Merge pull request #25 from IvanKuchin/development

Model output shape replaced from array to a single tensor

config.py

@@ -14,7 +14,7 @@
 BATCH_SIZE = 1
 BATCH_NORM_MOMENTUM = 0.8
 
-GRADIENT_ACCUMULATION_STEPS = None # https://www.tensorflow.org/api_docs/python/tf/keras/optimizers/Adam#args
+GRADIENT_ACCUMULATION_STEPS = 4 # https://www.tensorflow.org/api_docs/python/tf/keras/optimizers/Adam#args
 
 # Option 1) HU range for pancreas in CT scans from 30 to 400
 # https://radiopaedia.org/articles/windowing-ct?lang=us

tools/craft_network/att_gate.py

@@ -62,7 +62,7 @@ def call(self, inputs):
 
 
         __sum = self.add_g_x([theta_x, phi_g])
-        __activation_sum = tf.keras.layers.Activation("relu")(__sum)
+        __activation_sum = tf.keras.layers.LeakyReLU()(__sum)
 
         psi = tf.keras.layers.Activation("sigmoid")(self.psi(__activation_sum))
         psi = self.psi_upsample(psi)

tools/craft_network/att_unet.py

@@ -109,7 +109,7 @@ def craft_network(checkpoint_file = None, apply_batchnorm=True, apply_instanceno
 
     output_layer = tf.keras.layers.Conv3D(2, kernel_size = 1, padding = "same", kernel_initializer = "he_uniform", activation="softmax")(x)
 
-    model = tf.keras.models.Model(inputs = [inputs], outputs = [output_layer])
+    model = tf.keras.models.Model(inputs = [inputs], outputs = output_layer)
 
     if checkpoint_file and os.path.exists(checkpoint_file):
         print("Loading weights from checkpoint ", checkpoint_file)

tools/craft_network/att_unet_dsv.py

@@ -54,7 +54,7 @@ def craft_network(checkpoint_file = None, apply_batchnorm = True, apply_instance
     output_layer = tf.keras.layers.Conv3D(2, kernel_size = 1, padding = "same", kernel_initializer = "he_uniform", activation="softmax")(concat_layer)
     # output_layer = tf.keras.activations.softmax(output_layer)
 
-    model = tf.keras.models.Model(inputs = [inputs], outputs = [output_layer])
+    model = tf.keras.models.Model(inputs = [inputs], outputs = output_layer)
 
     if checkpoint_file and os.path.exists(checkpoint_file):
         print("Loading weights from checkpoint ", checkpoint_file)

tools/craft_network/dsv.py

@@ -22,6 +22,8 @@ def call(self, inputs, **kwargs):
         if self.scale_factor > 1:
             x = self.upsample(x)
 
+        x = tf.keras.activations.softmax(x)
+
         return x
 
     ##########################
@@ -44,7 +46,7 @@ def get_config(self):
     dsv_model = DSV(scale_factor = 4)
     result = dsv_model(x_shape)
 
-    model = tf.keras.models.Model(inputs = [inp], outputs = [result])
+    model = tf.keras.models.Model(inputs = [inp], outputs = result)
     outputs = model(rnd)
 
     model.summary()

tools/craft_network/unet_classic.py

@@ -60,7 +60,7 @@ def craft_network(checkpoint_file = None, apply_batchnorm=True):
 
     output_layer = tf.keras.layers.Conv3D(2, kernel_size = 1, padding = "same", kernel_initializer = "he_uniform")(x)
 
-    model = tf.keras.models.Model(inputs = [inputs], outputs = [output_layer])
+    model = tf.keras.models.Model(inputs = [inputs], outputs = output_layer)
 
     if checkpoint_file and os.path.exists(checkpoint_file):
         print("Loading weights from checkpoint ", checkpoint_file)

tools/craft_network/unet_shortcuts_every_layer.py

@@ -47,7 +47,7 @@ def craft_network(checkpoint_file = None, apply_batchnorm=True):
 
     output_layer = tf.keras.layers.Conv3D(2, kernel_size = 1, padding = "same", kernel_initializer = "he_uniform")(x)
 
-    model = tf.keras.models.Model(inputs = [inputs], outputs = [output_layer])
+    model = tf.keras.models.Model(inputs = [inputs], outputs = output_layer)
 
     if checkpoint_file and os.path.exists(checkpoint_file):
         print("Loading weights from checkpoint ", checkpoint_file)

train_segmentation.py

@@ -99,9 +99,9 @@ def main():
                                                                 mode = config.MONITOR_MODE)
     early_stopping = tf.keras.callbacks.EarlyStopping(monitor = config.MONITOR_METRIC, mode = config.MONITOR_MODE, patience = 200,
                                                       verbose = 1)
-    model.compile(optimizer = tf.keras.optimizers.Adam(
+    model.compile(optimizer = tf.keras.optimizers.AdamW(
                                                         learning_rate = config.INITIAL_LEARNING_RATE,
-                                                        gradient_accumulation_steps = config.GRADIENT_ACCUMULATION_STEPS,
+                                                        # gradient_accumulation_steps = config.GRADIENT_ACCUMULATION_STEPS,
                                                         ),
                   loss = __dice_loss,
                   metrics = [