Lecture 5: Convolutional Neural Network Models

Convolutional Neural Network Models

Convolutional Neural Network ILSVRC AlexNet (2012) ZFNet (2013) VGGNet (2014) GoogleNet 2014) ResNet (2015) Conclusion

 Convolutional Neural Network (CNN)is a multi-layer neural network  Convolutional Neural Network is comprised of one or more convolutional layers (often with a pooling layers) and then followed by one or more fully connected layers.

 Convolutional layer acts as a feature extractor that extracts features of the inputs such as edges, corners , endpoints.

 Pooling layer reduces the resolution of the image that reduce the precision of the translation (shift and distortion) effect.

 fully connected layer have full connections to all activations in the previous layer.  Fully connect layer act as classifier.

Output Image = ( ( (ImageSize+2*Padding)- KernalSize )/ Stride) +1

 Conv 3x3 with stride=1,padding=0 6x6 Image 4x4

 MaxPooling 2x2 with stride=2 4x4 Image 2x2

 MaxPooling 3x3 with stride=2 7x7 Image 3x3

 ImageNet Large Scale Visual Recognition Challenge is image classification challenge to create model that can correctly classify an input image into 1,000 separate object categories. Models are trained on 1.2 million training images with another 50,000 images for validation and 150,000 images for testing

 AlexNet achieve on ILSVRC 2012 competition 15.3% Top-5 error rate compare to 26.2% achieved by the second best entry.  AlexNet using batch stochastic gradient descent on training, with specific values for momentum and weight decay.  AlexNet implement dropout layers in order to combat the problem of overfitting to the training data.

Image Conv1 Pool1 Conv2 Pool2 Conv3 Conv4 Conv5 Pool3 FC1 FC2 FC3  AlexNet has 8 layers without count pooling layers.  AlexNet use ReLU for the nonlinearity functions  AlexNet trained on two GTX 580 GPUs for five to six days

Image 227x227x3 Conv11-96 Maxpool Conv5-256 Maxpool Conv3-384 Conv3-384 Conv3-256 Maxpool FC-4096 FC-4096 FC-1000

 Layer 0: Input image  Size: 227 x 227 x 3  Memory: 227 x 227 x 3

 Layer 0: 227 x 227 x 3  Layer 1: Convolution with 96 filters, size 11×11, stride 4, padding 0  Outcome Size= 55 x 55 x 96  (227-11)/4 + 1 = 55 is size of outcome  Memory: 55 x 55 x 96 x 3 (because of ReLU & LRN(Local Response Normalization))  Weights (parameters) : 11 x 11 x 3 x 96

 Layer 1: 55 x 55 x 96  Layer 2: Max-Pooling with 3×3 filter, stride 2  Outcome Size= 27 x 27 x 96  (55 – 3)/2 + 1 = 27 is size of outcome  Memory: 27 x 27 x 96

 Layer 2: 27 x 27 x 96  Layer 3: Convolution with 256 filters, size 5×5, stride 1, padding 2  Outcome Size = 27 x 27 x 256  Original size is restored because of padding  Memory: 27 x 27 x 256 x 3 (because of ReLU and LRN)  Weights: 5 x 5 x 96 x 256

 Layer 3: 27 x 27 x 256  Layer 4: Max-Pooling with 3×3 filter, stride 2  Outcome Size = 13 x 13 x 256  (27 – 3)/2 + 1 = 13 is size of outcome  Memory: 13 x 13 x 256

 Layer 4: 13 x 13 x 256  Layer 5: Convolution with 384 filters, size 3×3, stride 1, padding 1  Outcome Size = 13 x 13 x 384  the original size is restored because of padding (13+2 -3)/1 +1 =13  Memory: 13 x 13 x 384 x 2 (because of ReLU)  Weights: 3 x 3 x 256 x 384

