The model reaches a maximum accuracy of 91.53% in 41 epochs on CIFAR-10 dataset using ResNet-18 model.
Gradient-weighted Class Activation Map (GradCAM) was implemented for each convolution block to generate model prediction heatmaps (Examples shown below).
- Loss Function: Cross Entropy Loss (combination of
nn.LogSoftmax
andnn.NLLLoss
) - Optimizer: SGD
- Learning Rate: 0.01
- LR Step Size: 25
- LR Gamma: 0.1
- Batch Size: 64
- Epochs: 50
The following data augmentation techniques were applied to the dataset during training:
- Horizontal Flip
- Rotation
- CutOut
The albumentations
package was used to apply augmentation.
Some of the examples where the network was focusing while predicting the output is shown below:
Install the required packages
$ pip install -r requirements.txt
Upload the files in the root folder and select Python 3 as the runtime type and GPU as the harware accelerator.
- Shantanu Acharya (Canvas ID: 25180630)
- Rakhee (Canvas ID: 25180625)