CS194-26 Project 4: Classification and Segmentation

Vincent Zhu

The best results were achieved with a batch size of 64, with 32 channels for both convulational layers and 1568->68->10 features for the fully connected layers. I used Cross entropy loss and Adam with a learning rate of 0.005.

Training Accuracy Over 100 Epochs:

[<matplotlib.lines.Line2D at 0x7f0dfc2265c0>]

Test Set Accuracy: 89.69%

Per Class Accuracy:

'T-shirt/top': 83.6%

'Trouser': 98.7%

'Pullover': 81.2%

'Dress': 88.3%

'Coat': 89.9%

'Sandal': 98.7%

'Shirt': 67.5%

'Sneaker': 94.6%

'Bag': 98.2%

'Ankle Boot: 96.2%

Hardest classes to identify: "Shirt" and "Pullover"

Correctly identified images: (two of each class, in the order listed above)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

Misclassified images: (two of each class, same order)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

(3, 28, 28)

Visualized filter from first layer:

First convolutional layer