Udacity AI with Python project: Create-your-own-image-classifier

Purpose: To build an image classifier that will predict flower types using PyTorch. This was done for Udacity's AI Programming with Python Nanodegree.

Part 1 - Jupyter notebook

The full trainng and prediction workflow is contained in the Jupyter Notebook. The command line app in part 2 uses code developed in this notebook.

Part 2 - Comand line app

There are 2 python executables for the command line app, one for training and one for prediction. These executables use modules from 5 additional files as described below.

Training

train.py can be executed to build and train the image classifier

Mandatory argument

The user will need to specify one mandatory argument 'data_directory' contating the path to the training data directory.

Optional arguments:

--data_directory: the saving directory and filename for saving the checkpoint. Default is 'save_directory/checkpoint1'.
--arch: the user can choose which architecture to use for the neural network. The default architecture is vgg11.
--GPU: Allows the user to specifify if GPU will be used. Default is GPU = True.
--learning_rate: sets the learning rate for gradient descent: default is 0.01.
--hidden_units: an int specifying how many neurons the hidden-layer in the classifier will contain if so chosen. Default is 4096.
--epochs: specifies the number of epochs as integer. Set to 2 by default.

Prediction

predict.py can be executed to predict a flower type for a single image.

Mandatory arguments:

The user will need to specify the the path to the input image and the checkpoint filename to be loaded.

Optional arguments:

--top_k: let's the user specify the numer of top K-classes to output. Default is 3.
--GPU: Allows the user to specifify if GPU will be used. Default is GPU = True.
--category_names: allows user to provide path of JSON file mapping categories to names. Default is cat_to_name.json.

Additional files

train.py and predict.py use modules from the following files:

• get_input_args.py contains modules to accept inputs from the user via the command line.
• prepare_data.py contains modules for preparing training and validation data and preparation of the image before prediction.
• classifier.py contains modules to build and train the model and run a prediction.
• save_load_model.py contains mdolues to save and load the checkpoint.
• display_results.py contains modules to display the prediction results to the user.