Reproduction of GhostNet as described in GhostNet: More Features from Cheap Operations on ILSVRC2012 benchmark with PyTorch framework.
Download the ImageNet dataset and move validation images to labeled subfolders. To do this, you can use the following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
- batch size 1024
- iterations 450,000
- learning rate 0.4 (on 8 gpus)
- weight decay 0.00004
- dropout rate 0.2
- label smoothing 0.1
Clone the repo:
git clone https://github.com/diaomin/PyTorch-implementation-of-GhostNet/
Train the model:
python train.py --train-dir=/path/to/train/folder/ --val-dir=/path/to/val/folder/ --model-size=1.0x
WIP
Performance:
Models | MACs (M) | Params (M) | Top-1 Acc | Top-5 Acc |
---|---|---|---|---|
GhostNet 0.5x 224 | 2.6 | 42 | 65.5 | 86.3 |
GhostNet 1.0x 224 | 5.2 | 141 | 72.5 | 91.0 |
Curves:
- from left to right: loss, top-1, top5
- blue for training and orange for validation
- model size of 0.5x (450,000 iters)
- model size of 1.0x (300,000 iters)
- To train ghostnet with more iterations
- To train ghostnet with model size 2.0x