eilertskram/cnn-ablation-study

CNN-ablation-study

Objective

Data

Famous dataset used for vision learning, PETS. Images of varying sizes of different animals The images are named with the spesific animal's name and race(Model's target label), i.e cocker_spaniel_19.jpg

Dataloader

Resize

To save on time and computing, allowing to run multiple models on more epochs

Hypothesis: Squish will perform well on square images

Augmentation

Using crop and resizing method as mentioned before

Using btch_tfms to transform the pictures with a multiplier from 0 to 2.0

More info about augmentation, a good notebook is Introduction to Image Augmentation using fastai by Sanyam Bhutani

Dataloaders with augmentation multiplier with values 0.1, 0.5, 1.0, 1.5 and 2.0

Using Resnet18 with 10 epochs

After 5 epochs augmented data performs better

Architecture

Used Jeremy Howard's Notebook to find different architectures to try out

Limited which models avalible due to computing power and image size

Exploring the different levels of resnet-architectures, but including one model from different architecture

Surprising how bad efficinet is performing

The ablation study

Resnet26

From the early testing, resnet26 seems to perform the best. Will be chosen as the benchmark.

Augmentation impact

Resize 80

Impact does not seem that significiant. Suspecting its due to image size, so limited how much levels of augmentation and change there is to be had

Resize 128 lowering architecture

Codeblocks are the same for the resnet18 model, with 'resnet26' replaced with 'resnet18'

resnet26	resnet18

Dihedral