Morphological Neural Networks (MNN)

This is the repository for the current experiments on MNNs in the Machine Learning and Sensing Lab. These experiments are a continuation of the work done in Xu (2023).

Instructions To Train

Clone this repository:

git clone https://github.com/GatorSense/MorphExperiments
cd MorphExperiments

Use the train.py file to train a binary classifier on hit-miss filters on MNIST images of threes and not threes (fours and eights). The test accuracy will include all 10 classes found in the MNIST data, showcasing its ability to reject classes, even if not included in the training data.

For example:

python train.py --epochs 100 --lr 0.1 --use-comet

These are all arguments used in our script:

options:
  -h, --help            show this help message and exit
  --batch-size BATCH_SIZE
                        input batch size for training (default: 64)
  --test-batch-size TEST_BATCH_SIZE
                        input batch size for testing (default: 1000)
  --epochs EPOCHS       number of epochs to train (default: 10)
  --lr LR               learning rate (default: 0.1)
  --momentum MOMENTUM   SGD momentum (default: 0.5)
  --no-cuda             disables CUDA training
  --log-interval LOG_INTERVAL
                        how many batches to wait before logging training status
  --model-type MODEL_TYPE
                        type of layer to use (default: morph, could use conv or MCNN)
  --use-comet           uses comet.ml to log training metrics and graphics
  --model-filename MODEL_FILENAME
                        filename for saved model (default: None)

Instructions To Run Pretrained model

Simplly put the pretrained model.pth file in the same directory as demo.py and run that script.

Metrics

The following metrics are logged by train.py:

• Training and testing confusion matrices
• Histograms of feature map values throughout training per class
• True vs. prediction label heatmap on all test classes
• Hit and miss filters over training

All of the above are viewed on Comet.ml. To use this functionality, store the API key in .env file at the root directory with the following format:

COMET_API_KEY=your_api_key_here

Project Structure

MorphExperiments/
├─ models/
│  ├─ MNN.py             # Implementation of learnable Hit-Miss transformation
│  ├─ models.py          # Model classes implementation (MNN, CNN, MCNN)
├─ utils/
│  ├─ custom_dataset.py  # Custom datasets for binary classifiers
│  ├─ logger.py          # Function to log model weights thorugh training
│  ├─ plot.py            # Plots feature map histograms and filters through training
├─ .gitignore
├─ train.py              # Script to run for training loop
├─ demo.py               # Script to run pretrained model and show accuracy
└─ README.md

References

Xu, Weihuang (2023). Deep Morph-Convolutional Neural Network: Combining Morphological Transform and Convolution in Deep Neural Networks. [Doctoral dissertation, University of Florida]. UF Digital Collections. https://ufdc.ufl.edu/UFE0059487/00001/pd