From 6380205cea3b3d29fa6b7df3d9ca988b1050d9c0 Mon Sep 17 00:00:00 2001
From: Nikolai Hartmann <Nikolai.Hartmann@physik.uni-muenchen.de>
Date: Wed, 5 Dec 2018 09:38:51 +0100
Subject: [PATCH] make adv learning rate tunable
---
toolkit.py | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/toolkit.py b/toolkit.py
index 6e2b819..b0500fa 100755
--- a/toolkit.py
+++ b/toolkit.py
@@ -2027,10 +2027,12 @@ class ClassificationProjectDecorr(ClassificationProject):
def _init_from_args(self, name,
decorr_bins=10,
+ adv_lr=0.001,
**kwargs):
super(ClassificationProjectDecorr, self)._init_from_args(name, **kwargs)
self.decorr_binnings = []
self.decorr_bins = decorr_bins
+ self.adv_lr = adv_lr
self._write_info("project_type", "ClassificationProjectDecorr")
self._class_layers = None
self._adv_hidden_layers = None
@@ -2247,7 +2249,7 @@ class ClassificationProjectDecorr(ClassificationProject):
self.set_trainability(self.class_layers, False)
self.set_trainability(self.adv_layers, True)
self._model_adv.compile(
- optimizer=keras.optimizers.adam(lr=0.001),
+ optimizer=keras.optimizers.adam(lr=self.adv_lr),
loss="categorical_crossentropy",
)
return self._model_adv
--
GitLab