- Install conda.
- Install requirements.txt
conda create --name dl-env --file ./requirements.txt
. - Depending on environment cuda might or might not be needed consult this page and install torch and friends accordingly.
- Create an empty
checkpoints
directory if it does not exist already - Download the dataset
- Rename the dataset file to
landmark_images
if necessary.
The models can be trained from the 2 notebooks provided:
custom_cnn.ipynb
for out custom cnn modeltransfer_learning.ipynb
for the 3 transfer learning models
Running the notebook trains the corresponding model saving the best performing checkpoint each time and in the end of the training it wraps the model with a predicting wrapper to change the output of the model from percentages to a single prediction, runs a confusion matrix and saves the model with the wrapper.
The trained models are used in the SpotTheSpot application