Merge branch 'vgg16-fix' into 'master'

Fix issue with VGG16 from slim. The slim models adds the hot-encoded in the architecture function

See merge request !78

bob/learn/tensorflow/network/Vgg.py

@@ -9,7 +9,11 @@ VGG16 and VGG19 wrappers
 
 import tensorflow as tf
 from tensorflow.contrib.slim.python.slim.nets import vgg
+from tensorflow.contrib.layers.python.layers import layers as layers_lib
+from tensorflow.contrib import layers
 import tensorflow.contrib.slim as slim
+from tensorflow.python.ops import variable_scope
+from .utils import is_trainable
 
 
 def vgg_19(inputs,
@@ -40,7 +44,10 @@ def vgg_19(inputs,
 
 def vgg_16(inputs,
            reuse=None,                      
-           mode=tf.estimator.ModeKeys.TRAIN, **kwargs):
+           mode=tf.estimator.ModeKeys.TRAIN,
+           trainable_variables=None,
+           scope="vgg_16",
+           **kwargs):
     """
     Oxford Net VGG 16-Layers version E Example from tf-slim
 
@@ -57,9 +64,71 @@ def vgg_16(inputs,
            Estimator mode keys
     """
 
+    dropout_keep_prob = 0.5
+    end_points = {}
+    
+
+    with variable_scope.variable_scope(scope, 'vgg_16', [inputs]) as sc:
+        end_points_collection = sc.original_name_scope + '_end_points'
+        # Collect outputs for conv2d, fully_connected and max_pool2d.
+
+        with slim.arg_scope(
+            [layers.conv2d, layers_lib.fully_connected, layers_lib.max_pool2d], outputs_collections=end_points_collection):
+    
             with slim.arg_scope(
                 [slim.conv2d],
                     trainable=mode==tf.estimator.ModeKeys.TRAIN):
 
-        return vgg.vgg_16(inputs, spatial_squeeze=False)
+                name = "conv1"
+                trainable = is_trainable(name, trainable_variables, mode=mode)
+                net = layers_lib.repeat(
+                    inputs, 2, layers.conv2d, 64, [3, 3], scope=name, trainable=trainable)
+                net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
+                end_points[name] = net
+
+                name = "conv2"        
+                trainable = is_trainable(name, trainable_variables, mode=mode)
+                net = layers_lib.repeat(net, 2, layers.conv2d, 128, [3, 3], scope=name, trainable=trainable)
+                net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
+                end_points[name] = net        
+                
+                name = "conv3"
+                trainable = is_trainable(name, trainable_variables, mode=mode)
+                net = layers_lib.repeat(net, 3, layers.conv2d, 256, [3, 3], scope=name, trainable=trainable)
+                net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
+                end_points[name] = net        
+
+                name = "conv4"
+                trainable = is_trainable(name, trainable_variables, mode=mode)        
+                net = layers_lib.repeat(net, 3, layers.conv2d, 512, [3, 3], scope=name, trainable=trainable)
+                net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
+                end_points[name] = net        
+
+                name = "conv5"
+                trainable = is_trainable(name, trainable_variables, mode=mode)
+                net = layers_lib.repeat(net, 3, layers.conv2d, 512, [3, 3], scope=name, trainable=trainable)
+                net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
+                end_points[name] = net        
+                
+                net = layers.flatten(net)
+                
+                # Use conv2d instead of fully_connected layers.
+                name = "fc6"
+                trainable = is_trainable(name, trainable_variables, mode=mode)
+                net = layers.fully_connected(net, 4096, scope=name, trainable=trainable)
+                net = layers_lib.dropout(
+                    net, dropout_keep_prob, is_training=mode==tf.estimator.ModeKeys.TRAIN, scope='dropout6')
+                end_points[name] = net            
+
+                name = "fc7"            
+                trainable = is_trainable(name, trainable_variables, mode=mode)
+                net = layers.fully_connected(net, 4096, scope=name, trainable=trainable)
+                net = layers_lib.dropout(
+                    net, dropout_keep_prob, is_training=mode==tf.estimator.ModeKeys.TRAIN, scope='dropout7')
+
+                end_points[name] = net            
+  
+    # Convert end_points_collection into a end_point dict.
+    return net, end_points
+    #return vgg.vgg_16(inputs, spatial_squeeze=False)
 

bob/learn/tensorflow/test/test_architectures.py

@@ -100,6 +100,7 @@ def test_inceptionv1_adaptation():
 
 def test_vgg():
     tf.reset_default_graph()
+
     # Testing VGG19 Training mode
     inputs = tf.placeholder(tf.float32, shape=(1, 224, 224, 3))
     graph, _ = vgg_19(inputs)
@@ -116,10 +117,11 @@ def test_vgg():
     tf.reset_default_graph()
     assert len(tf.global_variables()) == 0
 
+
     # Testing VGG 16 training mode
     inputs = tf.placeholder(tf.float32, shape=(1, 224, 224, 3))
     graph, _ = vgg_16(inputs)
-    assert len(tf.trainable_variables()) == 32
+    assert len(tf.trainable_variables()) == 30
 
     tf.reset_default_graph()
     assert len(tf.global_variables()) == 0