This is the code for this video on Youtube by Siraj Raval on medical classification and a fork of his original work here.
The only addition made thus far is the addition of a Go-based image resizing tool that is able to use all cores of a given host machine/VM/container to perform the TTA.
This is performed through a simple approach of splitting the list of images into N-groups, where N is the number of logical cores assigned to the host machine as determined by the Go native library function runtime.NumCPU(). (see NumCPU()). Once the image file list is broken up into 'N' groups, 'N' go routines are started to process the images.
The Go routines are all synchronized via a Go waitgroup so that all of the image processing work is completed before the process is terminated.
When attempting to run the original resizing code on the large set of images, I realized how inefficiently the code was using the AWS EC2 instance on which I was running; despite the EC2 instance having 8 cores and sufficinent memory/storage/etc., the Python image processing script was only using a single core, which meant a waste of money for anything more than a single core EC2 with enough memory to sustain the processing.
For the Go version, while not a perfectly linear scaling, the resizing code was able to fully utilize the 8x vCPU EC2 instance on which I was testing vs. only getting 1 vCPU of usage with the original Python implemention.
I would like to work with other folks in the data science/ML/AI community on similar efforts if anyone is interested.