Merge pull request #68 from FR-DC/frrd-81

FRRD-81 FixMatch Visualization between Chestnut & Casuarina

notebooks/frrd-81/README.md

@@ -0,0 +1,20 @@
+# Jupyter Notebooks
+
+The notebook may not render Plotly on GitHub.
+View it on: https://nbviewer.org/
+
+## FRRD-80/81: Visualization of FixMatch
+
+This simply shows the embeddings of the FixMatch model.
+The embeddings are generated by the model and then visualized using LDA.
+
+We found that the model transfer doesn't necessarily equate to a well separated
+embedding space. This is likely due to the fact that the model is trained on
+too small of a dataset.
+
+We attribute much of performance degradation to our contextual shift when
+we chose our backbone model. We chose a model that was trained on ImageNet,
+which in retrospect, has a large contextual shift from our dataset.
+
+However, it's not all bad, we reinforced the idea that this shift is a real
+problem and that we need to address it in our future work.