cranjis-mcb / vision_architectures

Pytorch Implementation and Performance Analysis of the Popular Vision Architectures from Scratch.

VISION_ARCHITECTURES

1. Implementaion of the Popular Vision Architectures from Scratch.
2. Performance Comparison with the Pytorch Standard Architectures.
3. Implementation of the Popular Vision Data Augmentation Techniques from Scratch.

Implemented Architectures as of now

1. ResNet (Versions - 18, 34, 50, 101, 152) ✅
2. EfficientNet (Versions - b0-b7) ✅
3. EfficientNet-V2 (Versions - S, M, L) ✅

cnn_architectures.ipynb file

You can import the mentioned CNN Architectures from this file.

How to Use?

from cnn_architectures import ResNet18, ResNet34, ResNet50, ResNet101, ResNet152
from cnn_architectures import EffNet
from cnn_architectures import EffNetV2

# Initialize the model.
resnet = ResNet18(img_channel=3, num_classes=1000) # Resnets

version = 'b0' # b0-b7
effnet = EffNet(version, num_classes=1000) # Efficient-Nets

version = 's' # s, m, l
effnetV2 = EffNetV2(version, num_classes=1000) # EfficientNet-V2

Data Augmentations

This folder consists of the Pytorch Implementation of the Popular SOTA Vision Data Augmentations techniques.

Implemented Data Augmentations as of now

1. Cutout
2. Mixup
3. CutMix
4. AugMix
5. GridMask