Problem formulation:
Classify photos at a gate to a facility. Identify cars/trucks and front/back of cars.
Weekend one: Manual classification
- Created a quick tool for manual data classification. (React+Konva and Sinatra+SQLite).
- Manually classified 1100 images.
- Gave up on http://mlcourse.ai. Not enought time to do the homeworks. Will concentrate on finishing the specialization on Coursera first https://www.coursera.org/specializations/deep-learning.
Weekend two: Initial learning
- Created Jupyter notebook where data from manual classification (csv) are loaded,
- Loaded image data to the Jupyter notebook, display data, etc.
- Copied several CNN models in Keras, tried one by one.
- Trained a model with ~15 Keras elements, convolutional layers, dropout, cathegorical cross-entropy cost function, used Adam optimizer with default values for hyperparameters.
-
Accuracy train/test >90/>60, large variance. Seems the capacity of the model is more that sufficient. Probably need more training data.
-
Inspiration: take a look at keras image processing and feeding into model for training. Carry on with Coursera Sequential models course.
Week three: Improvements
- Prepared (a bit) more of training data.
- Looking at how to deploying models as Python service: Flask
- Experimantaion with GPU learning nvidia 1050, 2GB RAM, out of memory even with bach_size = 1 on my home desktop.
Should I consider some (systematic) hyperparameter search? Or just get more data?
Experiment with varying “decay” for Adam
To address oscilation in test set accuracy I tried to limit the learning rate by the learning decay parameter passed to the Adam optimizer.
Default setting: decay=0
Significant oscilation of the test set accuracy.
decay=0.01
The oscilation is limited, but the train accuracy dropped significantly. Test accuracy did not improve compared to the best values from the previous experiment.
decay=0.005
With this middle setting the oscilation was minimized and at the same time both train and test accuracy are slightly better.
