I worked on google colab, so the downloaded dataset and model checkpoints are saved in the drive.
Task 2
-
Use this dataset (https://www.dropbox.com/s/pan6mutc5xj5kj0/trainPart1.zip) to train a CNN. Use no other data source or pretrained networks, and explain your design choices during preprocessing, model building and training. Also, cite the sources you used to borrow techniques. A test set will be provided later to judge the performance of your classifier. Please save your model checkpoints.
-
Next, select only 0-9 training images from the above dataset, and use the pretrained network to train on MNIST dataset. Use the standard MNIST train and test splits (http://yann.lecun.com/exdb/mnist/). How does this pretrained network perform in comparison to a randomly initialized network in terms of convergence time, final accuracy and other possible training quality metrics? Do a thorough analysis. Please save your model checkpoints.
-
Finally, take the following dataset (https://www.dropbox.com/s/otc12z2w7f7xm8z/mnistTask3.zip), train on this dataset and provide test accuracy on the MNIST test set, using the same test split from part 2. Train using scratch random initialization and using the pretrained network part 1. Do the same analysis as 2 and report what happens this time. Try and do qualitative analysis of what's different in this dataset. Please save your model checkpoints.
Progress for task 2- part 3 Planned how to do it and downloaded the files