 Layer 5: 13 x 13 x 384  Layer 6: Convolution with 384 filters, size 3×3, stride 1, padding 1  Outcome Size = 13 x 13 x 384  the original size is restored because of padding  Memory: 13 x 13 x 384 x 2 (because of ReLU)  Weights: 3 x 3 x 384 x 384

 Layer 6: 13 x 13 x 384  Layer 7: Convolution with 256 filters, size 3×3, stride 1, padding 1  Outcome Size = 13 x 13 x 256  the original size is restored because of padding  Memory: 13 x 13 x 256 x 2 (because of ReLU)  Weights: 3 x 3 x 384 x 256

 Layer 7: 13 x 13 x 256  Layer 8: Max-Pooling with 3×3 filter, stride 2  Outcome Size = 6 x 6 x 256  (13 – 3)/2 + 1 = 6 is size of outcome  Memory: 6 x 6 x 256

 Layer 8: 6x6x256=9216 pixels are fed to FC  Layer 9: Fully Connected with 4096 neuron  Memory: 4096 x 3 (because of ReLU and Dropout)  Weights: 4096 x (6 x 6 x 256)

 Layer 9: Fully Connected with 4096 neuron  Layer 10: Fully Connected with 4096 neuron  Memory: 4096 x 3 (because of ReLU and Dropout)  Weights: 4096 x 4096

 Layer 10: Fully Connected with 4096 neuron  Layer 11: Fully Connected with 1000 neurons  Memory: 1000  Weights: 4096 x 1000

 Total (label and softmax not included)  Memory: 2.24 million  Weights: 62.37 million

 first use of ReLU  Alexnet used Norm layers  Alexnet heavy used data augmentation  Alexnet use dropout 0.5  Alexnet batch size is 128  Alexnet used SGD Momentum 0.9  Alexnet used learning rate 1e-2, reduced by 10

[227x227x3] INPUT [55x55x96] CONV1 : 96 11x11 filters at stride 4, pad 0 27x27x96] MAX POOL1 : 3x3 filters at stride 2 [27x27x96] NORM1: Normalization layer [27x27x256] CONV2: 256 5x5 filters at stride 1, pad 2 [13x13x256] MAX POOL2: 3x3 filters at stride 2 [13x13x256] NORM2: Normalization layer

[13x13x384] CONV3: 384 3x3 filters at stride 1, pad 1 [13x13x384] CONV4: 384 3x3 filters at stride 1, pad 1 [13x13x256] CONV5: 256 3x3 filters at stride 1, pad 1 [6x6x256] MAX POOL3: 3x3 filters at stride 2 [4096] FC6: 4096 neurons [4096] FC7: 4096 neurons [1000] FC8: 1000 neurons

 Implement AlexNet using TFLearn

 ZFNet the winner of the competition ILSVRC 2013 with 14.8% Top-5 error rate  ZFNet built by Matthew Zeiler and Rob Fergus  ZFNet has the same global architecture as Alexnet, that is to say 5 convolutional layers, two fully connected layers and an output softmax one. The differences are for example better sized convolutional kernels.

 ZFNet used filters of size 7x7 and a decreased stride value, instead of using 11x11 sized filters in the first layer (which is what AlexNet implemented).  ZFNet trained on a GTX 580 GPU for twelve days.  Developed a visualization technique named Deconvolutional Network “deconvnet” because it maps features to pixels.

AlexNet but: • CONV1: change from (11x11 stride 4) to (7x7 stride 2) • CONV3,4,5: instead of 384, 384, 256 filters use 512, 1024, 512

 Keep it deep. Keep it simple.  VGGNet the runner up of the competition ILSVRC 2014 with 7.3% Top-5 error rate.  VGGNet use of only 3x3 sized filters is quite different from AlexNet’s 11x11 filters in the first layer and ZFNet’s 7x7 filters.  two 3x3 conv layers have an effective receptive field of 5x5  Three 3x3 conv layers have an effective receptive field of 7x7  VGGNet trained on 4 Nvidia Titan Black GPUs for two to three weeks

