add caffe scale (#732)

fix #720

mmdnn/conversion/caffe/mapper.py

@@ -259,11 +259,15 @@ def map_scale(cls, node):
         # TODO: The gamma parameter has to be set (in node.data?) and this should work.
         # Also, mean should be set to 0, and var to 1, just to be safe.
         if node.data:
-            raise NotImplementedError
             scale_value = float(node.parameters.filler.value)
-            kwargs = {'scale' : False, 'bias' : False, 'gamma' : scale_value, 'epsilon': 0}
+            if node.parameters.bias_term:
+                bias_value = float(node.parameters.bias_filler.value)
+                kwargs = {'use_scale' : True, 'use_bias' : node.parameters.bias_term, 'gamma' : scale_value, 'beta': bias_value, 'epsilon': 0}
+            else:
+                kwargs = {'use_scale' : True, 'use_bias' : node.parameters.bias_term, 'gamma' : scale_value, 'epsilon': 0}
+
             cls._convert_output_shape(kwargs, node)
-            return Node.create('Scale', **kwargs)
+            return Node.create('Affine', **kwargs)
         else:
             return Node.create('Mul')
 

mmdnn/conversion/examples/caffe/extractor.py

@@ -44,6 +44,8 @@ class caffe_extractor(base_extractor):
                            'caffemodel' : 'http://dl.caffe.berkeleyvision.org/fcn16s-heavy-pascal.caffemodel'},
         'voc-fcn32s'    : {'prototxt' : MMDNN_BASE_URL + "caffe/voc-fcn32s_deploy.prototxt",
                            'caffemodel' : 'http://dl.caffe.berkeleyvision.org/fcn32s-heavy-pascal.caffemodel'},
+        'trailnet_sresnet': {'prototxt': 'https://raw.githubusercontent.com/NVIDIA-AI-IOT/redtail/master/models/pretrained/TrailNet_SResNet-18.prototxt',
+                            'caffemodel': 'https://raw.githubusercontent.com/NVIDIA-AI-IOT/redtail/master/models/pretrained/TrailNet_SResNet-18.caffemodel'}
     }
 
 
@@ -80,7 +82,7 @@ def inference(cls, architecture_name, architecture, path, image_path):
             img = np.transpose(img, (2, 0, 1))
             img = np.expand_dims(img, 0)
             net.blobs['data'].data[...] = img
-            predict = np.squeeze(net.forward()[net._layer_names[-1]][0])
+            predict = np.squeeze(net.forward()[net._output_list[-1]][0])
             predict = np.squeeze(predict)
             return predict
 

mmdnn/conversion/examples/imagenet_test.py

@@ -88,6 +88,7 @@ class TestKit(object):
             'voc-fcn8s'     : lambda path : TestKit.ZeroCenter(path, 500, True),
             'voc-fcn16s'    : lambda path : TestKit.ZeroCenter(path, 500, True),
             'voc-fcn32s'    : lambda path : TestKit.ZeroCenter(path, 500, True),
+            'trailnet_sresnet': lambda path: TestKit.ZeroCenter(path, (320, 180), True)
         },
 
         'tensorflow' : {
@@ -246,7 +247,11 @@ def __init__(self):
     @staticmethod
     def ZeroCenter(path, size, BGRTranspose=False):
         img = Image.open(path)
-        img = img.resize((size, size))
+        if isinstance(size, tuple):
+            h, w = size[0], size[1]
+        else:
+            h, w = size, size
+        img = img.resize((h, w))
         x = np.array(img, dtype=np.float32)
 
         # Reference: 1) Keras image preprocess: https://github.com/keras-team/keras/blob/master/keras/applications/imagenet_utils.py

mmdnn/conversion/keras/keras2_emitter.py

@@ -862,6 +862,23 @@ def emit_PRelu(self, IR_node, in_scope=False):
             )
             return code
 
+    def emit_Affine(self, IR_node, in_scope=False):
+        if in_scope:
+            raise NotImplementedError
+        else:
+            self.used_layers.add('Affine')
+            if IR_node.layer.attr.get('beta', None) is None:
+                bias = None
+            else:
+                bias = IR_node.layer.attr['beta'].f
+            code = "{:<15} = Affine(name='{}', scale={}, bias={})({})".format(
+                IR_node.variable_name,
+                IR_node.name,
+                IR_node.layer.attr['gamma'].f,
+                bias,
+                self.parent_variable_name(IR_node))
+            return code
+
     def emit_yolo(self, IR_node, in_scope=False):
         self.used_layers.add('Yolo')
         self.yolo_parameter = [IR_node.get_attr('anchors'),
@@ -1183,6 +1200,28 @@ def compute_output_shape(self, input_shape):
         return input_shape""")
 
 
+    def _layer_Affine(self):
+        self.add_body(0, '''
+from keras.engine import Layer, InputSpec
+from keras import initializers
+from keras  import backend as K
+
+class Affine(Layer):
+    def __init__(self, scale, bias=None, **kwargs):
+        super(Affine, self).__init__(**kwargs)
+        self.gamma = scale
+        self.beta = bias
+
+    def call(self, inputs, training=None):
+        input_shape = K.int_shape(inputs)
+        # Prepare broadcasting shape.
+        return self.gamma * inputs + self.beta
+
+    def compute_output_shape(self, input_shape):
+        return input_shape
+        ''')
+
+
     def _layer_Split(self):
         self.add_body(0, '''
 def __split(input, split_num, axis):