Flowers-Classification

Dataset: Oxford 102 Flowers

This dataset includes 8189 images of 102 (0 to 101) categories of flowers.

Of these,

Number of images in training data = 1020
Number of images in validation data = 1020
Number of images in testing data = 6149

Task

The task of this project is to use ResNet18 on Flowers Dataset for classification followed by GradCAM to get visual explanations of our model. Later, without making changes to epochs, improvise the model results.

Navigation

Flowers_ResNet18_V1.ipynb:

1. No Data Augmentation
2. Simpler model:

FFClassifier(
  (fc1): Linear(in_features=512, out_features=102, bias=True)
)

Flowers_ResNet18_V2.ipynb:

1. Data Augmentation techniques like -
   • Rotation by 45 degrees
   • Crop from original images and obtain subplots
   • Horizontal Flip
2. Deeper model:

FFClassifier(
  (fc1): Linear(in_features=512, out_features=4096, bias=True)
  (fc2): Linear(in_features=4096, out_features=102, bias=True)
)

This is done in order to improve our results. The training set has less number of images. Upon data augmentation, we generate more images to be used as training data. Also, deeper models might take longer to train every epoch but has the ability to learn better.

Accuracy

Accuracy for V1 = 0.7660369251329295
Accuracy for V2 = 0.8159778965354642

Weakest Class

As observed in both Flowers_ResNet18_V1.ipynb and Flowers_ResNet18_V2.ipynb, the weakest class is -

Class number = 2
Flower name = Canterbury Bells

GradCAM

Results obtained after applying GradCAM are stored in folders cam_V1 and cam_V2 respectively.

For an observation of class 2 which was predicted incorrectly for the simpler model (Flowers_ResNet18_V1.ipynb) and correctly for the improved model (Flowers_ResNet18_V2.ipynb), we can see the GradCAM results below:

Future Work

This work has shown great potential for a number of use cases. Some avenues to pursue ahead can be:

1. Data
   • Gathering more images for each flower type. This will increase training data and can lead to a better model.
   • Using Generative Adversarial Network (GAN) variants like Data Augmentation GAN (DAGAN) for generating synthetic training data.
2. Deep Learning Models
   • Making use of models like ResNet152, VGGNet, etc and compare the results.
3. Interpretability
   • Using approaches like Local Interpretable Model-agnostic Explanations (LIME) and Layer-wise Relevant Propagation (LRP) for interpreting the models.

References

1. Problem Statement by Fellowship.ai
2. Dataset: https://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html
3. Data Augmentation: https://discuss.pytorch.org/t/data-augmentation-in-pytorch/7925
4. Transfer Learning: https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html
5. ResNet18: https://medium.com/@josh_2774/deep-learning-with-pytorch-9574e74d17ad
6. Labels: https://gist.github.com/JosephKJ/94c7728ed1a8e0cd87fe6a029769cde1#file-oxford-102_flower_dataset_labels-txt
7. GradCAM: https://github.com/Caoliangjie/pytorch-gradcam-resnet50/blob/master/grad-cam.py