Tensorflow implementation for 'LCNN: Lookup-based Convolutional Neural Network'

This also have an implementations multi-gpu training codes for various models, so you can train your own model faster and predict images faster with Lookup Convolutions.

[x] Achieve MNist, ILSVRC2012 Baseline

[x] Training Imagenet on Multiple node with multiple gpus

[x] Training Code - Lookup-based Convolution Layer

[x] Same training result as the original paper

[x] Inference Code - Optimized Dense Matrix Operation by Implementing Custom Tensorflow Operation

[] Fast inference speed as the original paper

• Naive Lookup Convolution Processed

• [] TODO : OpenBlas or Eigen Implementation

Custom Operation have been implemented for LCNN's lookup convolution.

Source codes in /ops, and it should be build before run the inference code.

(Recommend tensorflow build with '-mavx -msse4.1 -msse4.2' options)

$ cp {tf-lcnn}/ops/* {tensorflow}/tensorflow/core/user_ops/
$ bazel build --config opt --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0" //tensorflow/core/user_ops:sparse_conv2d.so
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:{tensorflow}/bazel-bin/tensorflow/core/user_ops/

As you can see below timeline, this custom lookup convolutional operation has very little weight in the whole time, when compared relatively with normal convolutional layer.

• LCNN-Fast Configuration
• 1 core, single thread : all tests has been proceeded under this condition.

Alexnet's Fully connected layer was replaced with convolutional layer. Codes will be optimized soon and inference times will be updated.

• LCNN-Fast
  • Dictionary Size : 3, 30, 30, 30, 30, 512, 512
  • Lambda : 0.3
• LCNN-Accurate
  • Dictionary Size : 3, 500, 500, 500, 30, 1024, 1024
  • Lambda : 0.4

For LCNN Model, Two versions of networks were trained for experiments.

• LCNN-Fast
  • Sparsity : 0.083, 0.034, 0.008, 0.013, 0.027, 0.001, 0.002
• LCNN-Accurate
  • Sparsity : 0.129, 0.040, 0.029, 0.034, 0.071, 0.006, 0.007

The original paper was not evaluated on MNIST, but the dataset was suitable for rapid experiments.

Tests are in progress. Below is a partial result, and it will be updated soon.

• LCNN-Fast
  • Dictionary Size : 3, 30, 30, 30, 30, 512, 512
• LCNN-Mid
• LCNN-Accurate
  • Dictionary Size : 3, 500, 500, 500, 30, 1024, 1024

TODO : More tests on Resnet and etcs.

The experimental results from the original paper are as follows.

This code is very experimental and have been helped a lot from various websites.

[1] LCNN: Lookup-based Convolutional Neural Network

[2] https://openresearch.ai/t/lcnn-lookup-based-convolutional-neural-network

[3] author's code : https://github.com/hessamb/lcnn/blob/master/layers/PooledSpatialConvolution.lua

[1] ImageNet Classification with Deep Convolutional Neural Networks

[2] imagenet training on alexnet : https://github.com/dontfollowmeimcrazy/imagenet

[3] https://github.com/mouradmourafiq/tensorflow-convolution-models

[4] https://github.com/hpssjellis/easy-tensorflow-on-cloud9/blob/master/aymericdamien-Examples/examples/alexnet.py

[1] https://www.tensorflow.org/extend/adding_an_op

[2] https://davidstutz.de/implementing-tensorflow-operations-in-c-including-gradients/

[3] https://github.com/tensorflow/tensorflow/blob/8eaf671025e8cd5358278f91f7e89e2fbbe6a26b/tensorflow/core/kernels/conv_ops.cc#L94

[4] https://github.com/tensorflow/tensorflow/blob/r1.3/tensorflow/python/ops/sparse_ops.py

[5] https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/ops/nn_ops.cc#L503

[6] tensorflow/tensorflow#2412

[1] https://www.tensorflow.org/install/install_sources

[2] https://github.com/cjweeks/tensorflow-cmake

[3] tensorflow/tensorflow#2412

[1] Distributed Tensorflow : https://www.tensorflow.org/deploy/distributed

[2] Distributed Tensorflow Example : https://github.com/tensorflow/models/tree/master/inception

[3] https://research.fb.com/publications/imagenet1kin1h/

[1] https://stackoverflow.com/questions/34293714/can-i-measure-the-execution-time-of-individual-operations-with-tensorflow/37774470#37774470

[2] https://github.com/ppwwyyxx/tensorpack

[3] https://github.com/sorki/python-mnist

[4] imgaug : https://github.com/aleju/imgaug