Closed michaelmunje closed 3 years ago
Added methods to the Model class and a new DataSet class to implement transfer learning with a Pytorch CNN, though we need to ensue that the output layer has the same number of classes as the Deep Mars dataset.
I transfer learned an AlexNet model with output classes corresponding to our dataset on a GPU in a Google colab notebook. Since the file is too large to track on GitHub, I am attaching a link to a dropbox folder containing the model (and any future models I might train).
This model was trained on 700 epochs and had ~90% accuracy on a test set. I will perform some more detailed evaluation in a notebook.
We will train a multioutput CNN classifier using data from DeepMars (https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/16040/16400). This will require revisiting the Training method in the Model class.