Everyday new datasets show up on my feed that are interesting. Yesterday a new beauty dataset surfaced out of China SCUT-FBP5500 where they had made the dataset public. Wouldn’t it be fun to take a pass on this data? Could I beat the original researchers on their accuracy? Could I do all of this in less then 10 minutes of my time?
Deep learning is black magic to most people. The frame works we pick (i.e. Tensorflow, Keras, Theano, pytorch, Caffe2, etc…), the models (resnet152, ResNext, ResNextBestDense?, etc…), the training strategy, the GPU dependency stack, the data augmentation strategy, the learning rate decay, the optimizer selection…. the whole process is a mess. A mess of confusion between you and Getting Stuff Done (GSD).
At Ziff we have realized that all deep learning frameworks fall short and that the majority of people out there aren’t experienced enough to make all of the other decisions and optimizations around this problem. So we have automated the entire process from problem definition (Excel/CSV/SQL) to training, validation, and deep net deployment. No custom programming required, no parameter tuning, no trial and error. We want to react to our incoming data and not the process between data and value.
SCUT-FBP5500 Data Preview:
Baseline To Beat:
These researchers tried a variety of deep nets and based on the list of networks they tried they were not clueless with their approach. I would say this represents most people doing research with deep learning, somewhere between not-clueless and expert.
So the baseline to beat on unseen data (validation set) is r=0.8997 for beauty. Another thing I would like to do before we start is percentile norming between the man/woman/race subgroups, so our beauty scores are percentiles for the specific group (i.e. Asian Women). Here is a little python script that makes that fast and easy:
Before we pull the trigger I actually want to include gender prediction and race prediction in the same model. Those are interesting things to predict as well so why not. This information was given to us in the filename so I can pull that out and make a CSV that now has all three columns:
Typically, when you combine outputs you actually lose accuracy from competing objectives, so we might be shooting ourselves in the foot but here goes.
Submitting the CSV into our genetic boosting network creation pipeline produces validation scores of:
beauty r_value = 0.90721756
gender_accuracy = 98.868%
race_accuracy = 98.913%
So not only did we beat their r-value but we also included very solid gender and race predictions on the same model. Another bonus with our current state is the final deployment is already taken care of, take your pick, standalone Docker, iPhone (coreML exports), or the Ziff inference cloud with enterprise level SLAs.
What do you guys think about this beauty dataset that was released? Weird? Useful? Just because we can…. how about we don’t?
Beating Deep Learning Beauty Researchers With Excel was originally published in ziff inc on Medium, where people are continuing the conversation by highlighting and responding to this story.
Source: Deep Learning on Medium