AI image classification and machine learning utilizing the PyTorch framework

Project assets:

• Image Classifier Project.ipynb Jupyter Notebook
• Image Classifier Project.html HTML export of the Jupyter Notebook above.
• train.py to train a new network on a data set.
• predict.py to predict flower name from an image.

Example Image for Training

Image categories are found in cat_to_name.json and flower images can be downloaded in the gziped tar file flower_data.tar.gz from Udacity.

Get flower images:

mkdir -p flowers && cd flowers
curl https://s3.amazonaws.com/content.udacity-data.com/nd089/flower_data.tar.gz | tar xz

You should now have test, train and valid directories containing classification directories and flower images under the flowers directory.

Examples train.py

Help:

python ./train.py -h

Train on CPU with default vgg16:

python ./train.py ./flowers/train/

Train on GPU with densenet121 with one 500 node layer:

python ./train.py ./flowers/train --gpu --arch "densenet121" --hidden_units 500 --epochs5

Additional hidden layers with checkpoint saved to densenet201 directory.

python ./train.py ./flowers/train --gpu --arch=densenet201 --hidden_units 1280 640 --save_dir densenet201

Examples train.py

Help

python ./predict.py -h

Basic Prediction

python ./predict.py flowers/valid/5/image_05192.jpg checkpoint.pth

Prediction with Top 10 Probabilities

python ./predict.py flowers/valid/5/image_05192.jpg checkpoint.pth --tok_k 10

Prediction with GPU

python ./predict.py flowers/valid/5/image_05192.jpg checkpoint.pth --gpu

Part 1 Jupyter Notebook

Image Classifier Project.ipynb

To review the Image Classifier Project.ipynb notebook, launch Jupyter Notebook from the project root:

jupyter notebook

Part 2 Scripts

train.py

Options:

• Set directory to save checkpoints
  • python train.py data_dir --save_dir save_directory
• Choose architecture
  • python train.py data_dir --arch "vgg13"
• Set hyperparameters
  • python train.py data_dir --learning_rate 0.01 --hidden_units 512 256 --epochs 20
• Use GPU for training
  • python train.py data_dir --gpu

Help - python ./train.py -h:

usage: python ./train.py ./flowers/train --gpu --learning_rate 0.001 --hidden_units 3136 --epochs 5

Train and save an image classification model.

positional arguments:
  data_directory

optional arguments:
  -h, --help            show this help message and exit
  --save_dir SAVE_DIR   Directory to save training checkpoint file (default:
                        .)
  --save_name SAVE_NAME
                        Checkpoint filename. (default: checkpoint)
  --categories_json CATEGORIES_JSON
                        Path to file containing the categories. (default:
                        cat_to_name.json)
  --arch ARCH           Supported architectures: vgg11, vgg13, vgg16, vgg19,
                        densenet121, densenet169, densenet161, densenet201
                        (default: vgg16)
  --gpu                 Use GPU (default: False)

hyperparameters:
  --learning_rate LEARNING_RATE
                        Learning rate (default: 0.001)
  --hidden_units HIDDEN_UNITS [HIDDEN_UNITS ...], -hu HIDDEN_UNITS [HIDDEN_UNITS ...]
                        Hidden layer units (default: [3136, 784])
  --epochs EPOCHS       Epochs (default: 1)

predict.py

• Basic usage
  • python predict.py /path/to/image checkpoint
• Options
  • Return top KK most likely classes
    • python predict.py input checkpoint --top_k 3
  • Use a mapping of categories to real name
    • python predict.py input checkpoint --category_names cat_to_name.json
  • Use GPU for inference
    • python predict.py input checkpoint --gpu

Help - python ./predict.py -h:

usage: python ./predict.py /path/to/image.jpg checkpoint.pth

Image prediction.

positional arguments:
  path_to_image         Path to image file.
  checkpoint_file       Path to checkpoint file.

optional arguments:
  -h, --help            show this help message and exit
  --save_dir SAVE_DIR   Directory to save training checkpoint file (default:
                        .)
  --top_k TOP_K         Return top KK most likely classes. (default: 5)
  --category_names CATEGORIES_JSON
                        Path to file containing the categories. (default:
                        cat_to_name.json)
  --gpu                 Use GPU (default: False)