The implementation of group convolution in CNTK has been updated. The updated implementation moves away from creating a sub-graph for group convolution (using slicing and splicing), and instead uses cuDNN7 and MKL2017 APIs directly. This improves the experience both in terms of performance and model size.
As an example, for a single group convolution op with the following attributes:
- Input tensor (C, H, W) = (32, 128, 128)
- Number of output channels = 32 (channel multiplier is 1)
- Groups = 32 (depth wise convolution)
- Kernel size = (5, 5)
The comparison numbers for this single node are as follows:
| First Header | GPU exec. time (in millisec., 1000 run avg.) | CPU exec. time (in millisec., 1000 run avg.) | Model Size (in KB, CNTK format) |
|---|---|---|---|
| Old implementation | 9.349 | 41.921 | 38 |
| New implementation | 6.581 | 9.963 | 5 |
| Speedup/savings Approx. | 30% Approx. | 65-75% Approx. | 87% |
There is a breaking change in the depth_to_space and space_to_depth operators. These have been updated to match ONNX specification, specifically the permutation for how the depth dimension is placed as blocks in the spatial dimensions, and vice-versa, has been changed. Please refer to the updated doc examples for these two ops to see the change.
There is a breaking change in the arguments property in CNTK python API. The default behavior has been updated to return arguments in python order instead of in C++ order. This way it will return arguments in the same order as they are fed into ops. If you wish to still get arguments in C++ order, you can simply override the global option. This change should only affect the following ops: Times, TransposeTimes, and Gemm(internal).
- Updated CNTK's ONNX BatchNormalization op export/import to latest spec.