ramon349 / IR-redo

IR final project

This section deals with the use of the list loss function to train a neural net to generate an embedding of images. The embedding of images is thus used to idenitfy images that are similar for image search purposes

Step 1: Pre train on tiny-imagenet

run the scripts in train_tiny_scripts to generate an initial model that learns "shape" features and such to distinguish between natural images

Step 2: Train on medical images

run train_med_scripts to run training on the medical image dataset. Should provide a prevState variable to the bash scripts pointing to the last checkpoitn of step 1

Step 3. Generate embeddigns

Run bash script that loads model from step2 and generates embeddings of the data in the embeddigns folder

Step 4: evaluation

TODO 1. MAP across categories 2. gradCAM visualization of activation layers 3. TSNE plots of embedding space?