 Interesting to notice that the number of filters doubles after each maxpool layer. This reinforces the idea of shrinking spatial dimensions, but growing depth.  VGGNet used scale jittering as one data augmentation technique during training  VGGNet used ReLU layers after each conv layer and trained with batch gradient descent

Image Conv Conv Pool Conv Conv Pool Conv Conv Conv Pool Conv Conv Conv Pool Conv Conv Conv Pool FC FC FC Image Low Level Feature Mid Level Feature High Level Feature Classifier

Input 224x224x3 Conv3-64 Conv3-64 Maxpool Conv3-128 Conv3-128 Maxpool Conv3-256 Conv3-256 Conv3-256 Maxpool Conv3-512 Conv3-512 Conv3-512 Maxpool Conv3-512 Conv3-512 Conv3-512 Maxpool FC-4096 FC-4096 FC-1000 VGGNet 16

Input 224x224x3 Conv3-64 Conv3-64 Maxpool Conv3-128 Conv3-128 Maxpool Conv3-256 Conv3-256 Conv3-256 Conv3-256 Maxpool Conv3-512 Conv3-512 Conv3-512 Conv3-512 Maxpool Conv3-512 Conv3-512 Conv3-512 Conv3-512 Maxpool FC-4096 FC-4096 FC-1000 VGGNet 19

 Implement VGGNet16 using TFLearn

 GoogleNet is the winner of the competition ILSVRC 2014 with 6.7% Top-5 error rate.  GoogleNet Trained on “a few high-end GPUs with in a week”  GoogleNet uses 12x fewer parameters than AlexNet  GoogleNet use an average pool instead of fully connected layers, to go from a 7x7x1024 volume to a 1x1x1024 volume. This saves a huge number of parameters.

 GoogleNet used 9 Inception modules in the whole architecture  This 1x1 convolutions (bottleneck convolutions) allow to control/reduce the depth dimension which greatly reduces the number of used parameters due to removal of redundancy of correlated filters.  GoogleNet has 22 Layers deep network

 GoogleNet use inexpensive Conv1 to compute reduction before the expensive Conv3 and Conv5  Conv1 follow by Relu to reduce overfitting

Input 224x224x3 Conv7/2-64 Maxpool3/2 Conv1 Conv3/1- 192 Maxpool3/2 Inception3a 256 Inception3b 480 Maxpool3/2 Inception4a 512 Inception4b 512 Inception4c 512 Inception4d 528 Inception4e 832 Maxpool3/2 Inception5a 832 Inception5b 1024 Avgpool7/1 Dropout 40% FC-1000 Softmax- 1000 GoogleNet

Type Size/ Stride Output Depth Conv1 # Conv3 Conv3 # Conv5 Conv5 Pool Param Ops Conv 7x7/2 112x112x64 1 - - - - - - 2.7K 34M Maxpool 3x3/2 56x56x64 0 - - - - - - - - Conv 3x3/1 56x56x192 2 - 64 192 - - - 112K 360M Maxpool 3x3/2 28x28x192 0 - - - - - - - - Inception 3a - 28x28x256 2 64 96 128 16 32 32 159K 128M Inception 3b - 28x28x480 2 128 128 192 32 96 64 380K 304M Maxpool 3x3/2 14x14x480 0 - - - - - - - - Inception 4a - 14x14x512 2 192 96 208 16 48 64 364K 73M Inception 4b - 14x14x512 2 160 112 224 24 64 64 437K 88M Inception 4c - 14x14x512 2 128 128 256 24 64 64 463K 100M Inception 4d - 14x14x528 2 112 144 288 32 64 64 580K 119M

