pneumonia

Tensorflow Dataset API
-------------------------

The tensorflow dataset API helps us in writing custom and efficient input pipelines. 

Dataset API is used for: -

$ Creating a Stacked object for Input data and labels.

$ Applying transformations to the data.

$ Iterating over the object to fetch the data

ResNets
-----------------

It was the time when ‘AlexNet’ was released and people were just trying to build bigger and bigger networks. 

Bigger networks were actually getting better accuracies but there was a limit and after that the errors went higher i.e. no improvement in performance

Issues using Plain Convolutional Networks
------------------------------------------
Validation errors for two plain convolutional networks
18 layers.
34 layers.
The error for a 34 layers network is higher compared to 18 layers.

Adding more layers can lead to more complex learning functions which might lead to an overfitting situation

Deeper networks might face an issue backpropagating the gradients, this issue is most commonly called the ‘Vanishing Gradients’ problem.

solution
----------
Overfitting can be handled by using regularization and dropout where we deactivate few random neurons to help other neurons learn. 
‘Vanishing Gradients’ problem can be handled using ‘ReLu’ activation which maximizes the gradient flow, we can use ‘Batch Normalization’ to address this issue as well.

ResNet Background
-------------
“A deep network should perform at least equal to its shallow counterparts.
Consider a Network where Initial layers are the shallow network and remaining layers are Identity functions.”

Deeper ResNet
------------------
They are very similar to each other except for the skip connections part.

Plain Networks vs ResNets
-------------------------------
There is not much difference in validation errors of plain 18 layers and resnet

Difference starts showing up when we use deeper models like resnet 34 and resnet 50

Validation error for 34 layer resnet is significantly lower compared to its plain counterpart.

Difference will keep increasing with increase in the number of layers