several updates for radar nowcast challenge

Author: markveilletteLL

radar_nowcasting/RadarNowcastBenchmarks.ipynb

@@ -17,8 +17,8 @@
 
 def create_model(input_shape=(384,384,13),
                  start_neurons=32, 
-	             num_outputs=12, 
-	             activation='relu'):
+                 num_outputs=12, 
+                 activation='relu'):
     """
 
     Inputs are assumed to be in range [0-255].
@@ -27,18 +27,12 @@ def create_model(input_shape=(384,384,13),
 
     """
     inputs = layers.Input(shape=input_shape)
-    
-    # normalize inputs
-    ir069 = tf.keras.layers.Input(shape=(192,192,1)) 
-    ir107 = tf.keras.layers.Input(shape=(192,192,1)) 
-    lght  = tf.keras.layers.Input(shape=(48,48,1))
-    inputs = [ir069,ir107,lght]
 
     # Normalize inputs
-    ir069_norm = tf.keras.layers.Lambda(lambda x,mu,scale: (x-mu)/scale,
-                                        arguments={'mu':MEAN,'scale':SCALE})(inputs)
+    inputs_norm = tf.keras.layers.Lambda(lambda x,mu,scale: (x-mu)/scale,
+                                         arguments={'mu':MEAN,'scale':SCALE})(inputs)
 
-    encoder0_pool, encoder0 = encoder_block(inputs, start_neurons, activation=activation)
+    encoder0_pool, encoder0 = encoder_block(inputs_norm, start_neurons, activation=activation)
     encoder1_pool, encoder1 = encoder_block(encoder0_pool, start_neurons*2, activation=activation)
     encoder2_pool, encoder2 = encoder_block(encoder1_pool, start_neurons*4, activation=activation)
     encoder3_pool, encoder3 = encoder_block(encoder2_pool, start_neurons*8, activation=activation)
@@ -50,21 +44,21 @@ def create_model(input_shape=(384,384,13),
     outputs = layers.Conv2D(num_outputs, (1, 1), padding='same',
                                 activation='linear', name='output_layer')(decoder0)
 
-    return inputs, outputs
+    return tf.keras.models.Model(inputs=inputs,outputs=outputs,name='nowcast_model')
 
 
 def nowcast_mse(y_true,y_pred):
-	""" 
-	MSE loss normalized by SCALE*SCALE
-	"""
-	return mean_squared_error(y_true,y_pred)/(SCALE*SCALE)
+    """ 
+    MSE loss normalized by SCALE*SCALE
+    """
+    return mean_squared_error(y_true,y_pred)/(SCALE*SCALE)
 
 
 def nowcast_mae(y_true,y_pred):
-	"""
-	MAE normalized by SCALE
-	"""
-    retrun mean_absolute_error(y_true,y_pred)/(SCALE)
+    """
+    MAE normalized by SCALE
+    """
+    return mean_absolute_error(y_true,y_pred)/(SCALE)