Provides a pytorch implementation of sparse_image_warp and an example of spec_augment.
Original implementation comes from zcaceres.
But there are several errors in their implementation:
- In their implementation they didn't add boundary points to source and destination. The idea is to sample a point and its destination after warpping but we want to keep all boundary points the same. Since they didn't do this they need to add small noise instead of zeros to lhs matrix.
- cross_squared_distance_matrix
Some initial results of SpecAug on WSJ:
| System(#iters) | WER dev93 | WER eval92 |
|---|---|---|
| Transformer | 8.2 | 4.8 |
| Transformer+FM+TM | 7.3 | 4.6 |
| Transformer+FM+TM+TW | 7.0 | 5.0 |
| RNN | 8.6 | 5.1 |
| RNN+FM+TM | 7.8 | 5.2 |
| RNN+FM+TM+TW | 7.3 | 4.6 |
Example of time warpping(W changes from 10 to 160, 81 dimensional features):
Example of time warpping(10 sources to 10 destinations which indicates that speed changes 10 times):
