.
| result = batch.flatten_as_dict(input_batch) | ||
| assert len(result) == 9 # input keys are considered | ||
| assert ( | ||
| len([k for k in result if k.startswith("inputs.")]) == 4 |
There was a problem hiding this comment.
what is the difference between inputs. and features. keys ?
There was a problem hiding this comment.
inputs is what went into the batch-transformation, used for the mechanism to restore values of the batch that were not transformed by any of the branches.
|
|
||
| def test_in_parallel(self): | ||
| feat, target = torch.tensor([1, 2]), torch.tensor([3, 4]) | ||
| outputs = module_utils.module_test( |
There was a problem hiding this comment.
+1 to how we can now create different preprocessing blocks for each subset of the inputs! I love it!!
Goals ⚽
BatchBlockwill be used inside theModelto create theBatchobject. It's also useful for things like masking & padding.