Merge branch 'coreml-2' into 'main'

Coreml 2

See merge request vato007/fast-depth-tf!9

coreml.py

@@ -1,8 +1,8 @@
 import coremltools as ct
 
 
-def convert_coreml(model_path, save_path='../mobilenet_nnconv5.mlmodel'):
+def convert_coreml(model_path, save_path='mobilenet_nnconv5.mlmodel'):
-    mlmodel = ct.convert(model_path)
+    mlmodel = ct.convert(model_path, inputs=[ct.ImageType()])
     mlmodel.save(save_path)
 
 

dense_depth_functional.py

@@ -1,5 +1,4 @@
 import tensorflow.keras as keras
-import tensorflow_datasets as tfds
 
 import fast_depth_functional as fd
 
@@ -12,12 +11,23 @@ def dense_upsample_block(input, out_channels, skip_connection):
     x = keras.layers.Concatenate()([x, skip_connection])
     x = keras.layers.Conv2D(filters=out_channels,
                             kernel_size=3, strides=1, padding='same')(x)
+    x = keras.layers.LeakyReLU(alpha=0.2)(x)
     x = keras.layers.Conv2D(filters=out_channels,
                             kernel_size=3, strides=1, padding='same')(x)
     return keras.layers.LeakyReLU(alpha=0.2)(x)
 
 
 def dense_depth(size, weights=None, shape=(224, 224, 3)):
+    """
+    Make the dense depth network graph using keras functional api.
+
+    Note that you should use the dense depth loss function, and use Adam as the optimiser with a learning rate of 0.0001
+    (default learning rate of Adam is 0.001).
+    :param size:
+    :param weights:
+    :param shape:
+    :return:
+    """
     input = keras.layers.Input(shape=shape)
     densenet = dense_net(input, size, weights, shape)
 
@@ -37,6 +47,8 @@ def dense_depth(size, weights=None, shape=(224, 224, 3)):
     decoder = dense_upsample_block(
         decoder, densenet_output_channels // 16, densenet.get_layer('conv1/relu').output)
 
+    decoder = dense_upsample_block(decoder, int(densenet_output_channels / 32), input)
+
     conv3 = keras.layers.Conv2D(
         filters=1, kernel_size=3, strides=1, padding='same', name='conv3')(decoder)
     return keras.Model(inputs=input, outputs=conv3, name='dense_depth')

fast_depth_functional.py

@@ -72,6 +72,7 @@ def mobilenet_nnconv5(weights=None, shape=(224, 224, 3)):
     x = keras.layers.Conv2D(1, 1, padding='same')(x)
     x = keras.layers.BatchNormalization()(x)
     x = keras.layers.ReLU(6.)(x)
+    x = keras.layers.Reshape([shape[0], shape[1]])(x)
     return keras.Model(inputs=input, outputs=x, name="fast_depth")
 
 

losses.py

@@ -5,6 +5,10 @@ def dense_depth_loss_function(y, y_pred):
     """
     Implementation of the loss from the dense depth paper https://arxiv.org/pdf/1812.11941.pdf
     """
+    if len(y.shape) == 3:
+        y = tf.expand_dims(y, 3)
+    if len(y_pred.shape) == 3:
+        y_pred = tf.expand_dims(y_pred, 3)
     # Point-wise L1 loss
     l1_depth = tf.reduce_mean(tf.math.abs(y_pred - y), axis=-1)
 
@@ -15,6 +19,6 @@ def dense_depth_loss_function(y, y_pred):
                               tf.math.abs(dx_pred - dx), axis=-1)
 
     #  Structural Similarity (SSIM)
-    ssim = (1 - tf.image.ssim(y, y_pred, 500)) / 2
+    ssim = tf.clip_by_value((1 - tf.image.ssim(y, y_pred, 100)) / 2, 0, 1)
 
     return 0.1 * tf.reduce_mean(l1_depth) + tf.reduce_mean(gradient) + ssim

metric.py

@@ -1,7 +1,7 @@
 import tensorflow as tf
 
 
-def delta1_metric(y_true, y_pred):
+def delta1(y_true, y_pred):
     max_ratio = tf.maximum(y_pred / y_true, y_true / y_pred)
     return tf.reduce_mean(tf.cast(max_ratio < tf.convert_to_tensor(1.25), tf.float32))
 

train.py

@@ -19,7 +19,7 @@ def compile(model, optimiser=keras.optimizers.SGD(), loss=keras.losses.MeanSquar
                   loss=loss,
                   metrics=[keras.metrics.RootMeanSquaredError(),
                            keras.metrics.MeanSquaredError(),
-                           delta1_metric,
+                           delta1,
                            delta2,
                            delta3,
                            keras.metrics.MeanAbsolutePercentageError(),