AIOSYM / dl-model-gallery

Implementation of popular Deep Learning model

Deep Learning Model Gallery

Implementation of various popular Deep Learning models and architectures.

LeNet

Paper: [Gradient-based learning applied to document recognition] (http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf) (1998)

LeNet or LeNet-5 (because it has 5 layers that contain parameters) is the most basic form of Convolutional Neural Network (CNN) architecture. It was first used for handwritten digit recognition (MNIST dataset)

Architecture Summary

Layer	Kernel Size	Number of Filter	Stride	Padding	Activation Function	Output Size
Input	32 x 32 x 3
Convolution	5 x 5	6	1	0	-	28 x 28 x 6
Maxpool	2 x 2	-	2	0	sigmoid	14 x 14 x 6
Convolution	5 x 5	16	1	0	-	10 x 10 x 16
Maxpool	2 x 2	-	2	0	sigmoid	5 x 5 x 16
Fully Connected	120	-	-	-	-	120
Fully Connected	84	-	-	-	-	84
Fully Connected	10 (classes)	-	-	-	softmax	10

Framework	PyTorch	TensorFlow
Implemented	✅	☑️

AlexNet

Paper: [ImageNet Classification with Deep Convolutional Neural Networks] (https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf) (2012)

AlexNet competed in the ImageNet Large Scale Visual Recognition Challenge and achieved a top-5 error of 15.3%. This model has made the name of Deep Learning as well known as today. Also, there is a special layer called LocalResponseNorm(LRN). However, there are some experiments that proof that LRN does not contribute much.

Architecture Summary

Image source

Note: If you read the original paper, the figure above will be a bit different. Because the author use 2 GPUs for training and the figure illustrate distribution of each layter to those GPUs.

Layer	Kernel Size	Number of Filter	Stride	Padding	Activation Function	Output Size
Input	227 x 227 x 3
Convolution	11 x 11	96	4	0	ReLU	55 x 55 x 96
LocalResponseNorm	-	-	-	-	-	55 x 55 x 96
MaxPooling	3 x 3	-	2	0	-	27 x 27 x 96
Convolution	5 x 5	256	1	1	ReLU	27 x 27 x 256
LocalResponseNorm	-	-	-	-	-	27 x 27 x 256
MaxPooling	3 x 3	-	2	0	-	13 x 13 x 256
Convolution	3 x 3	384	1	1	ReLU	13 x 13 x 384
Convolution	3 x 3	384	1	1	ReLU	13 x 13 x 384
Convolution	3 x 3	256	1	1	ReLU	13 x 13 x 256
MaxPooling	3 x 3	-	2	0	-	6 x 6 x 256
Fully Connected	4096	-	-	-	ReLU	4096
Fully Connected	4096	-	-	-	ReLU	4096
Fully Connected	2 (classes)	-	-	-	softmax	2

Framework	PyTorch	TensorFlow
Implemented	✅	☑️

VGG-16

Paper: [Very Deep Convolutional Networks for Large-Scale Image Recognition] (https://arxiv.org/pdf/1409.1556.pdf) (2014)

VGG-16 is a model proposed by the researchers at the University of Oxford. This model uses many convolution layers on top of each other to extract features. VGG-16 here derived from the fact that there are 13 convolution layers and 3 fully connected layers (13+3=16). You might also see VGG-19 as well. So refer to the paper for more detail of the architecture.

Architecture Summary

Layer	Kernel Size	Number of Filter	Stride	Padding	Activation Function	Output Size
Input	224 x 224 x 3
Convolution	3 x 3	64	1	1	ReLU	224 x 224 x 64
Convolution	3 x 3	64	1	1	ReLU	224 x 224 x 64
MaxPooling	2 x 2	-	2	0	-	112 x 112 x 64
Convolution	3 x 3	128	1	1	ReLU	112 x 112 x 128
Convolution	3 x 3	128	1	1	ReLU	112 x 112 x 128
MaxPooling	2 x 2	-	2	0	-	56 x 56 x 128
Convolution	3 x 3	256	1	1	ReLU	56 x 56 x 256
Convolution	3 x 3	256	1	1	ReLU	56 x 56 x 256
Convolution	3 x 3	256	1	1	ReLU	56 x 56 x 256
MaxPooling	2 x 2	-	2	0	-	28 x 28 x 256
Convolution	3 x 3	512	1	1	ReLU	28 x 28 x 512
Convolution	3 x 3	512	1	1	ReLU	28 x 28 x 512
Convolution	3 x 3	512	1	1	ReLU	28 x 28 x 512
MaxPooling	2 x 2	-	2	0	-	14 x 14 x 512
Convolution	3 x 3	512	1	1	ReLU	14 x 14 x 512
Convolution	3 x 3	512	1	1	ReLU	14 x 14 x 512
Convolution	3 x 3	512	1	1	ReLU	14 x 14 x 512
MaxPooling	2 x 2	-	2	0	-	7 x 7 x 512
Fully Connected	4096	-	-	-	ReLU	4096
Fully Connected	4096	-	-	-	ReLU	4096
Fully Connected	2 (classes)	-	-	-	softmax	2

Framework	PyTorch	TensorFlow
Implemented	✅	☑️

GoogLeNet

Paper: [Going Deeper with Convolutions] (https://arxiv.org/abs/1409.4842) (2014)

GoogLeNet is a 22 layers deep network and this name is honor to Yann LeCuns pioneering LeNet 5 network. It introduces a new module called Inception.

Architecture Summary

Layer	Kernel Size/Stride/Padding	#Filter	For inception module						Activation Function	output size
			#1x1	#3x3	#5x5	pooling	#3x3(reduce)	#5x5(reduce)
Input	224x224x3
Convolution	7x7 / 2 / 0	64	-	-	-	-	-	-	ReLU	112x112x64
MaxPool	3x3 / 2	-	-	-	-	-	-	-	-	56x56x64
LocalResponseNorm	-	-	-	-	-	-	-	-	-	56x56x64
Convolution	1x1 / 1 / 0	64	-	-	-	-	-	-	ReLU	56x56x64
Convolution	3x3 / 1 / 1	192	-	-	-	-	-	-	ReLU	56x56x192
LocalResponseNorm	-	-	-	-	-	-	-	-	-	56x56x192
MaxPool	3x3 / 2	-	-	-	-	-	-	-	-	28x28x192
Inception(3a)	-	-	64	128	32	32	96	16	-	28x28x256
Inception(3b)	-	-	128	192	96	64	128	32	-	28x28x480
MaxPool	3x3 / 2	-	-	-	-	-	-	-	-	14x14x480
Inception(4a)	-	-	192	208	48	64	96	16	-	14x14x512
Inception(4b)	-	-	160	224	64	64	112	24	-	14x14x512
Inception(4c)	-	-	128	256	64	64	128	24	-	14x14x512
Inception(4d)	-	-	112	288	64	64	144	32	-	14x14x528
Inception(4e)	-	-	256	320	128	128	160	32	-	14x14x832
MaxPool	3x3 / 2	-	-	-	-	-	-	-	-	7x7x832
Inception(5a)	-	-	-	-	-	-	-	-	-	7x7x832
Inception(5b)	-	-	-	-	-	-	-	-	-	7x7x1024
AvgPool	7x7 / 1	-	-	-	-	-	-	-	-	1x1x1024
Dropout (40%)	-	-	-	-	-	-	-	-	-	1x1x1024
Fully Connected	2 (classes)	-	-	-	-	-	-	-	softmax	1x1x2

Framework	PyTorch	TensorFlow
Implemented	✅	☑️