• iterations count = 60000
• batch size = 1
• dataset: MNIST
• accuracy = 0.9436

MLP architecture:

1. Fully-connected layer [784✕800]
2. ReLU activation
3. Fully-connected layer [800✕10]
4. Softmax

• dataset: MNIST

• dataset: CIFAR-10

• batch_size: 128
• Accuracy calc by samples with IoU >= 0.5

Worst sample on test set (IoU = 0.00599):

Best sample on test set (IoU = 0.98055):

• Accuracy on test set = 0.9534
• Only classification accuracy = 0.9862

• MAP: 0.9653