Merge pull request #120 from DeepRank/NeuralNet_class_weights

Added class weights to the CrossEntropy Loss

.travis.yml

@@ -1,5 +1,5 @@
 sudo: false
-distr : "trusty"
+distr : "bionic"
 
 language: generic
 

deeprank/learn/NeuralNet.py

@@ -29,6 +29,7 @@ class NeuralNet():
 
     def __init__(self, data_set, model,
                  model_type='3d', proj2d=0, task='reg',
+                 class_weights = None,
                  pretrained_model=None,
                  cuda=False, ngpu=0,
                  plot=True,
@@ -59,6 +60,10 @@ def __init__(self, data_set, model,
                 The loss function, the target datatype and plot functions
                 will be autmatically adjusted depending on the task.
 
+            class_weights (Tensor): a manual rescaling weight given to
+                each class. If given, has to be a Tensor of size #classes.
+                Only applicable on 'class' task.
+
             pretrained_model (str): Saved model to be used for further
                 training or testing.
 
@@ -108,6 +113,8 @@ def __init__(self, data_set, model,
         # pretrained model
         self.pretrained_model = pretrained_model
 
+        self.class_weights = class_weights
+
         if isinstance(self.data_set, (str, list)) and pretrained_model is None:
             raise ValueError(
                 'Argument data_set must be a DeepRankDataSet object\
@@ -169,7 +176,7 @@ def __init__(self, data_set, model,
             self._plot_scatter = self._plot_scatter_reg
 
         elif self.task == 'class':
-            self.criterion = nn.CrossEntropyLoss(reduction='sum')
+            self.criterion = nn.CrossEntropyLoss(weight = self.class_weights, reduction='mean')
             self._plot_scatter = self._plot_boxplot_class
             self.data_set.normalize_targets = False
 

setup.py

@@ -17,7 +17,7 @@
         'pandas',
         'matplotlib',
         'torchsummary',
-        'torch',
+        'torch < 1.4.0',
         'pdb2sql >= 0.2.1',
         'freesasa==2.0.3.post7;platform_system=="Linux"',
         'freesasa==2.0.3.post6;platform_system=="Darwin"'

test/test_learn.py

@@ -65,7 +65,7 @@ def test_learn_3d_reg_mapfly():
         model.train(
             nepoch=5,
             divide_trainset=0.8,
-            train_batch_size=5,
+            train_batch_size=2,
             preshuffle_seed=2019,
             num_workers=0)
 
@@ -116,7 +116,7 @@ def test_learn_3d_reg():
         model.train(
             nepoch=5,
             divide_trainset=0.8,
-            train_batch_size=5,
+            train_batch_size=2,
             num_workers=0,
             preshuffle_seed=2019,
             save_model='all')
@@ -161,7 +161,7 @@ def test_learn_3d_class():
         model.train(
             nepoch=5,
             divide_trainset=0.8,
-            train_batch_size=5,
+            train_batch_size=2,
             num_workers=0,
             save_epoch='all')
 
@@ -210,7 +210,7 @@ def test_learn_2d_reg():
         model.train(
             nepoch=5,
             divide_trainset=0.8,
-            train_batch_size=5,
+            train_batch_size=2,
             num_workers=0)
 
     @unittest.skipIf(skip, "torch fails on Travis")