MobileNetV2

Implementation of MobileNetV2 with pyTorch, adapted from MobileNetV2-PyTorch and pytorch-mobilenet-v2.

Theory

 You can find the paper of MobileNetV2 at Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segment.

Usage

 This project uses Python 3.7.3 and PyTorch 1.0.1.

Prepare data

 The ImageNet dataset is used in this project and is put as follows (Copied from miraclewkf/MobileNetV2-PyTorch where you can find the files ILSVRC2012_img_train and ILSVRC2012_img_val).

├── train.py # train script
├── MobileNetV2.py # network of MobileNetV2
├── read_ImageNetData.py
├── ImageData
  ├── ILSVRC2012_img_train
    ├── n01440764
    ├──    ...
    ├── n15075141
  ├── ILSVRC2012_img_val
  ├── ILSVRC2012_dev_kit_t12
    ├── data
      ├── ILSVRC2012_validation_ground_truth.txt
      ├── meta.mat

Train

• Train from scratch:

CUDA_VISIBLE_DEVICES=0,1 python train.py --batch-size 128

• Train from one checkpoint(for example, train from epoch_200.pth, the --start-epoch parameter is corresponding to the epoch of the checkpoint):

CUDA_VISIBLE_DEVICES=2,3 python train.py --batch-size 256 --resume /media/data2/chenjiarong/MobileNetV2/output/epoch_200.pth --start-epoch 200 --num-epochs 300

Pretrained models

 In pretrained, achieving an accuracy of 71.62%.

Experiments

training setting:

1. number of epochs: 400
2. learning rate schedule: learning rate decay rate of 0.98 per epoch, initial lr=0.045
3. weight decay: 4e-5
4. remove dropout
5. batch size: 256
6. optimizer: SGD

MobileNetV2

	Madds	Parameters	Top1-acc
Offical 1.0	300 M	3.4 M	72.0%
Ours 1.0 (Madds&Parameters calculated by thop)	328.78 M	3.5 M	71.62%
Ours 1.0 (Madds&Parameters calculated by torchsummaryX)	300.79 M	3.5 M	71.62%