In this task, you should train Inception V3 with pytorch using transfer learning to classify CIFAR10 dataset.

Load the dataset¶

Load the CIFAR10 dataset bellow. You can use either torchvision.datasets.CIFAR10 or sklearn.datasets.fetch_openml() or any other way to load the dataset.

Get Inception V3¶

Instantiate Inception V3 model (pretrained on imagenet) from torchvision's model zoo.

Modify for CIFAR10¶

Imagenet has 1000 classes, but CIFAR10 has 10. So, the decider layers of the model must be changed so that we can use the model for CIFAR10. Therefore, adapt net.fc and net.AuxLogits.fc with CIFAR10.

Freeze the convolutional part¶

In order to apply transfer learning, freeze all layers except the deciders. Freezing consist of disabling optimization by disabling grad calculation. Also in batch normalization layers, updating mooving average and variance must be disabled. Note that you must later filter out the frozen parameter for optimizer.

Instantiate the optimizer¶

Create the optimizer filtering out freezed parameters.

Train the model¶

Write your train/validation loop bellow. Then train the model until it converges. Feel free to add extra cells.

Draw the training curves¶

Draw two diagrams for train and validat ion, one showing loss of each epoch, and another showing accuracy of each epoch.

Evaluate your model¶

Evaluate the best epoch's model (according to the validation accuracy) on the test set, and report the accuracy. Is your model overfitted?

Draw misclassified images¶

Draw 20 misclassified images from test set with expected and predicted labels.

Plot the confusion matrix¶

Plot the confusion matrix for the test set.