Fashion-MNIST_Classification

Course project (mandatory task) of SJTU CS3612: Machine Learning, 2023 spring.

Attention: Discussion & reference welcomed, but NO PLAGIARISM !!!

Task objective：

You should design one neural network by yourself. Specifically, each designed neural network should contain 12-35 layers, including convolutional layers, ReLU layers, Batch Normalization layers, fully connected layers, and maxpoolinglayers.

• You should not directly use classical neural networks (including but not limited to VGG-11/16/19, AlexNet, ResNet-18/24/32/36/44/56/102, DenseNet, GoogLeNet, and InceptionNet). Moreover,you should not design a new neural network by just adding or removing several layers from the above classical neural networks.
• You should train your designed neural networks on the dataset.
• You should use both PCA and t-SNE to visualize features on the designed neural network.

Designed network architecture：

Visualization：

• PCA
• t-SNE
• t-SNE with real image display

Run codes：

• Reproduce the best accuracy on test set with pretrained weight

python main.py --eval

• Reproduce the whole training process with default parameterization

python main.py --train

• Show visualization results

python visualize.py

For further information, refer to Section 1 of the project report here.