This implementation of bilinear upsampling is considerably faster than the native PyTorch one in half precision (fp16). It is also slightly faster for single precision (fp32). See the "Performance" section below.
- PyTorch 1.0.0+
- CUDA 10.0+
- GPU with compute capability 7.0+ (Tesla V100, Titan V, GeForce RTX 2070/2080/2080Ti, etc.)
Currently this implementation is equivalent to PyTorch's bilinear upsampling with align_corners=True. The case of align_corners=False hasn't been implemented yet.
$python setup.py install
The script test.py demonstrates use. The TL;DR is that the following calls are equivalent:
- PyTorch:
result = torch.nn.functional.interpolate(data, scale_factor=2, mode='bilinear', align_corners=True) - This library:
from bilinear_upsampling import Interpolation
interp = Interpolation()
result = interp(data, scale_factor=2)
Tensor dimensions: (128, 3, 224, 224)
Test env: Titan V GPU, Intel Core i7-7800X CPU @ 3.50GHz
fp16
| Direction | PyTorch | This Implementation |
|---|---|---|
| forward | 685 us | 482 us |
| backward | 15.11 ms | 4.17 ms |
fp32
| Direction | PyTorch | This Implementation |
|---|---|---|
| forward | 788 us | 629 us |
| backward | 1.92 ms | 1.49 ms |