12 - Tiny-ImageNet and YOLO v2 Anchor Boxes

Session 12 - Tiny ImageNet and YOLO v2 Anchor Boxes

Part 1 - ResNet18 on Tiny ImageNet Dataset

The model reaches a maximum accuracy of 55.42% on Tiny-ImageNet using ResNet 18 model.

Parameters and Hyperparameters

• Loss Function: Cross Entropy Loss (combination of nn.LogSoftmax and nn.NLLLoss)
• Optimizer: SGD
  • Momentum: 0.9
  • Learning Rate: 0.01
• Reduce LR on Plateau
  • Patience: 2
  • Factor: 0.1
  • Min LR: 1e-6
• Epochs: 50
• Batch Size: 128

Data Augmentation

The following data augmentation techniques were applied to the dataset during training:

• Horizontal Flip
• Vertical Flip
• Random Rotate
• CutOut

Change in Training and Validation Accuracy

Part 2 - Finding YOLO v2 Anchor Boxes

Finding anchor boxes on a dataset of 50 dogs using K-Means Clustering ALgorithm. The dataset was annotated and exported in COCO JSON Format. It can be found here.

Plot of Number of clusters vs Mean IoU

After running the algorithm on the dataset, it was found that the best k can have the value of 3 or 4.

Number of Clusters (k)	Mean IoU	Cluster Plot	Anchor Boxes
3	0.75
4	0.79

Project Setup

On Local System

Install the required packages
$ pip install -r requirements.txt

On Google Colab

Upload the directory tensornet and the file dogs.json in the root folder and select Python 3 as the runtime type and GPU as the harware accelerator.

Group Members

• Shantanu Acharya (Canvas ID: 25180630)
• Rakhee (Canvas ID: 25180625)