selvaraj-sembulingam / advanced-convolutions-and-data-augmentation

Advanced Convolutions and Data Augmentation

This folder contains an implementation of a Convolutional Neural Network (CNN) using some of the advanced convolutions and data augmentation techniques on CIFAR10 dataset.

Folder Structure

└── README.md
└── src/
    └── data_setup.py
    └── utils.py
    └── engine.py
    └── model_builder.py
└── models/
    └── S9Model1.pth
    └── incorrect_images.png
    └── loss_accuracy_plot.png
└── train.py
└── S9.ipynb

How to Run the code

Clone the repo and run Change your current directory to S9

python train.py

Receptive Field Calculations

	r_in	n_in	j_in	s	r_out	n_out	j_out		kernal_size	padding
Conv	1	32	1	1	3	32	1		3	1
Conv	3	32	1	1	5	32	1		3	1
Conv (Dilated)	5	32	1	1	9	30	1		5	1
Conv	9	30	1	1	11	30	1		3	1
Conv	11	30	1	1	13	30	1		3	1
Conv (Stride2)	13	30	1	2	15	15	2		3	1
Conv	15	15	2	1	19	15	2		3	1
Conv	19	15	2	1	23	15	2		3	1
Conv (Stride2)	23	15	2	2	27	8	4		3	1
Conv (DWS)	27	8	4	1	35	8	4		3	1
Conv	35	8	4	1	43	8	4		3	1
Conv	43	8	4	1	51	8	4		3	1
GAP	51	8	4	1	79	1	4		8	0

Convolutions

Normal Convolution

Strided Convolution (stride = 2)

Dilated Convolution (dilation=2)

Depthwise Separable Convolution

Data Augmentations Used

1. Horizontal Flip
2. ShiftScaleRotate
3. CoarseDropout

Training and Testing Results

• Total Parameters: 150,866
• Best Train Accuracy: 82.18
• Best Test Accuracy: 86.78

Key Results

1. Has the architecture to C1C2C3C40
2. Used Dilated kernels here instead of MP or strided convolution in first block
3. Total RF is more than 44 (79)
4. One of the layers must use Depthwise Separable Convolution
5. One of the layers must use Dilated Convolution
6. Used GAP and added FC after GAP to target #of classes
7. Used albumentation library and applied - horizontal flip - shiftScaleRotate - coarseDropout
8. Achieved more than 85% accuracy (86.78), in 50 epochs.
9. Total Params to be less than 200k. (150,866)
10. Created a modular code

Incorrect Classified Images

References

• https://towardsdatascience.com/a-comprehensive-introduction-to-different-types-of-convolutions-in-deep-learning-669281e58215
• https://www.learnpytorch.io/05_pytorch_going_modular/