Image-Classifier

Training an Image Classifier in PyTorch framework by using Transfer Learning with Pre-Trained CNN Model Architectures.

Table of Content

1. Introduction
2. Prerequisites
   1. Environment
   2. Data Pre-Processing
   3. Label Mapping
3. Application
   1. Jupyter Notebook
   2. Command Line Application
      1. Training
      2. Prediction
4. References
5. Author
6. License

Introduction

This repo consisting of a Command Line Application used in training and prediction of images

In this project, you train an image classifier to predict the type of flower species

Once trained, the model can be further trained on any set of labelled images to predict any kind of images, not limited to flower species

Prerequisites

Environment

1. If you use Anaconda, you could create an environment with all required packages directly from requirements.txt by using the command:

   $ conda create --name <env> --file req.txt
   

2. If you don't, here are the required packages:

   • cudatoolkit==8.0
   • numpy==1.13.3
   • pandas==0.22.0
   • python==3.6.9
   • pytorch==0.4.0
   • torchvision==0.2.1
   • Image

   First make sure you've the latest pip version by the command:

   python -m pip install --upgrade pip
   

   Then you could use pip to install the packages

   pip install python==3.6.9 numpy==1.13.3
   

   Hint: to install cunda use the command:

   conda install cudatoolkit=8.0 -c pytorch
   

   or

   pip install cudatoolkit==8.0 -c pytorch
   

Data Pre-Processing

You could work on the flowers dataset or any other one but note that the dataset must be labeled and divided in folders where each folder is named by its class number.

Example

• flowers\train\52\image_04221.jpg

  • Image Name: image_04221.jpg
  • Class Number: 52
  • Dataset of: Training

• flowers\valid\1\image_06756.jpg

  • Image Name: image_06756.jpg
  • Class Number: 1
  • Dataset of: Validation

Label Mapping

You'll also need to load in a mapping from category label (class number) to category name. You can find this in the file cat_to_name.json in case of using flowers dataset. It's a JSON object which you can read in with the json module. This will give you a dictionary mapping the integer encoded categories to the actual names of the flowers.

Application

Command Line Application

• Training
• Prediction

Training

You would use train.py file to train a new Deep Neural Network on a dataset of images and saves the model to a checkpoint.

• Required Arguments

  • data_dir ---> directory path of the datasets.

• Optional Arguments

  • -s or --save_dir ---> directory path to save the Trained Model inside it. -- Default = work directory
  • -a or --arch ---> choosing a CNN Model Architecture. -- Default = vgg19
  • -l or --learning_rate ---> choosing a learning rate for DNN. -- Default = 0.0001
  • --hidden_units ---> number of hidden units of hidden layers (Must be integers). -- Default = 1024
  • -d or --drop_prob ---> drop probability of the hidden units of the hidden layers. -- Default = 0.2
  • -e or --epochs ---> choosing number of model trainings (Must be integer). -- Default = 20
  • -g or --gpu ---> choosing GPU for training or inference.

• Basic Usage

  python train.py datasets_directory
  

• Other Examples

  python train.py datasets_directory -s checkpoints_directory --arch vgg16
  

  python train.py datasets_directory -l 0.001 --hidden_units 2048 512
  

  python train.py datasets_directory -d 0.1 -e 10 -g
  

• Supported CNN Architectures

  Architectures
  AlexNet
  VGG11
  VGG13
  VGG16
  VGG19
  ResNet18
  ResNet34
  ResNet50
  ResNet101
  ResNet152
  DenseNet121
  DenseNet161
  DenseNet169
  DenseNet201

• Output

  • While Training: Printing out current epoch, training loss, validation loss, and validation accuracy.
    • Ex: Epoch: 8/8.. Training Loss: 0.599.. Validation Loss: 0.782.. Validation Accuracy: 0.809
  • After Training: A checkpoint that contains the trained DNN wights, biases, and hyper parameters.
    • Ex: resnet18.pth

Prediction

You would use predict.py file to predict the class of an image using the checkpoint of any saved model, and the probability of the topmost likely classes.

• Required Arguments

  • input ---> path of the flower that you want to predict its label.
  • checkpoint ---> path of the trained DNN Model.

• Optional Arguments

  • -k or --top_k ---> choosing top K most likely classes. -- Default = 1
  • -c or --category_names ---> choosing a mapping of categories to real names.
  • -g or --gpu ---> choosing GPU for training or inference.

• Basic Usage

  python predict.py input_image_path checkpoint
  

• Other Examples

  python predict.py input_image_path checkpoint -g --top_k 3
  

  python predict.py input_image_path checkpoint --category_names cat_To_name.json
  

  python predict.py input_image_path checkpoint -g -c cat_To_name.json -k 3
  

• Output

  Printing the most likely image class and it's associated probability.

References

• CS231n Convolutional Neural Networks for Visual Recognition - Transfer Learning
• PYTORCH TUTORIAL - TRANSFER LEARNING FOR COMPUTER VISION
• PYTORCH TUTORIAL - FINETUNING TORCHVISION MODELS
• Udacity - Frequently Asked Questions for Classifying Images Project

Author

• Tyler Newman
  • LinkedIn

License

MIT License