Embedding training parameters - max_steps about multitaskais HOT 1 CLOSED

Hi,

1. On the Brittany (Ushant) dataset, the Embedding layer converges around -5.2, after ~480000 time steps.
2. No we didnt let it train until max_steps. The purpose of setting max_steps = 1e9 is to let the Embedding layer runs "forever", we manually stop the training when we observe that the model converge.

We havent tested the Trajectory reconstruction module on the Brittany dataset yet. The most important of the proposed model is anomaly detection. We spent a very little time on trajectory reconstruction and vessel type identification. These two tasks were added in the paper to demonstrate that with the proposed architecture, we can easily add any task-specific sub-module to the system without any effort.

You can notice that the Trajectory reconstruction module was poorly implemented. We believed that there is much room for improvement there, especially how to embed/decode the "four-hot vector".

from multitaskais.