Type Size/ Stride Output Depth Conv1 # Conv3 Conv3 # Conv5 Conv5 Pool Param Ops Inception 4e - 14x14x832 2 256 160 320 32 128 128 840K 170M Maxpool 3x3/2 7x7x832 0 - - - - - - - - Inception 5a - 7x7x832 2 256 160 320 32 128 128 1072K 54M Inception 5b - 7x7x1024 2 384 192 384 48 128 128 1388K 71M Avgpool 7x7/1 1x1x1024 0 - - - - - - - - Dropout .4 - 1x1x1024 0 - - - - - - - - Linear - 1x1x1024 1 - - - - - - 1000K 1M Softmax - 1x1x1024 0 - - - - - - - - Total Layers 22

 ResNet the winner of the competition ILSVRC 2015 with 3.6% Top-5 error rate.  ResNet mainly inspired by the philosophy of VGGNet.  ResNet proposed a residual learning approach to ease the difficulty of training deeper networks. Based on the design ideas of Batch Normalization (BN), small convolutional kernels.  ResNet is a new 152 layer network architecture.  ResNet Trained on an 8 GPU machine for two to three weeks

 The idea of Batch Normalization, instead of just normalizing the inputs to the network, BN normalize the inputs to layers within the network. It's called "batch" normalization because during training, we normalize each layer's inputs by using the mean and variance of the values in the current mini-batch.

 Residual network  Keys:  No max pooling  No hidden fc  No dropout  Basic design (VGG-style)  All 3x3 conv (almost)  Batch normalization

Conv Layers Preserving base information can treat perturbation

 Residual Bottleneck consist of a 1×1 layer for reducing dimension, a 3×3 layer, and a 1×1 layer for restoring dimension.  Residual Bottleneck used for deep deep network layers.

 Residual block Conv3- 64 Conv3- 64 2Conv3- 64

Image Conv7/ 2-64 Pool/2 Conv3- 64 Conv3- 64 Conv3- 64 Conv3- 64 Conv3/ 2-128 Conv3- 128 Conv3- 128 Conv3- 128 Conv3/ 2-256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3/ 2-512 Conv3- 512 Conv3- 512 Conv3- 512 Avg pool FC-1000 ResNet 18

Image Conv7/2- 64 Pool/2 2Conv3- 64 2Conv3- 64 2Conv3/2- 128 2Conv3- 128 2Conv3/2- 256 2Conv3- 256 2Conv3/2- 512 2Conv3- 512 Avg pool FC-1000 ResNet 18

Image Conv7/2 -64 Pool/2 Conv3- 64 Conv3- 64 Conv3- 64 Conv3- 64 Conv3- 64 Conv3- 64 Conv3/2 -128 Conv3- 128 Conv3- 128 Conv3- 128 Conv3- 128 Conv3- 128 Conv3- 128 Conv3- 128 Conv3/2 -256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3- 256 Conv3/2 -512 Conv3- 512 Conv3- 512 Conv3- 512 Conv3- 512 Conv3- 512 Avg pool FC-1000 ResNet 34

Image Conv7/2- 64 Pool/2 2Conv3- 64 2Conv3- 64 2Conv3- 64 2Conv3/2 -128 2Conv3- 128 2Conv3- 128 2Conv3- 128 2Conv3/2 -256 2Conv3- 256 2Conv3- 256 2Conv3- 256 2Conv3- 256 2Conv3- 256 2Conv3/2 -512 2Conv3- 512 2Conv3- 512 Avg pool FC-1000 ResNet 34

Image Conv7/2- 64 Pool/2 2Conv3- 64 2Conv3- 128 2Conv3- 256 2Conv3- 512 Avg pool FC-1000 ResNet 18 2 2 2 2 ResNet 34 3 4 6 3

