Field Museum Botany / Grainger Center CNN Classification

models directory

Contains the files used to define neural network layer architectures.

---

Class Hierarchy:

The CNNModel class (within cnnmodel.py) is an abstract template class which sets up a CNN architecture (layers), and compiles the model with standard hyperparameters.

CNNModel contains three abstract methods which must be implemented in a child class:

• add_convolutional_layers()
• add_hidden_layers()
• add_output_layers()

(Note: These methods should always be executed sequentially -- they are separated for human readability.)

---

Directory layout

Files:

• modeltrainingarguments.py
  • Sets up, parses, and validate command-line arguments for model training.
• model_training.py
  • ModelTrainer helper class for training Keras models, coordinating between the executing training script, images on the local filesystem, and a NN model architecture (child class of CNNModel).
• cnnmodel.py
  • Contains the abstract template class CNNModel -- see Class Hierarchy above.
• smithsonian.py
  • Contains the SmithsonianModel class, a child class of CNNModel based on the layer architecture as published in the paper "Applications of deep convolutional neural networks to digitized natural history collections," Schuettpelz, Frandsen, et al. 2017.