THIS CODE IS NO LONGER MAINTAINED
Based on the CIFAR10 example on TensorFlow. Reconfiguring it to run my own images and data.
I removed the "LICENSE HEADERS" in the code because it was getting kinda cluttered. My apologies if I violate any license laws. I'm new to this open source thing, so drop me an email at samuelchin91@gmail.com and tell me that I'm doing it wrong and this is not allowed. This code is for personal use and nothing else.
Notable Improvements
- Refactored all flags to be put in python flags file.
- Made a small hack to evaluate the test set sequentially (we lose out on speed greatly). This way, we can single out those images that have been classified wrongly. Still imperfect and a work in progress.
- Previous CIFAR10 evaluation code runs the test set in batches. In my opinion, this is unnecessary. We only want to run every image in the test set once.
- Refactored code to allow for 3 channel and 1 channel images, or X channels for that matter, since it's just a variable.
Successful Runs:
- Street View House Number Data Set from http://ufldl.stanford.edu/housenumbers/
- After training, I cropped my own images in. 100% accuracy. Wow!
- My self-engineered Wally data set.
- 100% accuracy too, but the data set creation is off.
- Assortment of others that yield promising results.
P.S. I apologize for messy code. This is experimental. Trying to hack stuff together to make it work. Expect a refactored general purpose module soon!