@@ -14,13 +14,13 @@ def get_handler_name(node_kind):
1414
1515
1616class NodeMapper (object ):
17-
17+
1818 @classmethod
1919 def _convert_output_shape (cls , kwargs , node ):
2020 shape = TensorShape ()
2121 dim = shape .dim .add ()
2222 dim .size = - 1
23-
23+
2424 if len (node .output_shape ) > 2 :
2525 for i in node .output_shape [2 :]:
2626 dim = shape .dim .add ()
@@ -31,22 +31,22 @@ def _convert_output_shape(cls, kwargs, node):
3131 dim = shape .dim .add ()
3232 dim .size = node .output_shape [1 ]
3333 kwargs ['_output_shapes' ] = [shape ]
34-
34+
3535 @classmethod
3636 def get_kernel_params (cls , node ):
37- kwargs = {}
37+ kwargs = {}
3838 if node .kernel_parameters .p_h > 0 or node .kernel_parameters .p_w > 0 :
39- padding = [0 , 0 , node .kernel_parameters .p_h , node .kernel_parameters .p_h , node .kernel_parameters .p_w , node .kernel_parameters .p_w , 0 , 0 ]
39+ padding = [0 , node .kernel_parameters .p_h , node .kernel_parameters .p_w , 0 , 0 , node .kernel_parameters .p_h , node .kernel_parameters .p_w , 0 ]
4040 elif node .kernel_parameters .s_h > 1 or node .kernel_parameters .s_w > 1 :
41- padding = [0 , 0 , (node .kernel_parameters .s_h - 1 ) // 2 , node .kernel_parameters .s_h // 2 , ( node .kernel_parameters .s_w - 1 ) // 2 , node .kernel_parameters .s_w // 2 , 0 , 0 ]
41+ padding = [0 , (node .kernel_parameters .s_h - 1 ) // 2 , ( node .kernel_parameters .s_w - 1 ) // 2 , 0 , 0 , node .kernel_parameters .s_h // 2 , node .kernel_parameters .s_w // 2 , 0 ]
4242 else :
4343 padding = None
44-
45- kwargs ['padding ' ] = 'VALID'
44+
45+ kwargs ['auto_pad ' ] = 'VALID'
4646 kwargs ['strides' ] = [1 , node .kernel_parameters .s_h , node .kernel_parameters .s_w , 1 ]
4747 cls ._convert_output_shape (kwargs , node )
48-
49- return kwargs , {'paddings ' : padding , 'mode' : 'CONSTANT ' , 'constant_values' : 0.0 }
48+
49+ return kwargs , {'pads ' : padding , 'mode' : 'constant ' , 'constant_values' : 0.0 }
5050
5151
5252 @classmethod
@@ -60,7 +60,7 @@ def map_data(cls, node):
6060 dim .size = i
6161 dim = shape .dim .add ()
6262 dim .size = node .output_shape .channels
63-
63+
6464 kwargs = {'shape' : shape } # Ignore the dimension of batch size
6565 cls ._convert_output_shape (kwargs , node )
6666 return Node .create ('DataInput' , ** kwargs )
@@ -74,29 +74,30 @@ def map_input(cls, node):
7474 def map_convolution (cls , node ):
7575 kwargs , padding = cls .get_kernel_params (node )
7676 parent , _ = node .get_only_parent ()
77- kwargs ['filter ' ] = [node .kernel_parameters .k_h , node .kernel_parameters .k_w , parent .output_shape .channels , node .parameters .num_output ]
78- kwargs ['use_bias' ] = node .parameters .bias_term
77+ kwargs ['kernel_shape ' ] = [node .kernel_parameters .k_h , node .kernel_parameters .k_w , parent .output_shape .channels , node .parameters .num_output ]
78+ kwargs ['use_bias' ] = node .parameters .bias_term
7979 group = node .parameters .group
8080 if group != 1 :
8181 kwargs ['group' ] = group
82-
83- if padding ['paddings ' ] != None :
84- return [Node .create ('Pad' , ** padding ), Node .create ('Convolution ' , ** kwargs )]
82+
83+ if padding ['pads ' ] != None :
84+ return [Node .create ('Pad' , ** padding ), Node .create ('Conv ' , ** kwargs )]
8585 else :
86- return Node .create ('Convolution' , ** kwargs )
86+ kwargs ['pads' ] = [0 ] * 8
87+ return Node .create ('Conv' , ** kwargs )
8788
8889
8990 @classmethod
9091 def map_deconvolution (cls , node ):
9192 raise NotImplementedError ()
9293 kwargs = cls .get_kernel_params (node )
9394 parent , _ = node .get_only_parent ()
94- kwargs ['filter ' ] = [node .kernel_parameters .k_h , node .kernel_parameters .k_w , parent .output_shape .channels , node .parameters .num_output ]
95+ kwargs ['kernel_shape ' ] = [node .kernel_parameters .k_h , node .kernel_parameters .k_w , parent .output_shape .channels , node .parameters .num_output ]
9596 group = node .parameters .group
9697 if group != 1 :
97- kwargs ['group' ] = group
98+ kwargs ['group' ] = group
9899 return Node .create ('deconv' , ** kwargs )
99-
100+
100101 @classmethod
101102 def map_crop (cls , node ):
102103 offset = node .parameters .offset
@@ -105,13 +106,13 @@ def map_crop(cls, node):
105106 return Node .create ('crop' , ** kwargs )
106107 else :
107108 return Node .create ('crop' )
108-
109+
109110 @classmethod
110111 def map_relu (cls , node ):
111112 kwargs = {}
112113 cls ._convert_output_shape (kwargs , node )
113114 return Node .create ('Relu' , ** kwargs )
114-
115+
115116 @classmethod
116117 def map_pooling (cls , node ):
117118 kwargs , padding = cls .get_kernel_params (node )
@@ -122,21 +123,22 @@ def map_pooling(cls, node):
122123 else :
123124 # Stochastic pooling, for instance.
