- Kunal Bharadwaj - D19016
- Pooja More - D19021
- Veera Vignesh - D19036
- Vighnes Tamse -D19037
- DurjayDas - D19038
We used CIFAR-10 dataset for our experiments with ResNet Architecture. Experiments where conducted with base model as ResNet-18 to tune various parameters and the best hyperparameters where then progressively applied to deeper models.
EDA of the data was done to understand the various class of the image and effect of transformation on the image.
Hyper Parameters:
- Batch Size - 32,64,128,256,512
- Epoch - 10,20,40,60,80
- Optimizer - Adam (with & without weight decay), SGD (with & without momentum and weight decay)
- Transformation - RandomCrop,Random Rotation, Padding, Colorjitter, Normalization
- Learning Rate - Dynamic learning rate
Experiments conducted are available in the folder. Based on the experiments HyperParameters where choosen for further experiments
| Model | Accuracy |
|---|---|
| ResNet-18 | 91.07% |
| ResNet-34 | 91.96 |
| ResNet-152 | 87.01 |