Shufflenet implementation (Forward-pass) in tensorflow based on https://arxiv.org/abs/1707.01083
Using pre-trained model from tensorpack model zoo (ShuffleNetV1-1x-g=8.npz). This model utilizes g=8
and has a BNReLU after the first conv2d
(Conv1) in Stage1, not mentioned in the paper but can be found by looking at the entries in the model.
TODO: Optimize self.batch_normalization
since this is taking ~100ms
~40ms
to compute in later stages.
TODO: Current FLOPS at 145.25M
. Check for possibility of optimization to ~140.0M
as stated in the paper.