mobileNet-v2_cifar10

a pytorch implement of mobileNet v2 on cifar10

architecture

• The origin mobileNet architecture is designed specifically for ImageNet where images' size is 224x224x3. To make it fit cifar10's size (32x32x3), I have disabled some downsample layer, i.e. replace the first few layers which have stride 2 with stride 1, as highlighted below.

  

training

• Run python3 train.py to start training

• Run python3 plot.py to show training curve

• Or you can run python3 train_ddp.py using ddp and cosine scheduler(final validacc = 94.71%)

• I have trained this model (width multiplier = 1, more setups can be seen in train.py) on two titan x, which takes about 6 hours, the weights and logs are available in folder bak

result

• The model can achieve max / mean accuracy 94.69% / 94.52% on validation set. Here the "mean accuracy" refers to mean of last 10 accuracy.

• accuracy & loss - iterations curves are shown below:

  

dependency

• python 3+
• pytorch 0.4.0