The objective of this part is to test different architectures to compare two digits visible in a two-channel image.
It aims at showing:
-
The impact of weight sharing,
-
The use of an auxiliary loss to help the training.
It is implemented with PyTorch only code, without using other external libraries such as scikit-learn or numpy.
Implemented a deep network such that, given as input a series of 21414 tensor, corresponding to pairs of 14 * 14 grayscale images, it predicts for each pair if the first digit is lesser or equal to the second.
The goal of this project is to design a small deep learning framework without using "torch.nn", autograd and over advance lybrary.
We want to create a module that given a points in [0,1]^2 should determine if the points is in the disk of center (0.5,0.5) and radius 1/sqrt(2*pi).
The network should have 3 hidden layer of 25 units and can work with 2 possible activation function (ReLU and Tanh). To train the network, we're using stochastic gradient descent.