pytorch-mobilenet

Mobilenet converted from tensorflow. Training not supported yet.

Convert

Download tensorflow checkpoint point from https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md and https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/README.md

Then run:

python convert_v1.py --depth_multiplier 0.50 --image_size 224 --tensorflow_model mobilenet_v1_0.50_224.ckpt

This will generate a pth file which has the same name as the tensorflow checkpoint.

Similar to v2.

(There is currently a mismatch between official mobilenetv2 model code and the official pretrained weights. The last layer in the conv_defs didn't do depth mutliplyication when multiplier is less than 1 according to the weights.)

Preprocessing:

The preoprocessing follows the tensorflow incecption: transforms.Normalize([0.5, 0.5, 0.5],[0.5, 0.5, 0.5]).

The current problem

To match the behavior of tensorflow slim conv2d, we manually calculate the paddings, this may lead to slower speed.

Converted models

You can download the converted models from link.

Benchmarks

Missing.

Training(not supported).