Asym helps handle asymmetric pytorch data (e.g. those with variable length) easily. Aside from simple features like
- batching (with padding)
- unbatching (with unpadding)
, asym also provides sophisticated features such as
- grouping (explained in the following section)
Unbatching a batched data after applying a torch.nn.Module cannot be done without additional information, since it could be unclear which part of the tensors should be unpadded. Asym achieves this by requiring the user to write a shallow wrapper class of the module (inherit from AnnotatedModule) that annotates the input and output dimension types. This allows unpadding only those parts of the output tensors that correspond to the padded length index of the input tensors.
With the help of dimension type annotations, some unusual (but sometimes useful) padding methods can be easily denoted and used by users:
- CDimPadder : Pad such that "constant dimensions" (jointly) take a given tensor value at padded positions.
Suppose you have a vastly asymmetric set of data (Let's say some have sequence length 1000 while others have sequence length <10), and you want to pass them to a module (torch.nn.Module) and get the results of module.forward() application. Conventional ways to do that would include:
- Processing them one by one, in which case we cannot exploit devices'(e.g. GPU's) full capacity.
- Processing them in mini-batches, which may require gross amounts of padding if you plainly do so.
A solution for this dilemma is to group the set in such a way that each group contains only the data with similar lengths, process the resulting mini-batches, and get the results back. Asym allows users to easily carry out this procedure with their custom Grouper classes.
See precompute_grouper_thresholds.py
Run
python3 examples.py
(later, simple examples will be here in README)
- A new length dimension cannot be introduced in an AnnotatedModule application. This means, for example, after applying a CNN layer that shrinks image width and height, data cannot be properly unbatched. (Since you don't know which part along the shrinked length dimension should be unpadded) This will be fixed soon.
This was made with the support of Deargen Inc.