I created this project to become familiar with the ResNet structure; in particular, I ultimately create a ResNet50 model (as seeen in the ResNet50 notebook) I then use the network to train on the dataset for a Kaggle Competition to try and learn which camera model was used to take a particular photograph.
The data comes from two sources, which are explained in the More Data notebook (from Kaggle and scraped from Flickr). In my implementation, I perform the same types
of manipulation as described in the competition (resize, compression, gamma adjustment) and then crop the images to be 512x512x3. As I don't have a whole lot of system
memory, I introduce a load function that will load a random subset of the data based on a ratio parameter to be used; this is repeated for each training epoch.
The ResNet50 structure is followed pretty closely, with only the final fully-connected layer being different. Since the network only accepts 224x224x3 images, I sample each of the training images via a center crop and then a few more times from random positions. The network was then trained for roughly 30 epochs of ~20k images.
Ultimately, I only achieved about a 50% test accuracy. I did not have more time to explore why, but I suspect a few reasons:
- Errors in my data generation
- Not enough training time
- No data pre-processing Really, I likely would've gotten further had I taken the time to build out a confusion matrix for both altered and unaltered images to see where the biggest issues were.
Other than the More Data and ResNet50 notebooks, there are 5 notebooks; they are marginally functional at best, but I included them as they track my progress leading
up to the final two notebooks.