Merge pull request #3 from exosports/lossfix

Fixed using custom loss functions

Author: mdhimes

MARGE.py

@@ -333,26 +333,30 @@ def MARGE(confile):
             model_evaluate = None
             if "lossfunc" in conf.keys():
                 # Format: path/to/module.py function_name
-                lossfunc = conf["lossfunc"].split()  # [path/to/module.py, function_name]
-                if lossfunc[0] == 'mse':
-                    lossfunc = keras.losses.MeanSquaredError
+                #      or string identifier
+                lossfunc = conf["lossfunc"].split()  # [path/to/module.py, function_name] or [string identifier]
+                # Handle string identifiers:
+                if lossfunc[0] in ['mse', 'mean_squared_error']:
+                    lossfunc = keras.losses.MeanSquaredError()
                     lossfunc.__name__ = 'mse'
-                elif lossfunc[0] == 'mae':
-                    lossfunc = keras.losses.MeanAbsoluteError
+                elif lossfunc[0] in ['mae', 'mean_absolute_error']:
+                    lossfunc = keras.losses.MeanAbsoluteError()
                     lossfunc.__name__ = 'mae'
-                elif lossfunc[0] == 'mape':
-                    lossfunc = keras.losses.MeanAbsolutePercentageError
+                elif lossfunc[0] in ['mape', 'mean_absolute_percent_error', 'mean_absolute_percentage_error']:
+                    lossfunc = keras.losses.MeanAbsolutePercentageError()
                     lossfunc.__name__ = 'mape'
-                elif lossfunc[0] == 'msle':
-                    lossfunc = keras.losses.MeanSquaredLogarithmicError
+                elif lossfunc[0] in ['msle', 'mean_squared_logarithmic_error', 'mean_squared_log_error']:
+                    lossfunc = keras.losses.MeanSquaredLogarithmicError()
                     lossfunc.__name__ = 'msle'
-                elif lossfunc[0] in ['maxmse', 'm3se', 'mslse', 'mse_per_ax', 'maxse']:
+                    # Handle the custom loss functions in lib/losses.py
+                elif lossfunc[0] in ['maxmse', 'm3se', 'mslse', 'mse_per_ax', 'maxse', 'smape', 'maxsape']:
                     lossname = lossfunc[0]
                     lossfunc = getattr(losses, lossfunc[0])
                     lossfunc.__name__ = lossname
                 elif lossfunc[0] in ['heteroscedastic', 'heteroscedastic_loss']:
                     lossfunc = functools.partial(losses.heteroscedastic_loss, D=np.product(oshape), N=batch_size)
                     lossfunc.__name__ = 'heteroscedastic_loss'
+                # Handle user-supplied custom loss functions
                 else:
                     if lossfunc[0][-3:] == '.py':
                         lossfunc[0] = lossfunc[0][:-3]   # path/to/module

lib/NN.py

@@ -185,7 +185,7 @@ def __init__(self, ftrain_TFR, fvalid_TFR, ftest_TFR,
         self.stop_file   = stop_file
 
         if lossfunc is None:
-            lossfunc = keras.losses.MeanSquaredError
+            lossfunc = keras.losses.MeanSquaredError()
             lossfunc.__name__ = 'mse'
         #else:
         #    self.lossfunc = lossfunc
@@ -382,7 +382,7 @@ def __init__(self, ftrain_TFR, fvalid_TFR, ftest_TFR,
 
         # Compile model
         self.model.compile(optimizer=Adam(learning_rate=self.lengthscale, amsgrad=True),
-                           loss=lossfunc())
+                           loss=lossfunc)
         if self.verb:
             self.model.summary(print_fn=logger.info)
 

lib/loss/losses.py

@@ -88,3 +88,14 @@ def heteroscedastic_loss(true, pred, D, N):
 
     return K.mean(quad + log_det, 0)
 
+def smape(y_true, y_pred):
+    epsilon = 1e-8  # Small constant to prevent division by zero
+    numerator = tf.abs(y_pred - y_true)
+    denominator = tf.abs(y_true) + tf.abs(y_pred) + epsilon
+    return K.mean(numerator / denominator) * 200
+
+def maxsape(y_true, y_pred):
+    epsilon = 1e-8  # Small constant to prevent division by zero
+    numerator = tf.abs(y_pred - y_true)
+    denominator = tf.abs(y_true) + tf.abs(y_pred) + epsilon
+    return K.max(numerator / denominator) * 200