MobileNetV3 PyTorch implementation

This is an unofficial implementation of MobileNetV3 in PyTorch. Currently this repo contains the small and large versions of MobileNetV3, but I plan to also implement detection and segmentation extensions.

How this repo is organized

• net/ folder contains model definitions
• data/ folder is used when running training code
• models/ will store checkpoints & training log
• train_*.py runs training

How to use the models:

Models are found under net folder. You can load models as follows:

from net.mobilenetv3 import MobileNetV3Large, MobileNetV3Small

model_large = MobileNetV3Large(n_classes=1000) # Or use small

Train on CIFAR10

The script train_cifar10.py pulls the CIFAR10 dataset using torchvision datasets and trains a MobileNetV3 on it. Note that the dimension was upsampled to 224x224 in order to match the dimensions.

You can start the training by simply executing:

python train_cifar10.py

Performance

(WIP)

Train on ImageNet

To train the models on ImageNet, run train_imagenet.py. The script assumes that the imagenet dataset is placed under data/imagenet/ folder.

Performance

(WIP)

SSDLite (WIP)

Implemented in net/ssdlite.py

TODO:

• [ ] Training code for ImageNet
• [ ] Detection: SSDLite
• [ ] Segmentation: Lite R-ASPP