From e67e13f254bf44a126f912d76c473f6c9ab128d9 Mon Sep 17 00:00:00 2001
From: Nikolai Hartmann <Nikolai.Hartmann@physik.uni-muenchen.de>
Date: Mon, 19 Nov 2018 16:14:30 +0100
Subject: [PATCH] use mse by default for regression
---
toolkit.py | 33 +++++++++++++++++++++++----------
1 file changed, 23 insertions(+), 10 deletions(-)
diff --git a/toolkit.py b/toolkit.py
index 620d3f6..3bda85e 100755
--- a/toolkit.py
+++ b/toolkit.py
@@ -182,7 +182,7 @@ class ClassificationProject(object):
the first time. This seed (increased by one) is used again before
training when keras shuffling is used.
- :param loss: loss function name
+ :param loss: loss function name (or list of names in case of regression targets)
:param mask_value: value that is used for non-existent entries (e.g. 4th jet pt in events with 3 jets)
@@ -346,6 +346,9 @@ class ClassificationProject(object):
self.shuffle_seed = shuffle_seed
self.balance_dataset = balance_dataset
self.loss = loss
+ if self.regression_branches and (not isinstance(self.loss, list)):
+ self.loss = [self.loss]+["mean_squared_error"]*len(self.regression_branches)
+
self.mask_value = mask_value
self.apply_class_weight = apply_class_weight
self.normalize_weights = normalize_weights
@@ -865,7 +868,9 @@ class ClassificationProject(object):
idx = self.train_val_idx[1]
x_val, y_val, w_val = self.x_train[idx], self.y_train[idx], self.w_train_tot[idx]
x_val_input = self.get_input_list(self.transform(x_val))
- return x_val_input, y_val, w_val
+ y_val_output = self.get_output_list(y_val)
+ w_val_list = self.get_weight_list(w_val)
+ return x_val_input, y_val_output, w_val_list
@property
@@ -874,7 +879,9 @@ class ClassificationProject(object):
idx = self.train_val_idx[0]
x_train, y_train, w_train = self.x_train[idx], self.y_train[idx], self.w_train_tot[idx]
x_train_input = self.get_input_list(self.transform(x_train))
- return x_train_input, y_train, w_train
+ y_train_output = self.get_output_list(y_train)
+ w_train_list = self.get_weight_list(w_train)
+ return x_train_input, y_train_output, w_train_list
@property
@@ -917,6 +924,14 @@ class ClassificationProject(object):
return np.hsplit(y, len(self.target_fields)+1)
+ def get_weight_list(self, w):
+ "Repeat weight n times for regression targets"
+ if not self.target_fields:
+ return w
+ else:
+ return [w]*(len(self.target_fields)+1)
+
+
def yield_batch(self):
"Batch generator - optionally shuffle the indices after each epoch"
x_train, y_train, w_train = self.x_train, self.y_train, self.w_train_tot
@@ -936,7 +951,8 @@ class ClassificationProject(object):
w_batch = w_train[shuffled_idx[start:start+int(self.batch_size)]]
x_input = self.get_input_list(self.transform(x_batch))
y_output = self.get_output_list(y_batch)
- yield (x_input, y_output, w_batch)
+ w_list = self.get_weight_list(w_batch)
+ yield (x_input, y_output, w_list)
def yield_single_class_batch(self, class_label):
@@ -955,12 +971,9 @@ class ClassificationProject(object):
shuffled_idx = class_idx
# yield them batch wise
for start in range(0, len(shuffled_idx), int(self.batch_size/2)):
- x_batch = x_train[shuffled_idx[start:start+int(self.batch_size/2)]]
- y_batch = y_train[shuffled_idx[start:start+int(self.batch_size/2)]]
- w_batch = w_train[shuffled_idx[start:start+int(self.batch_size/2)]]
- x_input = self.get_input_list(self.transform(x_batch))
- y_output = self.get_output_list(y_batch)
- yield (x_input, y_output, w_batch)
+ yield (x_train[shuffled_idx[start:start+int(self.batch_size/2)]],
+ y_train[shuffled_idx[start:start+int(self.batch_size/2)]],
+ w_train[shuffled_idx[start:start+int(self.batch_size/2)]])
def yield_balanced_batch(self):
--
GitLab