cscn89 / shufflenet-v2-tensorflow

a lightweight convolutional neural network

ShuffleNet v2

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

model	accuracy	top 5 accuracy
0.5x	0.607	0.822
1.0x	0.688	0.886

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

model	accuracy	images/second
ShuffleNet v2 0.5x	0.607	3192
ShuffleNet v2 1.0x	0.689	2349
ShuffleNet v2 1.5x	-	1855
ShuffleNet v2 2.0x	-	1570
MobileNet v1 0.5x	0.633	3317
MobileNet v1 0.75x	0.684	2187
MobileNet v1 1.0x	0.709	1685
MobileNet v2 0.35x	0.603	2722
MobileNet v2 0.75x	0.698	1527
MobileNet v2 1.0x	0.718	1292

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