A basic dense neural network implemented with PyTorch to classify images in this flower dataset https://archive.ics.uci.edu/ml/datasets/iris
Required libraries: argparse numpy Pillow torch torchvision
The model can be trained with the train.py script.
Usage: python3 train.py [--options] [data_directory]
Default: python3 train.py flowers
-s, --save-directory specifies the directory to which trained models will be saved (./checkpoints by default)
-a, --arch specifies the architecture to use for transfer learning (vgg16 (default), alexnet, googlenet)
-g, --gpu trains the model using the gpu (disabled by default)
-l, --learn-rate specify learn rate (0.003 by default)
-H, --hidden-units specify the number of nodes in the hidden layer (2508 by default)
-e, --epochs specify the number of training epochs (20 by default)
The flowers subdirectory folder structure (train, test, valid) is important and should be maintained if the flowers folder is renamed.
Once trained, the model is saved to the specified save directory.
Predictions with the trained model are made with the predict.py script.
Usage: python3 predict.py [--options] [image_path] [checkpoint_path]
Where [image_path] is the relative path to an image to predict and [checkpoint_path] is the relative path to the saved model.
Example: python3 predict.py flowers/test/1/image_06743.jpg checkpoints/model_vgg16_1.pth
Sample output:
image_06743
1. pink primrose: 31.488%
2. hibiscus: 24.475%
3. tree mallow: 11.602%`
-k, --top_k specify the number of category predictions to show (3 by default)
-g, --gpu make predictions using the gpu (disabled by default)
-c, --category_names path to json file specifying the map of category numbers to names (cat_to_names.json by default)