Convolutional neural network for Cifar10 dataset, built with PyTorch in python
Train and test several CNN models for cifar10 dataset.
The CNNs overall performance can be evaluated with this Python script. Therefore, the entire test set will be forward passed through the network and the predictions are compared to the labels of each picture. The evaluation tool states, how well the network performs in each category.
The CustomImageClassifier.py script can be used to make predictions for custom pictures using the trained CNN. Any JPG image can be loaded and a probability for each of the ten classes is calculated an printed. The classifier outputs, which object is most likely in the image.
- three convolutional layers
- two fully connected layers
- four convolutional layers
- two fully connected layers
poor performance compared to other models
- five convolutional layers
- two fully connected layers
best performance among all models (84,4% accuracy on the test set)
- six convolutional layers
- two fully connected layers
Training process
- data augmentation: affine transformations (rotations, translations), random b/w images
- regularization and dropout
Built with python 3.7.4 (Anaconda). Requires PyTorch and MatplotLib.