Faster-RCNN from TensorFlow with dilated convolutions

samples/dnn/tf_text_graph_faster_rcnn.py

@@ -48,10 +48,42 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
 
     removeIdentity(graph_def)
 
+    nodesToKeep = []
     def to_remove(name, op):
+        if name in nodesToKeep:
+            return False
         return op == 'Const' or name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep) or \
                (name.startswith('CropAndResize') and op != 'CropAndResize')
 
+    # Fuse atrous convolutions (with dilations).
+    nodesMap = {node.name: node for node in graph_def.node}
+    for node in reversed(graph_def.node):
+        if node.op == 'BatchToSpaceND':
+            del node.input[2]
+            conv = nodesMap[node.input[0]]
+            spaceToBatchND = nodesMap[conv.input[0]]
+
+            # Extract paddings
+            stridedSlice = nodesMap[spaceToBatchND.input[2]]
+            assert(stridedSlice.op == 'StridedSlice')
+            pack = nodesMap[stridedSlice.input[0]]
+            assert(pack.op == 'Pack')
+
+            padNodeH = nodesMap[nodesMap[pack.input[0]].input[0]]
+            padNodeW = nodesMap[nodesMap[pack.input[1]].input[0]]
+            padH = int(padNodeH.attr['value']['tensor'][0]['int_val'][0])
+            padW = int(padNodeW.attr['value']['tensor'][0]['int_val'][0])
+
+            paddingsNode = NodeDef()
+            paddingsNode.name = conv.name + '/paddings'
+            paddingsNode.op = 'Const'
+            paddingsNode.addAttr('value', [padH, padH, padW, padW])
+            graph_def.node.insert(graph_def.node.index(spaceToBatchND), paddingsNode)
+            nodesToKeep.append(paddingsNode.name)
+
+            spaceToBatchND.input[2] = paddingsNode.name
+
+
     removeUnusedNodesAndAttrs(to_remove, graph_def)
 
 
@@ -225,6 +257,26 @@ def to_remove(name, op):
     detectionOut.addAttr('variance_encoded_in_target', True)
     graph_def.node.extend([detectionOut])
 
+    def getUnconnectedNodes():
+        unconnected = [node.name for node in graph_def.node]
+        for node in graph_def.node:
+            for inp in node.input:
+                if inp in unconnected:
+                    unconnected.remove(inp)
+        return unconnected
+
+    while True:
+        unconnectedNodes = getUnconnectedNodes()
+        unconnectedNodes.remove(detectionOut.name)
+        if not unconnectedNodes:
+            break
+
+        for name in unconnectedNodes:
+            for i in range(len(graph_def.node)):
+                if graph_def.node[i].name == name:
+                    del graph_def.node[i]
+                    break
+
     # Save as text.
     graph_def.save(outputPath)