This week

• [x] Set up Tensorboard for tracking
• [x] Implement a simple Train-Evaliation-Testing template
• [x] Use our synthetic dataset to check if a demographic model is better than a global one
• [ ] Test some model implementation with synthetic data

Day 2

In order to make the experimenting easier, I'd like to set up Tensorboard to track the performances, and also implement the Train-Evaluation-Testing chain in TFF.

So we need to investigate some aspects:

• [x] Save model checkpoints during training
• [x] Save metrics using TB
• [x] Perform Evaluation on the validation set at each round
• [x] Perform testing at the end of the training

To do this i will work on the previous Colab notebook:

Google Colaboratory

Checkpoints

Unfortunately with TFF models we can't use the common Checkpoint class we used with Keras, the state returned by the iterative_process is not trackable.

On Stackoverflow I found this answer: https://stackoverflow.com/questions/58785825/how-to-save-model-in-tensorflow-federated

But to use this method we need classes defined in tff.python.research..., which is not included in the pip package.