Non-overlapped Patch-aware training and inference

SKKU-ESLAB / Auto-Compression

Automatic DNN compression tool with various model compression and neural architecture search techniques

MIT License

20 stars 20 forks source link

Closed nugabom closed 8 months ago

nugabom commented 8 months ago

Non-overlapped patch-aware training and inference implementation

Training
- Patch-aware convolution module
- Patch-aware convolution training code
Inference
- Zero skipping 3x3 convolution
- Zero spkiing 3x3 depthwise convolution (stride=1, 2)