All code for this assignment can be found here.

Question 1 - Aj R Laddha

In this question we were asked to make a dataset consisting of lines varying in features like length, width, colour and angle which made a total of 96 classes.

Feature	Values
Length	2 - long and short
Width	2 - thick and thin
Colour	2 - blue and black
Angle	12 - 0-180 in steps of 15

The images (28x28 px) had to consist of these lines on a black background. The intraclass variation was induced with translation of the line. Having 1000 images in each class the dataset had a total of 96000 images. A video displaying all the images in a 9x9 grid was also made.

Question 2 - Kaustubh Verma

Question 3 - Nikhil T R

Description

In order to fully understand a fully connected layer, in this problem we are asked to create a dense layer api on top of lower level tensorflow functions like multiply. From the very beginning I’ve been complaining about having to work on tf1 when tf2 is being released, especially when tf1 api is being ridiculed by many.
However, I am ambivalent. I also believe that doing this will help me appreciate deep learning on a deeper level (lol). The underlying concepts will not change, and this should be my only focus.
Using this api, I am asked to make a model to classify a given MNIST data set and the dataset that we’ve made in Question 1.

My work

The api that I’ve designed basically uses everything from tensorflow except the layer api. My layer api can be found here here. I implemented only dense as that was the only requirement. The dense function creates a new tf.variable_scope for a dense layer. This gives me the added advantage of naming weights as w and bias as b everytime. It also accepts values for the standard deviation of the normal distribution from which the Tensors are initialized. For usage, refer to the notebooks here