124125 raise ConversionError ('Unsupported pooling type.' )
125- kwargs ['window_shape ' ] = [1 , node .kernel_parameters .k_h , node .kernel_parameters .k_w , 1 ]
126+ kwargs ['kernel_shape ' ] = [1 , node .kernel_parameters .k_h , node .kernel_parameters .k_w , 1 ]
126127 cls ._convert_output_shape (kwargs , node )
127-
128- if padding ['paddings ' ] != None :
128+
129+ if padding ['pads ' ] != None :
129130 return [Node .create ('Pad' , ** padding ), Node .create ('Pool' , ** kwargs )]
130131 else :
132+ kwargs ['pads' ] = [0 ] * 8
131133 return Node .create ('Pool' , ** kwargs )
132-
134+
133135
134136 @classmethod
135137 def _add_flatten_layer (cls , node ):
136138 shape = TensorShape ()
137139 dim = shape .dim .add ()
138- dim .size = - 1
139-
140+ dim .size = - 1
141+
140142 dim = shape .dim .add ()
141143 dim .size = 1
142144 for i in node .output_shape [1 :]:
@@ -149,40 +151,40 @@ def map_inner_product(cls, node):
149151 #TODO: Axis
150152 assert node .parameters .axis == 1
151153 #TODO: Unbiased
152- kwargs = {'use_bias' : node .parameters .bias_term , 'units' : node .parameters .num_output }
153-
154+ kwargs = {'use_bias' : node .parameters .bias_term , 'units' : node .parameters .num_output }
155+
154156 # check if need the Flatten layer
155157 parent , _ = node .get_only_parent ()
156158 ret = []
157159 if parent .output_shape .height > 1 or parent .output_shape .width > 1 :
158160 ret .append (cls ._add_flatten_layer (parent ))
159161 ret .append (Node .create ('FullyConnected' , ** kwargs ))
160162 return ret
161-
163+
162164 @classmethod
163165 def map_softmax (cls , node ):
164166 return Node .create ('Softmax' )
165-
167+
166168 @classmethod
167169 def map_lrn (cls , node ):
168170 params = node .parameters
169171 assert params .local_size % 2 == 1
170172 kwargs = {'size' : int ((params .local_size + 1 ) / 2 ), 'alpha' : params .alpha , 'beta' : params .beta , 'k' : params .k }
171173 cls ._convert_output_shape (kwargs , node )
172174 return Node .create ('LRN' , ** kwargs )
173-
175+
174176 @classmethod
175177 def map_concat (cls , node ):
176178 kwargs = {'axis' : (2 , 3 , 1 , 0 )[node .parameters .axis ]}
177179 cls ._convert_output_shape (kwargs , node )
178180 return Node .create ('Concat' , ** kwargs )
179-
181+
180182 @classmethod
181183 def map_dropout (cls , node ):
182184 kwargs = {'keep_prob' : node .parameters .dropout_ratio }
183185 cls ._convert_output_shape (kwargs , node )
184186 return Node .create ('Dropout' , ** kwargs )
185-
187+
186188 @classmethod
187189 def map_batch_norm (cls , node ):
188190 scale_offset = len (node .data ) == 4
@@ -191,12 +193,12 @@ def map_batch_norm(cls, node):
191193 kwargs ['epsilon' ] = epsilon
192194 cls ._convert_output_shape (kwargs , node )
193195 return Node .create ('batch_normalization' , ** kwargs )
194-
196+
195197 @classmethod
196198 def map_eltwise (cls , node ):
197199 operations = {0 : 'mul' , 1 : 'sum' , 2 : 'max' }
198200 op_code = node .parameters .operation
199- try :
201+ try :
200202 return Node .create (operations [op_code ])
201203 except KeyError :
202204 raise ConversionError ('Unknown elementwise operation: {}' .format (op_code ))
0 commit comments