An implementation of
ShuffleNet introduced in TensorFlow. According to the authors,
ShuffleNet is a computationally efficient CNN architecture designed specifically for mobile devices with very limited computing power. It outperforms
Google MobileNet by
small error percentage at much lower FLOPs.
Link to the original paper: ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
The paper uses the group convolution operator. However, that operator is not implemented in TensorFlow backend. So, I implemented the operator using graph operations.
This issue was discussed here: Support Channel groups in convolutional layers #10482
Reshaping the input tensor from (N, H, W, C) into (N, H, W, G, C').
Performing matrix transpose operation on the two dimensions (G, C').
Reshaping the tensor back into (N, H, W, C).
N: Batch size, H: Feature map height, W: Feature map width, C: Number of channels, G: Number of groups, C': Number of channels / Number of groups
Note that: The number of channels should be divisible by the number of groups.
Python 3 or above tensorflow 1.3.0 numpy 1.13.1 tqdm 4.15.0 easydict 1.7 matplotlib 2.0.2
python main.py --config config/test.json
The model have successfully overfitted TinyImageNet-200 that was presented in CS231n - Convolutional Neural Networks for Visual Recognition. I'm working on ImageNet training..
The paper has achieved 140 MFLOPs using the vanilla version. Using the group convolution operator implemented in TensorFlow, I have achieved approximately 270 MFLOPs. The paper counts multiplication+addition as one unit, so roughly dividing 270 by two, I have achieved what the paper proposes.
To calculate the FLOPs in TensorFlow, make sure to set the batch size equal to 1, and execute the following line when the model is loaded into memory.
tf.profiler.profile( tf.get_default_graph(), options=tf.profiler.ProfileOptionBuilder.float_operation(), cmd='scope')
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.