Previous weeks - Testing MODELED samples

1 minute read

Testing networks

So now we have the generated dataset and the network model trained with that data, we need to test it and see if our model is valid. For I have been using the scripts main_test.py and net_test_config.yml located in link, very similar to train scripts.

Basically, what the test script do is the finction:

to_test_net.test(testX, testY, gap, data_type, dim)

where:

textX -> Array of 20 values (x,y) textY -> Array of 1 value (x,y) gap -> Gap between the last value of testX and testY, meaning how many frames after is the network going to predict the pixel values. data_type -> raw or modeled dim -> dimensions of the frame, the default I have been using is 120x80

First I started with simple modeled images around 2000-4000 samples (80% training, 10% test, 10% validation) with DOF 1 and 2 (linear and parabolic)

The results differ a lot from the ones Nuria achieve. I was getting around 3% - 6% mean squared error for LSTM1 with 2000 samples and her results were around 0,1% - 0,3% The problem was that I was using only a few samples comparing with her (2000 vs 20000, 10000 vs 100000) and also some configuration wasn’t working properly, like batch size, dropout and I was suffering overfitting in some cases.

For next week I have to take the appropriate changes and get the right model to predict my frames in real life. I planned a meeting with her to get all the info neccesary.