model robustness testing

notes/models.md

@@ -0,0 +1,37 @@
+# models
+
+## no polarisation flipping
+
+```py
+config_path="data/20241229-163*-128-16384-50000-*.ini"
+model=".models/best_20241230_011907.tar"
+```
+
+```py
+config_path="data/20241229-163*-128-16384-80000-*.ini"
+model=".models/best_20241230_103752.tar"
+```
+
+```py
+config_path="data/20241229-163*-128-16384-100000-*.ini"
+model=".models/best_20241230_164534.tar"
+```
+
+## with polarisation flipping
+
+polarisation flipping: signal is randomly rotated by 180°. polarization rotation can be detected by adding a tone on one of the polarisations, but only to mod 180° with a direct detection setup. the randomly flipped signal should allow the network to hopefully learn to compensate for dispersion, pmd independently from the polarization rot. the training data includes the flipped signal as well, but no indication if the polarisation is flipped.
+
+```py
+config_path="data/20241229-163*-128-16384-50000-*.ini"
+model=".models/best_20241231_000328.tar"
+```
+
+```py
+config_path="data/20241229-163*-128-16384-80000-*.ini"
+model=".models/best_20241231_163614.tar"
+```
+
+```py
+config_path="data/20241229-163*-128-16384-100000-*.ini"
+model=".models/best_20241231_170532.tar"
+```