Image Classifier

This is a project from Udacity Data Scientist Nanodegree Program. An Image Classifier that will classify flowers name from images.

Overview

In this project, I built a classifier comprises pretrained ImageNet model with new defined classifier architecture.

The classifier architecture used here is 3 hidden layers with units for each of them are 4096, 2048, and 512 respectively. This user-defined architecture gives good results classifying images with 85% accuracy only in 15 epochs and 0.0003 for the learning rate. Thanks to transfer learning and pretrained model provided by PyTorch.

Table of Contents

1. Getting Started
   1. Requirements
   2. Installation
   3. Usage
2. Experiment Results
3. References

Getting Started

These instruction show what requirements that need to be installed and how to start or inferencing.

Requirements

To be able to use this repository, you need to have:

1. PyTorch (torch and torchvision)
2. Pillow
3. Numpy
4. Matplotlib (if you want to visualize the image)

Also, you must have downloaded the flowers dataset provided here. If you have downloaded the dataset, you need to make your dataset to match PyTorch data loader format.

Usage

• First, make sure you have the dataset in a proper directory structure

• There are optional arguments that can be passed as user input: --gpu, -hunits, --epochs, --learning_rate, --arch, and --save_dir.

  • --save_dir pass directory where you want to save the trained model,
  • --gpu pass a boolean whether activate the gpu or not,
  • --hidden_units or -hunits pass a list of hidden unit sizes to be the classifier of the model,
  • --learning_rate needs you to pass the learning rate, and
  • --arch is your choice of pretrained model architecture given in the torchvision.models

• dataset is required as positional argument as dataset directory for training the model.

• You can create and train your model using train.py with some optional arguments above, for example:

python train.py flowers --arch 'vgg16' -hunits 4096 -hunits 2048 -hunits 512 --learning_rate 0.0003 -saved models --gpu

If you run the above command, you will get a pretrained model of vgg16 with self-defined classifier with hidden units [4096, 2048, 512] that will revise predefined classifier in the pretrained model, learning rate 0.0003, train the network with gpu (the program will check the availability first), and save the trained model in folder named models.

References