Torchvision Experiments

In this repository, I have actually experimented with different transformations and experimented them on CIFAR10 dataset using torchvision library.

Image Transformations

I have tried Image transformations on CIFAR10 as well as 3 images taken from the internet. I tried the following transformations:

1. A RandomHorizontalFlip
2. A RandomRotation
3. ColorJitter (tested different brightness, contrast, saturation and hue)
4. Grayscale

You can view the Notebook here

CIFAR10 Experiments

Using the knowledge from above experiment about Image transformations, I did another experiment with the CIFAR10 dataset. You can find the Notebook here

Data transformations

We will try different image transformations with our data and prepare a detailed table of comparison on accuracy, speed, and efficiency with each set of transformations.

1. ColorJitter: We will vary the brightness, contrast, hue and saturation of an image and see which looks best.
2. Rotation: We will try to rotate the images randomly, and look at the model accuracy in test cases.
3. Grayscale: We will also try to randomly grayscale some of our data and see if our model can distinguish easily.
   We will try to experiment with 5 set of image transformations, (No Transformation, ColorJitter, Rotation,Rotation+Grayscale, and Rotation+ColorJitter).

Model Architectures

We will also experiment with different model architectures. (All we will use are pre-trained).

1. Our all-favourite ResNet50.
2. Another ResNet34.
3. ResNeXt50: Born out of Facebook AI and UC Diego.

Hyper Parameters.

1. Batchsize: 64/128
2. CyclicLR
3. Optimizer: Ranger

Conclusion

The Results table.

Architecture (batch size)	No Transformation	ColorJitter	Rotation	Rot.+Color.	Rot.+Grayscale
ResNet50 (128)	99.282%	72.542%	-	-	-
ResNeXt50 (128)	98.942%	-	-	-	-
Rot. = Rotation
Color. = ColorJitter

1. Ranger Optimizer just worked wonders and whithin 5 epochs got a 99%+ accuracy on ResNet50.
2. ResNet50 scores slightly better than ResNeXt50 on our dataset (in 5 epochs) but ResNeXt50 takes considerably more time for training than the former.

Model	Training Time	Utilized time %
ResNet50	18m 34s	70.64%
ResNeXt50	26m 17s	100%

ResNet50 performs slightly better while taking 30% lesser time than ResNeXt50.

Therefore, it will be better for us to use ResNet50 over ResNeXt50 because of time constrains.

Thank you so much for your time looking at my repository!