This repo is for learning purposes.
- an upgrade from the simple convolutional notebook
- use multi-headed attention to encode images as a sequence of patches
- final layer classifier learns to classify based on the encoded representations of the image
Chest X-Ray Medical Diagnosis Notebook:
- train classifier on medical image data
- deal with class imbalance (very common issue)
- more practice with convolutions, channels, filters, and dimensions
- train the model on (anchor, positive, negative) with Triplet Loss for 30 epochs
- then train it with just 5 epochs on Cross-entropy Loss
- the encoder of the embeddings used is that of a Transformer