-
1st Part - Built from scratch the following types of layers: 2D convolution, max pooling, and linear. Also constructed 2d batch normalization
-
2nd Part - Implemented a ResNet-18 architecture to perform classification on the CIFAR-10 dataset. The final model achieved a ~87% accuracy
In addition to the code, Deep_Learning_CNN.ipynb answers the following questions:
- Given such a network with a large number of trainable parameters, and a training set of a large number of data, what do you think is the best strategy for hyperparameter searching?
- Compare the feature maps from low-level layers to high-level layers, what do you observe?
- Use the training log, reported test set accuracy and the feature maps, analyse the performance of your network. If you think the performance is sufficiently good, explain why; if not, what might be the problem and how can you improve the performance
- What are the other possible ways to analyse the performance of your network?