Project 4

Kevin Zhang

1.

Training graph

Per class accuracies

Accuracy of t-shirt/top : 80 %

Accuracy of trouser : 97 %

Accuracy of pullover : 83 %

Accuracy of dress : 90 %

Accuracy of coat : 73 %

Accuracy of sandal : 95 %

Accuracy of shirt : 70 %

Accuracy of sneaker : 95 %

Accuracy of bag : 97 %

Accuracy of ankle boot : 96 %

It appears shirt is the hardest class to get right. This makes sense because there there is another class, t-shirt/top, which is very similar.

Correctly Classified Images

T-shirt/Top

Trouser

Pullover

Dress

Coat

Sandal

Shirt

Sneaker

Bag

Ankle boot

Incorrectly Classified Images

T-shirt/Top

Predicted labels: Shirt, shirt

Trouser

Predicted labels: Dress, Dress

Pullover

Predicted labels: Coat, coat

Dress

Predicted labels: T-shirt/top, Coat

Coat

Predicted labels: Pullover, pullover

Sandal

Predicted labels: Sneaker, sneaker

Shirt

Predicted labels: t-shirt, top

Sneaker

Predicted labels: Ankle boot, ankle boot

Bag

Predicted labels: T-shirt/top, Pullover

Ankle boot

Predicted labels: Sneaker, sneaker

Filter visualizations

These are the monochrome representations of filters from one stack of filters from the second convolutional layer of my network.

2.

Architecture

Hyperparameters

Adam optimizer, lr=1e-3, weight_decay=1e-5, batchsize=10

Training + Validation loss

Average precision

Example

The model seems to get have low precision on pillars and be overconfident with pillar pixels, which makes sense because both are rare in the training set.