CNN using tensorflow

@dkilike

Hi I am a full-time professional Deep Learning engineer from New Zealand. My job is to develop many CNN-based software for different medical image classification and segmentation. I believe my experience could help you increase the accuracy of your project. Sorts of things you may want to try (currently in top of my head): different architectures of CNN, image pre-processing (including image enhancement), data augmentation, fine-tuning hyper-parameters, different loss functions, etc, etc. More ideas will be raised if necessary. The major reason I apply for this job is that I like practicing my neural network skills in different fields with real-world dataset, outside my work hours. Here is a link of my work style and skills showcase: [login to view URL] Due to the medical ethics, I cannot showcase my real work project. BTW, I am living in New Zealand. This means that I work while you are sleeping in Germany. When you wake up everyday, you can find my daily work progress. This builds a nice 24-hour team.

@CROlakiluk

I am interested what images you are trying to classify, how much of them is there and how good are they classified. 80% is actually not high goal in modern Deep Learning CNN based models, but I guess that accuracy here more depends on dataset quality, rather than a model or training as such. I guess that model should be fine tuned (hyper-parameters finding) to increase accuracy, but it is also possible that cleaning training dataset could also help some.