Pytorch implementation of Google's EfficientNet-lite. Provide imagenet pre-train models.
In EfficientNet-Lite, all SE modules are removed and all swish layers are replaced with ReLU6. It's more friendly for edge devices than EfficientNet-B series.
Model details:
| Model | Params | MAdds | Top1 Acc(Official) | Top1 Acc(This repo) | Top5 Acc |
|---|---|---|---|---|---|
| efficientnet-lite0 | 4.7M | 407M | 75.1% | 71.73% | 90.17% |
| efficientnet-lite1 | 5.4M | 631M | 76.7% | 74.71% | 92.01% |
| efficientnet-lite2 | 6.1M | 899M | 77.6% | 77.14% | 93.54% |
| efficientnet-lite3 | 8.2M | 1.44B | 79.8% | 78.91% | 94.37% |
| efficientnet-lite4 | 13.0M | 2.64B | 81.5% | 80.34% | 95.06% |
| Pre-train Model |
|---|
| efficientnet-lite0 Download Link |
| efficientnet-lite1 Download Link |
| efficientnet-lite2 Download Link |
| efficientnet-lite3 Download Link |
| efficientnet-lite4 Download Link |
python train.py --model_name efficientnet_lite0 --train_dir YOUR_TRAINDATASET_PATH --val_dir YOUR_VALDATASET_PATH
python train.py --eval --eval_resume YOUR_MODEL_PATH --model_name efficientnet_lite0 --train_dir YOUR_TRAINDATASET_PATH --val_dir YOUR_VALDATASET_PATH
eval reaults:
efficientnet_lite0
TEST Iter 0: loss = 2.100231, Top-1 err = 0.282700, Top-5 err = 0.098280, val_time = 120.648957
efficientnet_lite1
TEST Iter 0: loss = 2.076898, Top-1 err = 0.252940, Top-5 err = 0.079880, val_time = 126.869352
efficientnet_lite2
TEST Iter 0: loss = 1.929238, Top-1 err = 0.228660, Top-5 err = 0.064640, val_time = 142.668548
efficientnet_lite3
TEST Iter 0: loss = 1.782202, Top-1 err = 0.210920, Top-5 err = 0.056260, val_time = 147.359098
efficientnet_lite4
TEST Iter 0: loss = 1.714834, Top-1 err = 0.196580, Top-5 err = 0.049440, val_time = 158.336004