Flowers102-Resnet50

FellowshipAi project

might takes some time to load, due to many plots and gifs, please be patient

The Notebook has everything from downloading dataset to fine tuning the model and inference.

Preprocessing

• download.py
• data_processing.py

Dataset

• Flowers102

Augmentation

• RandomResizedCrop
• RandomHorizontalFlip
• RandomVerticalFlip
• randomRotation
• CenterCrop

Trnasformations

• resize = 224
• ToTensor
• Normalize
  • mean = [0.485, 0.456, 0.406]
  • std = [0.229, 0.224, 0.225]

Model

• model.py

Architecture

• Resnet50 - backbone
• MlpHead - head
  • 3 layers
    • 2048 -> 512 -> 102 hidden units
    • ReLU activation
    • Dropout 0.5
    • BatchNorm

Train

• CrossEntropyLoss
• Adam optimizer
  • 0.01 learning rate
• 30 epochs

Fine-Tuning

• CrossEntropyLoss
• Radam optimizer
  • 3e-4 learning rate
• Cyclical Learning Rates
  • min_lr = 5e-8
  • max_lr = 3e-3
  • mode = 'triangular'
• 120 epochs

Results

1. EDA

Classest Distribution

• Class imbalance is present in the dataset

Sample Images

Training History

Highest Accuracy

• Test-Set

Experiment	Accuracy	Loss
Resnet50 - 30 Epochs	85.0838%	0.549521
Resnet50 - 120 Epochs	89.2956%	0.386665

• Validation-Set

Experiment	Accuracy	Loss
Resnet50 - 30 Epochs	87.2511%	0.451927
Resnet50 - 120 Epochs	91.0849%	0.312610

• Train-Set

Experiment	Accuracy	Loss
Resnet50 - 30 Epochs	88.8839%	0.389938
Resnet50 - 120 Epochs	96.7953%	0.105484

Graphs

LOSS

• 30 epochs
• 120 epochs (fine tuning)

ACCURACY

• 30 epochs
• 120 epochs (fine tuning)

GradCam

Train Sample

• grad camp for 1 image & top 4 classes

• CAM for all classes

Val Sample

• idx 75
• grad camp for 1 image & top 4 classes

• CAM for all classes

Test Sample

• idx 75
• grad camp for 1 image & top 4 classes

• CAM for all classes

Workflow