May

Part 1

Got tensorflow 2 running (Fedora 29). Tried to get tensorflow 2 running also on windows, but so far w/o success. Notes: need 2 packages from NVidia, most likely Visual Studio run-time, python 3.6 (not 3.7 for now it seems).

Created a script for learning using Keras ImageDataGenerator, splitted data into directories by class and fixed a number of misclassification. With these changes it seems that I am getting much better results. (surely over 80, maybe 90% validation accuracy, need to properly measure to make sure).

I also created a script that utilizes Talos for hyper parameter seach. Need to run it with the best model and interpret the results.

Looking at 2 Coursera courses on Tensorflow 2. The theory is known to me and so far pretty shalow. But the hands on with Ternsorflow 2 and various interesing datasets are being introduced, that is interesting. Introduction Tensorflow and Conv nets with Tensorflow.

Part 2

Influences by the Tensorflow courses I migrated the models I have from Keras with Tensorflow 1.x backend to the new Tensorflow 2.0 Keras interface. Got learning running on Google Colab using ImageDataGenerator.

Following https://www.tensorflow.org/tensorboard/r2/hyperparameter_tuning_with_hparams I created a notebook that uses Grid Search (really a bunch of nested for loops) to run 432 experiments searching for the model hyperparameters.

The experiments showed that with the 3.500 training examples verious (parametrized) versions of the model can get above 80% test accuracy in under 20 epochs.

Next steps

• Interpret results from latest experiments,
• analyze missclassified examples (so far I postponed this because I found that I have errors in the training data, so I first fixed that).
• Deploy as a service.
• Ddd Talos to the Colab notebooks to play with more interesting hyperparameter search techniques.