Final project of the KTH DD2412 - Deep Learning, Advanced course.

The scope of this project is to replicate FixMatch, a method developed by Google researchers in order to perform semi-supervised learning.

The tasks we performed are the followings:

1. Replication of the method using PyTorch and training on CIFAR-10 and SVHN datasets.
2. Unbalance the supervised classes number of samples when training.
3. Varying the labeled-unlabeled data ratio.
4. Employing the merge of different datasets of the same nature as unlabeled data, SVHN and MNIST in this case.

The main outcomes obtained in each task can be seen below. In order to get a deeper insight of our findings, go to FixMatch & SLL Exploration.

• Task 1:

Accuracy CIFAR-10 40 labels.	Accuracy CIFAR-10 250 labels.

Accuracy CIFAR-10 4000 labels.	Accuracy SVHN 250 labels.

• Task 2:

Confusion matrix SVHN class 3 not scaled.	Confusion matrix SVHN class 3 downscaled 50%.	Confusion matrix SVHN class 3 upscaled 50%

• Task 3:

Accuracy CIFAR-10 25% unlabeled data.	Accuracy CIFAR-10 50% unlabeled data.

Accuracy CIFAR-10 75% unlabeled data.	Accuracy CIFAR-10 100% unlabeled data.

• Task 4:

Confusion matrix SVHN merged unlabeled data.	Confusion matrix MNIST merged unlabeled data.