search

nmheim / torsk

An echo state network (ESN) for video prediction
10 stars 5 forks source link

Recreate Fig. 5.13 of thesis automatically #48

Closed nmheim closed 5 years ago

nmheim commented 5 years ago

check the effect of varying time prediction length. 100 steps ahead was used in thesis. is shorter or longer better? try 20, 40, 60, ..., 120, ... 240

nmheim commented 5 years ago

Not sure if a 1:1 recreation of the thesis figure makes sense, as we are now using IMED, which result in one error value per frame. Additionally that plot, in its current form, shows only one row of the given frames over time, so I guess its not really suitable for the cases of new anomalies for which we don't know the location.

Anyway, I started predicting they mackey circle as they are easier to work with. The averaged IMED values for all the predictions (including anomalies) look like this. (cycle based prediction can probably still be improved, the cycle_length is set to 200 atm) It is not really a nice method, but based on that plot I would estimate that we can predict 50 frames reasonably well.

Similar plot to Fig. 5.13:

vpl50

  1. plot: prediction IMEDs of all runs
  2. plot: mean prediction IMED
  3. plot: normality score based on 2. plot (large_window=20, small_window=3)
  4. plot: blue: mackey sequence used to create the moving circle, yellow: anomaly

meh, not so nice. decreasing to 25-step-ahead prediction:

vpl25

much better, but the normality score is not reeeeally low.

Anyway I think it is a reasonably good result, considering that I am definitely not able to see an anomaly in the mackey_video.mp4 in this folder (they are at frames 2300, 2600, 2900) with the naked eye

nmheim commented 5 years ago

this is done in torsk detect-row