| Model | Acc@1 | Acc@5 | #Params | FLOPs | Inf Time* |
|---|---|---|---|---|---|
| ResNet-50 | 76.40 | 93.15 | 25.6M | 4.12G | 1.41ms |
| LIP-ResNet-50 | 78.19 | 93.96 | 23.9M | 5.33G | 1.88ms |
| +1.79 | +0.81 | -6.6% | +29.4% | +33% | |
| ResNet-101 | 77.98 | 93.98 | 44.5M | 7.85G | 2.29ms |
| LIP-ResNet-101 | 79.33 | 94.60 | 42.9M | 9.06G | 2.77ms |
| +1.46 | +0.62 | -3.6% | +15.4% | +21% | |
| DenseNet-121 | 75.62 | 92.56 | 8.0M | 2.88G | 1.49ms |
| LIP-DenseNet-121 | 76.64 | 93.16 | 8.7M | 4.13G | 1.80ms |
| +1.02 | +0.60 | +8.8% | +43.4% | +21% |
* average inference time of a single image. The results are calculated by repeated inference with batch size 32 on a single Titan XP card.
** LIP models here denote the full LIP architectures with Bottleneck-128 logit modules.
Our training and testing code main.py is modified from PyTorch official example one. You can refer to this for preparing ImageNet dataset and dependencies.
The code main.py is specialized in single node, multiple GPUs training for the faster speed. You can configure settings in train-lip_resnet50 in Makefile and then train LIP-ResNet-50 by simply
make train-lip_resnet50
You can resort to PyTorch official example one if the command above fails. For that, you need to modify the learning rate schedule to be consistent with the paper, i.e., lr decays 10x at the 30, 60, 80-th epoch.
Alike, you can evaluate the models by
make val-lip_resnet50
make val-lip_resnet101
make val-lip_densenet121
You can place our pretrained models in this directory and evaluate them. The results should be
LIP-ResNet-50
Epoch [0] * Acc@1 78.186 Acc@5 93.964
LIP-ResNet-101
Epoch [0] * Acc@1 79.330 Acc@5 94.602
LIP-DenseNet-121
Epoch [0] * Acc@1 76.636 Acc@5 93.156