StatAssist & GradBoost: A Study on Optimal INT8 Quantization-aware Training from Scratch
- CIFAR10
- CIFAR100
- Imagenet_1K (ILSVRC2015)
- AlexNet
- VGGNet
- ResNet
- ShuffleNetV2
- MobileNetV2
- MobileNetV3
- MobileNetV3_ReLU
- Train
Modify setting/train.json before run.
python train.py
- Test
Modify setting/evaluate.json before run.
python evaluate.py
- Measured using AMD Ryzen Threadripper 1950X 16-Core Processor, 4 thread & PyTorch 1.6.0
- fbgemm
| Model | Params | Flops | Size (FP) | Size(quantized) | Latency(FP) | Latency(quantized) | Rate(%) |
|---|---|---|---|---|---|---|---|
| resnet18 | 11.68 M | 1.82 B | 46.81 MB | 11.78 MB | 366ms | 245 ms | -33.06 |
| shufflenet_v2_x0_5 | 1.36 M | 43.65 M | 5.56 MB | 1.48 MB | 157 ms | 89 ms | -43.31 |
| shufflenet_v2_x1_0 | 2.27 M | 150.6 M | 9.24 MB | 2.44MB | 255 ms | 372 ms | +45.88 |
| mobilenet_v2_ReLU6 | 3.50 M | 320.24 M | 14.21 MB | 3.58 MB | 282 ms | 143 ms | -49.29 |
| mobilenet_v2_ReLU | 3.50 M | 320.24 M | 14.21 MB | 3.58 MB | 282 ms | 112 ms | -60.28 |
| mobilenet_v3_large_HS | 5.47 M | 233.68 M | 22.07 MB | 5.81 MB | 286 ms | 160 ms | -44.05 |
| mobilenet_v3_small_HS | 2.70 M | 64.93 M | 10.91 MB | 2.92 MB | 164 ms | 99 ms | -39.63 |
| mobilenet_v3_large_ReLU | 5.47 M | 233.68 M | 22.06 MB | 5.57 MB | 258 ms | 131 ms | -49.22 |
| mobilenet_v3_small_ReLU | 2.70 M | 64.93 M | 10.90 MB | 2.77 MB | 148 ms | 79 ms | -46.62 |