Used Label: Randomly Images are flipped and labels for the flipped images are 1, otherwise 0.
Used Library: Pytorch
Used Model: ResNet18 from torchvision
Number of classes: 2 (0 means No flip and 1 means Flipped image)
Train Test Split: 70-30
Some sample Images and Labels from created dataset
Alt sample image png
Specific information about the Dataset:
The following files from the dataset were removed during the training because those files are broken/ corrupted.
./kagglecatsanddogs_3367a/PetImages/Cat/666.jpg
./kagglecatsanddogs_3367a/PetImages/Cat/10404.jpg
./kagglecatsanddogs_3367a/PetImages/Dog/11702.jpg
Training Information:
Total Epoch: 5
Pretrained: True (ImageNet weight) (Every layer is trainable)
Image Size: 224 x 224
Batch Size: 128
Optimizer: SGD
Learning Rate: 0.001 (Constant throughout the training)
Momentum: 0.9
Loss: CrossEntropy Loss
Result:
Accuracy: 98.4266
F1: 98.4271
Recall: 98.4261
Precision: 98.4265
Confusion Matrix:
Alt confusion matrix png
Some Misclassified Images (Randomly Selected):
Alt misclassified image png
Some possible improvements:
Most of the misclassified images are occluded by some other objects or partly visible. One possible improvement could be to improve this type of image in the training dataset.
Hyperparameter tuning is another option, we could try to see whether the performance improves or not. For example, instead of using a constant learning rate, we could try a cyclical learning rate. This type of learning rate helps the model overcome local minima.
If we consider the rightmost image in the above figure, we see that the pose of the cat is different than most of the images in the training set. I think Augmentation like CutMix will be helpful in this scenario.
