Implement an architecture with CNN and LSTM

JankowskiDaniel / Neural-Deep-Retina

0 stars 0 forks source link

Implement an architecture with CNN and LSTM #15

Closed kapiblue closed 1 month ago

JankowskiDaniel commented 2 months ago

The first experiment shows that the model can overfit well on the training set, however, at the same time is lost in the testing data. Training outputs_and_targets Testing outputs_and_targets_test

JankowskiDaniel commented 2 months ago

A few general observations:

StandardScaler() works better than MinMaxScaler(),
MinMaxScaler() (range [0,1]) was tested also with Sigmoid() on the output, but something was not working correctly (only 0 predictions)
No surprise, the longer the sequence for the LSTM, the better results,
Sometimes, for different channels, the predictions on the test set are very similar (look at the picture, channel 7 and 8) The picture is from the test set, the model trained with sequence = 16.

kapiblue commented 2 months ago

Two questions:

which timestep did you predict? The last of the sequence of images or the following one?
for how many epochs did you train the model?

JankowskiDaniel commented 2 months ago

I tested both options: predict the following value, and in the second configuration, predict the last value of the provided sequence. Both configurations produce very similar results, as presented above.
Each model in my configuration was trained for 100 epochs.