After successfuly obtaining trained model using fastAI’s vision learner and testing its working, some method of visualization of its working and bench marking was required.

Confusion matrix

Confusion matrix 5 EPoch

epochFinetune5confusionmat

Confusion matrix shows performance of our trained model based on the prediction and the actual result similar to TP VS FP table, but with multiple lables. Based on the result, there is clear distinction pattern of

It is also apparant that there exists some misdetection of trucks and automobiles both ways predicted and actual. This is due to visual similarity between automobiles and trucks, but more importantly, definition of mobiles may include truck in general. To avoid this issue, it is recommended to define clearer specification of automobiles, or narrow down the definition of automobiles to multiple lables such as “Car”, “Vans” and “Road rollers” etc…

T-SNE or T-Distributed Stochastic Neighbor Embedding provides visual representaion by utilizing non-linear dimension reduction.

5 epoch

epoch5TSNE

15 epoch

epoch15TSNE