Deep Learning Models Repository

This repository includes deep learning models that I implemented.

The code/ directory contains:

• for classification
  • VGG16 implementation with Global Average Pooling instead of Dense Layers
  • VGG16 with Global Average Pooling and MobileNet like structure
  • ResNet 18
  • ResNet 50
  • GoogleLeNet(Inception v1)
  • DenseNet
• for object detection
  • RCNN

The implementations are able to manage 2D images (so images in grayscale). In particular I got interested into the possibility to reduce the amount of parameters keeping the generalization power of the network.

Requirement

Before training any network please install the requirements with the command: pip3 install -r requirements.txt The packages required are the following:

• pandas
• pillow
• tqdm
• opencv-python-headless
• tensorflow=2.1.0
• tensorflow-datasets (for object detection)

How to train a classification network

In order to train one of the provided networks please run the following command:

python3 main.py -train classification -imgPath=../data/ -m resNet18

where:

• -train tells to the main that the user wants to train a network and expects as an argument classification
• -imgPath expects the path where the data is located (read the following section for further details)
• -m expects the name of one of the supported models
  • VGG16GlobalAverage
  • vgg16Depthwise (default)
  • resNet18
  • resNet50
  • googleLeNet
  • denseNet

The path to the data points to a directory where only 2 subdirectories are present:

• train
• validation

In each subdirectories the images are divided per class. The complete structure should be like the following:

data
+-- train/
|   +-- class1/
|   +-- class2/
|   +-- class3/
|   +-- class4/
+-- validation/
|   +-- class1/
|   +-- class2/
|   +-- class3/
|   +-- class4/

How to train an object detection network

In order to train one of the provided networks please run the following command:

python3 main.py -train OD

where:

• -train tells to the main that the user wants to train a network and expects as an argument OD
• only model currently available is RCNN

The data used for training and validation is the VOC 2007 that is automatically downloaded by the script.

Networks details

A summary of number of parameters for different architectures follows:

Model	Number of Parameters
VGG16 with Grayscale input	165,728,963
VGG16 with Grayscale input and Global Average Pooling	14,715,075
VGG16 with Grayscale input and Global Average Pooling MobileNet like	1,414,083
ResNet18	11,172,099
ResNet50 with shortcut projection	38,054,147
ResNet50 with zero padding	20,721,155
GoogleLeNet(Inception v1)	7,414,969
DenseNet	3,202,499

One should notice that ResNet18 has less parameter compared to VGG16 without the MobileNet improvement and it converges much faster than VGG16.

The focus for object detection models is not on the complexity and therefore the number of parameters is not reported.

Scientific papers

In the following table one can find reference papers for each architecture:

Name	Reference Papers
Vgg	https://arxiv.org/pdf/1409.1556.pdf
ResNet	https://arxiv.org/pdf/1512.03385.pdf
GoogleLeNet	https://arxiv.org/pdf/1409.4842.pdf
DenseNet	https://arxiv.org/pdf/1608.06993.pdf
Rcnn	https://arxiv.org/pdf/1311.2524.pdf