ShuffleNet v2

This is an implementation of ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design .

You can download ImageNet trained checkpoints from here.

How to use the pretrained models

You only need two things:

1. File architecture.py. It contains a definition of the graph.
2. Checkpoint. You can load it into the graph using tf.train.Saver or tf.train.init_from_checkpoint.

For an example of using the pretrained model see: inference_with_trained_model.ipynb.

Speed benchmarks

All measurements were done using batches of size 8, images of size 224x224, and NVIDIA GTX 1080 Ti.
See benchmark_speed.ipynb for the code.

MobileNet v1 results are taken from here. MobileNet v2 results are taken from here.

Notes

1. Using moving averages of weights doesn't increase accuracy for some reason.

Requirements

1. for using the pretrained models: tensorflow 1.10
2. for dataset preparation: pandas, Pillow, tqdm, opencv, ...

How to train

1. Prepare ImageNet. See data/README.md.
2. Set the right parameters in the beginning of train.py file.
3. Run python train.py.
4. Run tensorboard to see the loss curves. Examples of loss curves are in images/.
5. Use evaluation.ipynb for the final evaluation on ImageNet.

Credit

The training code is heavily inspired by:

1. https://github.com/tensorflow/models/tree/master/official/resnet
2. https://cloud.google.com/tpu/docs/inception-v3-advanced

Other implementations

1. miaow1988/ShuffleNet_V2_pytorch_caffe
2. tensorpack/examples/ImageNetModels