GiteonCaulfied
commented
1 year ago Hi @amartinhuertas @rhyshawkins
I've updated the auto-encoder by removing one convolutional layer and feed it with all 10000 temperature fields. The size of the latent space is now increase to 6x23x45, which is about 13 times smaller than the size of the original input (1x201x401). As expected, the ConvAE is now able to capture more features and the performance is better than the last commit (less spiky edges for the 10-region-color-map).
One problem is that the training now takes about 2-3 hours to complete since I use all 10000 temperature fields to do the training/testing/validation. (I stopped the training process manually after 180 epochs since it took too much time) So I may have to use Gadi if we want to test with a larger dataset.
sghelichkhani
amartinhuertas
Hi @amartinhuertas
In today's meeting we've mainly discussed about two things: implementation of the current auto-encoder and potential ways to build the neural network to predict the temperature field for the next time stamps.
For the auto-encoder, the current one doesn't perform quite well when it comes to some "complex" temperature fields. Therefore, Rhys suggested to increase the size of latent space by either removing one convolutional layer or decrease the stride value so that the latent space should be able to capture more features of the original temperature field.
Also, the current auto-encoder is only trained for those temperature fields that have a specific pair of consecutive timestamps. (So only 160 input for the training set in this case)After the discussion, we decided that it would be better if we feed all the temperature field (all 10000 of them) as the training/testing/validation data. In this way, we no longer need to build 99 auto-encoders when it comes to 100 timestamps and we can just use one auto-encoder to do all the work. I will test both these things before our next meeting.
For the neural network to predict the temperature field after we transform the input temperature field using auto encoders, there could be two ways of doing this. One is to train multiple NNs and make each of them specifically working for a single set of consecutive timestamps (one NN for 0->1, one NN for 1->2, ...), the other is to only train one huge NN to do all this work (same NN for 0->1 and 1->2 and ...). We've discussed about this in our last meeting and decided to use the first method two weeks ago but Rhys said he preferred the second method in today's meeting. Let's leave the final decision to the meeting on next Monday. In the meantime, I will just focus on the implementation of the auto-encoders for now.