分享
004-MNIST New Test
输入“/”快速插入
004-MNIST New Test
1.
Setting
•
Normalize data to (0-0.25), Gaussian blur, to simulate spectrum addition
•
channel to a single value, which is then followed by a linear layer.
•
Logistic regression, BCE loss.
2.
Experiment 1: EMNIST.MNIST
2.1 Config
•
ENIST MNIST 0-9 training
•
Random r/2 + 0.1 on the clean and all mixting
•
White: Gaussian noise
•
Square: corrupted by small squares
•
Mix: A clean mixed with 1-3 images that are all in the target set
•
Scale: one corruption image * scale 1-3
•
Overlap: 1-3 corruption images overlapped
To make the description clearer, we define the following terms:
•
Mix-Aug: mixing like 'data-aug mix' in the previous experiment
•
Data-Aug: corruption with non-target noise and patterns
•
Noise-Aug: Simply non-target noise and patterns, hope the output being 0
•
Prob-Aug: Prob-mix as in the previous experiment