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

This will fix bob.paper.fargo_tbiom_2019!1 (merged)

bob/learn/tensorflow/network/Vgg.py

@@ -9,7 +9,10 @@ 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 .utils import is_trainable
 
 
 def vgg_19(inputs,
@@ -40,7 +43,9 @@ def vgg_19(inputs,
 
 def vgg_16(inputs,
            reuse=None,                      
-           mode=tf.estimator.ModeKeys.TRAIN, **kwargs):
+           mode=tf.estimator.ModeKeys.TRAIN,
+           trainable_variables=None,
+           **kwargs):
     """
     Oxford Net VGG 16-Layers version E Example from tf-slim
 
@@ -57,9 +62,65 @@ def vgg_16(inputs,
            Estimator mode keys
     """
 
+    dropout_keep_prob = 0.5
+    end_points = {}
     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        
+        
+        # Use conv2d instead of fully_connected layers.
+        name = "fc6"        
+        trainable = is_trainable(name, trainable_variables, mode=mode)
+        net = layers.conv2d(net, 4096, [7, 7], padding='VALID', 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.conv2d(net, 4096, [1, 1], scope='fc7', trainable=trainable)
+        net = layers_lib.dropout(
+            net, dropout_keep_prob, is_training=mode==tf.estimator.ModeKeys.TRAIN, scope='dropout7')
+            
+        net = layers.flatten(net)
+        
+        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)