randl / mobilenetv3-pytorch

Implementation of MobileNetV3 in pytorch

MobileNetV3 in PyTorch

An implementation of MobileNetV3 in PyTorch. MobileNetV3 is an efficient convolutional neural network architecture for mobile devices. For more information check the paper: Searching for MobileNetV3

Usage

Clone the repo:

git clone https://github.com/Randl/MobileNetV3-pytorch
pip install -r requirements.txt

Use the model defined in MobileNetV3.py to run ImageNet example:

python3 -m torch.distributed.launch --nproc_per_node=8 imagenet.py --dataroot "/path/to/imagenet/" --sched clr -b 128 --seed 42 --world-size 8 --sync-bn```

To continue training from checkpoint

python imagenet.py --dataroot "/path/to/imagenet/" --resume "/path/to/checkpoint/folder"

Results

WIP

Classification Checkpoint	MACs (M)	Parameters (M)	Top-1 Accuracy	Top-5 Accuracy	Claimed top-1	Claimed top-5	Inference time
MobileNetV3 Large x1.0 224	219.80	5.481	73.53	91.14	75.2	-	~258ms
mobilenet_v2_1.0_224	300	3.47	72.10	90.48	71.8	91.0	~461ms

Inference time is for single 1080 ti per batch of 128.

You can test it with

python imagenet.py --dataroot "/path/to/imagenet/" --resume "results/mobilenetv3large-v1/model_best0.pth.tar" -e

Other implementations

• https://github.com/d-li14/mobilenetv3.pytorch : 73.152% top-1, with more FLOPs
• https://github.com/xiaolai-sqlai/mobilenetv3 : 75.45% top-1, even more FLOPs
• https://github.com/rwightman/gen-efficientnet-pytorch : 75.634% top-1, seems to be right FLOPs

Code used

• DropBlock implementation by miguelvr
• FLOPS calculator by warmspringwinds
• Utility function for divisibility by Tensorflow
• Squeeze-Excitation block by jonnedtc
• Custom cross-entropy by eladhoffer
• Shadow weights implementation by eladhoffer

mobilenetv3-pytorch's Issues

load_state_dict_from_url generates error

load_state_dict_from_url will get a web page not file. It will can't parse the downloaded file.