HarDNeXt-Pytorch
HarDNeXt: A Stage Receptive Field and Connectivity Aware Convolution Neural Network
HarDNeXt-MSEG for Medical Image Segmentation in 0.913 mDice / 127 FPS, 0.910 mDice / 150 FPS on Kvasir-SEG dataset
Image Classification (ImgaeNet 2012)
Medical Image Segmentation (Kvasir-SEG)
Performance on ImageNet2012 Validation Set
| Model Name | MAC(B) | FPS (2080 Ti)(FP32) | EPI | Top-1 Acc (FP32) |
|---|---|---|---|---|
| ResNet-34 | 3.67 | 347 | 0.435 | 73.3 |
| DenseNet-121 | 2.88 | 97 | 1.08 | 74.65 |
| HarDNet-39 | 2.12 | 245 | 0.487 | 74.4 |
| HarDNeXt-28 | 2.07 | 359 | 0.354 | 74.09 |
| HarDNeXt-32 | 2.11 | 324 | 0.397 | 74.5 |
| HarDNeXt-39 | 2.84 | 299 | 0.466 | 75.36 |
| Model Name | MAC(B) | FPS (2080 Ti)(FP32) | EPI | Top-1 Acc (FP32) |
|---|---|---|---|---|
| ResNet-50 | 4.12 | 257 | 0.643 | 76.15 |
| DenseNet-169 | 3.42 | 71 | 1.571 | 76 |
| HarDNet-68 | 4.26 | 149 | 0.836 | 76.5 |
| HarDNeXt-50 | 3.51 | 215 | 0.619 | 76.32 |
| Model Name | MAC(B) | FPS (2080 Ti)(FP32) | EPI | Top-1 Acc (FP32) |
|---|---|---|---|---|
| ResNet-101 | 7.84 | 141 | 1.207 | 77.3 |
| DenseNet-201 | 4.37 | 58 | 1.92 | 77.2 |
| ResNeXt-50 | 4.27 | 166 | 0.92 | 77.62 |
| HarDNeXt-56 | 6.32 | 182 | 0.89 | 77.3 |
Training
python ./src/main.py --model_name hardnext --arch 39 /imagenet/data/path
Evaluation
python ./src/main.py --model_name hardnext --arch 39 --evaluation --resume /pretrained weight/data/path