 Residual Bottleneck Conv1-64 Conv3-64 Conv1- 256 Conv1-64 Conv3-64 Conv1-256

Image Conv7/2- 64 Pool/2 Conv1-64 Conv3-64 Conv1- 256 Conv1-64 Conv3-64 Conv1- 256 Conv1-64 Conv3-64 Conv1- 256 Conv1/2- 128 Conv3- 128 Conv1- 512 Conv1- 128 Conv3- 128 Conv1- 512 Conv1- 128 Conv3- 128 Conv1- 512 Conv1- 128 Conv3- 128 Conv1- 512 Conv1/2- 256 Conv3- 256 Conv1- 1024 Conv1- 256 Conv3- 256 Conv1- 1024 Conv1- 256 Conv3- 256 Conv1- 1024 Conv1- 256 Conv3- 256 Conv1- 1024 Conv1- 256 Conv3- 256 Conv1- 1024 Conv1- 256 Conv3- 256 Conv1- 1024 Conv1/2- 512 Conv3- 512 Conv1- 2048 Conv1- 512 Conv3- 512 Conv1- 2048 Conv1- 512 Conv3- 512 Conv1- 2048 Avg pool FC-1000 ResNet 50

Image Conv7/2-64 Pool/2 Conv1-64 Conv3-64 Conv1-256 Conv1-64 Conv3-64 Conv1-256 Conv1-64 Conv3-64 Conv1-256 Conv1/2-128 Conv3-128 Conv1-512 Conv1-128 Conv3-128 Conv1-512 Conv1-128 Conv3-128 Conv1-512 Conv1-128 Conv3-128 Conv1-512 Conv1/2-256 Conv3-256 Conv1-1024 Conv1-256 Conv3-256 Conv1-1024 Conv1-256 Conv3-256 Conv1-1024 Conv1-256 Conv3-256 Conv1-1024 Conv1-256 Conv3-256 Conv1-1024 Conv1-256 Conv3-256 Conv1-1024 Conv1/2-512 Conv3-512 Conv1-2048 Conv1-512 Conv3-512 Conv1-2048 Conv1-512 Conv3-512 Conv1-2048 Avg pool FC-1000 ResNet 50

Image Conv7/2-64 Pool/2 Conv1-64 Conv3-64 Conv1-256 Conv1-128 Conv3-128 Conv1-512 Conv1-256 Conv3-256 Conv1-1024 Conv1-512 Conv3-512 Conv1-2048 Avg pool FC-1000 ResNet 50 3 4 6 3 ResNet 101 3 4 23 3 ResNet 152 3 8 36 3

