Implementation of AlexNet

Under Development

• ! ! ! It needs an evaluation on ImageNet

This project is an unofficial implementation of AlexNet, using C Program Language Without Any 3rd Library, according to the paper "ImageNet Classification with Deep Convolutional Neural Networks" by Alex Krizhevsky,et al.

Only support CPU now

Features

• Effective matrix multiply, w.r.t L1/L2 cache
• img2col implementation of convolutional layer
• Multi-thread CPU Apply MT in operator's forward computation and backward computation to boost speed
• Efficient Memory Usage & Minimal Memory Occupation Memory is allocated on demand for efficient memory usage. If an array isn't necessary for computations afterward, it's memory will be freed instantly to minimize memory occupation.
• Modular layer Define each layer seperately

Architecture

----------------------------------------------------------------
            Layer                Output Shape         Param #
================================================================
            Conv2d-1           [N, 64, 55, 55]          23,296
              ReLU-2           [N, 64, 55, 55]               0
         MaxPool2d-3           [N, 64, 27, 27]               0
            Conv2d-4          [N, 192, 27, 27]         307,392
              ReLU-5          [N, 192, 27, 27]               0
         MaxPool2d-6          [N, 192, 13, 13]               0
            Conv2d-7          [N, 384, 13, 13]         663,936
              ReLU-8          [N, 384, 13, 13]               0
            Conv2d-9          [N, 256, 13, 13]         884,992
             ReLU-10          [N, 256, 13, 13]               0
           Conv2d-11          [N, 256, 13, 13]         590,080
             ReLU-12          [N, 256, 13, 13]               0
        MaxPool2d-13            [N, 256, 6, 6]               0
          Dropout-14                 [N, 9216]               0
           Linear-15                 [N, 4096]      37,752,832
             ReLU-16                 [N, 4096]               0
          Dropout-17                 [N, 4096]               0
           Linear-18                 [N, 4096]      16,781,312
             ReLU-19                 [N, 4096]               0
           Linear-20                 [N, 1000]       4,097,000
================================================================
("N" stands for "batch size")
Total params: 61,100,840
Trainable params: 61,100,840
Non-trainable params: 0
----------------------------------------------------------------

Usage (Only for Linux now)

Install

git clone https://github.com/Dynmi/AlexNet.git
cd AlexNet
make clean && make all

Train

( The data loader only supports .jpeg and .png$ images now. For image dataset, go to http://www.image-net.org/)

1. Create file images.list in the directory ./, each line contains info of one image, like this: class_id image_path .

For example:

0 /home/haris/Documents/AlexNet/images/0/1.jpeg
1 /home/haris/Documents/AlexNet/images/1/1.jpeg
2 /home/haris/Documents/AlexNet/images/2/1.jpeg
3 /home/haris/Documents/AlexNet/images/3/1.jpeg
4 /home/haris/Documents/AlexNet/images/4/1.jpeg
5 /home/haris/Documents/AlexNet/images/5/1.jpeg

2. Run the command for training

./alexnet train -batchsize <batch-size> -epochs <epochs> -load_pretrained <weights-path> -save <weights-path>

For example:

./alexnet train -batchsize 8 -epochs 10 -load_pretrained ./alexnet_pretrained.weights -save ./temp.weights 

Inference

$./alexnet inference -input <image-path> -load <weights-path>

For example:

./alexnet inference -input ./0001.jpeg -load ./alexnet_pretrained.weights

Todo List

• Try & Compare on ImageNet

• CUDA speed boosting

Speed Benchmark for Ops

Experiments were done on a laptop --- Huawei MagicBook14

Conv2D Forward

	this	DarkNet
Scene1	0.13s	0.44s
Scene2	0.21s	0.66s

Scene1: Input shape [4,224,224,3], weight shape [64,3,11,11]

Scene2: Input shape [4,57,57,128], weight shape [256,128,3,3]

Full-connected Forward

	this	DarkNet
Scene1	0.07s	0.24s
Scene2	0.11s	0.52s

Scene1: Input shape [4,2048], weight shape [2048,1024]

Scene2: Input shape [4,4096], weight shape [4096,4096]

Original AlexNet Details

divide the learning rate by 10 when the validation error rate stopped improving with the current learning rate. The learning rate was initialized at 0.01 and reduced three times prior to termination. We trained the network for roughly 90 cycles through the training set of 1.2 million images.

Reference

• AlexNet, NIPS2012

• Darknet

• PyTorch