Term 3 out of 3 Project 2
Note: due to the excessive size of the training data, this repo does not contain any. Likewise, the trained Tensorflow model is not included.
git clone <this-repo-url>
# source a new python3 environment, if desired
pip install -r requirements
# obtain training data
python main.py
High-level tuning parameters such as number of epochs and learning rate can be found in config.py
A few things to note: I found that it is too easy to exhaust the GPU memory of a GTX 1050 while implementing a Fully Convolutional Network. With this in mind, the loss value is computed on the same single image and the Mean_IOU accuracy is computed on the same batch of 10 images.
To avoid this potential bias in the future, it would be ideal to increase the working memory of the GPU. Also, randomly choosing the sample images would help generalize the progress (accuracy/loss) the model makes.