Layer Output 18-Layer 34-Layer 50-Layer 101-Layer 152-Layer Conv-1 112x112 7x7/2-64 Conv-2 56x56 3x3 Maxpooling/2 𝟐𝐱 𝟑𝐱𝟑, 𝟔𝟒 𝟑𝐱𝟑, 𝟔𝟒 𝟑𝐱 𝟑𝐱𝟑, 𝟔𝟒 𝟑𝐱𝟑, 𝟔𝟒 𝟑𝐱 𝟏𝐱𝟏, 𝟔𝟒 𝟑𝐱𝟑𝐱𝟔𝟒 𝟏𝐱𝟏𝐱𝟐𝟓𝟔 𝟑𝐱 𝟏𝐱𝟏, 𝟔𝟒 𝟑𝐱𝟑𝐱𝟔𝟒 𝟏𝐱𝟏𝐱𝟐𝟓𝟔 𝟑𝐱 𝟏𝐱𝟏, 𝟔𝟒 𝟑𝐱𝟑𝐱𝟔𝟒 𝟏𝐱𝟏𝐱𝟐𝟓𝟔 Conv-3 28x28 𝟐𝐱 𝟑𝐱𝟑, 𝟏𝟐𝟖 𝟑𝐱𝟑, 𝟏𝟐𝟖 𝟒𝐱 𝟑𝐱𝟑, 𝟏𝟐𝟖 𝟑𝐱𝟑, 𝟏𝟐𝟖 𝟒𝐱 𝟏𝐱𝟏, 𝟏𝟐𝟖 𝟑𝐱𝟑𝐱𝟏𝟐𝟖 𝟏𝐱𝟏𝐱𝟓𝟏𝟐 𝟒𝐱 𝟏𝐱𝟏, 𝟏𝟐𝟖 𝟑𝐱𝟑𝐱𝟏𝟐𝟖 𝟏𝐱𝟏𝐱𝟓𝟏𝟐 𝟖𝐱 𝟏𝐱𝟏, 𝟏𝟐𝟖 𝟑𝐱𝟑𝐱𝟏𝟐𝟖 𝟏𝐱𝟏𝐱𝟓𝟏𝟐 Conv-4 14x14 𝟐𝐱 𝟑𝐱𝟑, 𝟐𝟓𝟔 𝟑𝐱𝟑, 𝟐𝟓𝟔 𝟔𝐱 𝟑𝐱𝟑, 𝟐𝟓𝟔 𝟑𝐱𝟑, 𝟐𝟓𝟔 𝟔𝐱 𝟏𝐱𝟏, 𝟐𝟓𝟔 𝟑𝐱𝟑𝐱𝟐𝟓𝟔 𝟏𝐱𝟏𝐱𝟏𝟎𝟐𝟒 𝟐𝟑𝐱 𝟏𝐱𝟏, 𝟐𝟓𝟔 𝟑𝐱𝟑𝐱𝟐𝟓𝟔 𝟏𝐱𝟏𝐱𝟏𝟎𝟐𝟒 𝟑𝟔𝐱 𝟏𝐱𝟏, 𝟐𝟓𝟔 𝟑𝐱𝟑𝐱𝟐𝟓𝟔 𝟏𝐱𝟏𝐱𝟏𝟎𝟐𝟒 Conv-5 7x7 𝟐𝐱 𝟑𝐱𝟑, 𝟓𝟏𝟐 𝟑𝐱𝟑, 𝟓𝟏𝟐 𝟑𝐱 𝟑𝐱𝟑, 𝟓𝟏𝟐 𝟑𝐱𝟑, 𝟓𝟏𝟐 𝟑𝐱 𝟏𝐱𝟏, 𝟓𝟏𝟐 𝟑𝐱𝟑𝐱𝟓𝟏𝟐 𝟏𝐱𝟏𝐱𝟐𝟎𝟒𝟖 𝟑𝐱 𝟏𝐱𝟏, 𝟓𝟏𝟐 𝟑𝐱𝟑𝐱𝟓𝟏𝟐 𝟏𝐱𝟏𝐱𝟐𝟎𝟒𝟖 𝟑𝐱 𝟏𝐱𝟏, 𝟓𝟏𝟐 𝟑𝐱𝟑𝐱𝟓𝟏𝟐 𝟏𝐱𝟏𝐱𝟐𝟎𝟒𝟖 1x1 Avgpool-FC1000-Softmax Flops 𝟏. 𝟖𝐱𝟏𝟎𝟗 𝟑. 𝟔𝐱𝟏𝟎𝟗 𝟑. 𝟖𝐱𝟏𝟎𝟗 𝟕. 𝟔𝐱𝟏𝟎𝟗 𝟏𝟏. 𝟑𝐱𝟏𝟎𝟗

 Pre Activation ResNet  Paper: Identity Mappings in Deep Residual Networks Original Proposed

 Implement ResNet using TFLearn

26.2 15.3 14.8 7.3 6.7 3.6 0 5 10 15 20 25 30 Before 2012 AlexNet 2012 ZFNet 2013 VGGNet 2014 GoogleNet 2014 ResNet 2015

facebook.com/mloey mohamedloey@gmail.com twitter.com/mloey linkedin.com/in/mloey mloey@fci.bu.edu.eg mloey.github.io

Lecture 5: Convolutional Neural Network Models

More Related Content

What's hot

Similar to Lecture 5: Convolutional Neural Network Models

More from Mohamed Loey

Recently uploaded

Lecture 5: Convolutional Neural Network Models