Merge pull request #12082 from dkurt:dnn_ie_faster_rcnn

samples/dnn/tf_text_graph_common.py

@@ -0,0 +1,25 @@
+import tensorflow as tf
+from tensorflow.core.framework.node_def_pb2 import NodeDef
+from google.protobuf import text_format
+
+def tensorMsg(values):
+    if all([isinstance(v, float) for v in values]):
+        dtype = 'DT_FLOAT'
+        field = 'float_val'
+    elif all([isinstance(v, int) for v in values]):
+        dtype = 'DT_INT32'
+        field = 'int_val'
+    else:
+        raise Exception('Wrong values types')
+
+    msg = 'tensor { dtype: ' + dtype + ' tensor_shape { dim { size: %d } }' % len(values)
+    for value in values:
+        msg += '%s: %s ' % (field, str(value))
+    return msg + '}'
+
+def addConstNode(name, values, graph_def):
+    node = NodeDef()
+    node.name = name
+    node.op = 'Const'
+    text_format.Merge(tensorMsg(values), node.attr["value"])
+    graph_def.node.extend([node])

samples/dnn/tf_text_graph_faster_rcnn.py

@@ -6,6 +6,8 @@ from tensorflow.core.framework.node_def_pb2 import NodeDef
 from tensorflow.tools.graph_transforms import TransformGraph
 from google.protobuf import text_format
 
+from tf_text_graph_common import tensorMsg, addConstNode
+
 parser = argparse.ArgumentParser(description='Run this script to get a text graph of '
                                              'SSD model from TensorFlow Object Detection API. '
                                              'Then pass it with .pb file to cv::dnn::readNetFromTensorflow function.')
@@ -93,21 +95,6 @@ while True:
     if node.op == 'CropAndResize':
         break
 
-def tensorMsg(values):
-    if all([isinstance(v, float) for v in values]):
-        dtype = 'DT_FLOAT'
-        field = 'float_val'
-    elif all([isinstance(v, int) for v in values]):
-        dtype = 'DT_INT32'
-        field = 'int_val'
-    else:
-        raise Exception('Wrong values types')
-
-    msg = 'tensor { dtype: ' + dtype + ' tensor_shape { dim { size: %d } }' % len(values)
-    for value in values:
-        msg += '%s: %s ' % (field, str(value))
-    return msg + '}'
-
 def addSlice(inp, out, begins, sizes):
     beginsNode = NodeDef()
     beginsNode.name = out + '/begins'
@@ -151,17 +138,25 @@ def addSoftMax(inp, out):
     softmax.input.append(inp)
     graph_def.node.extend([softmax])
 
+def addFlatten(inp, out):
+    flatten = NodeDef()
+    flatten.name = out
+    flatten.op = 'Flatten'
+    flatten.input.append(inp)
+    graph_def.node.extend([flatten])
+
 addReshape('FirstStageBoxPredictor/ClassPredictor/BiasAdd',
            'FirstStageBoxPredictor/ClassPredictor/reshape_1', [0, -1, 2])
 
 addSoftMax('FirstStageBoxPredictor/ClassPredictor/reshape_1',
            'FirstStageBoxPredictor/ClassPredictor/softmax')  # Compare with Reshape_4
 
-flatten = NodeDef()
-flatten.name = 'FirstStageBoxPredictor/BoxEncodingPredictor/flatten'  # Compare with FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd
-flatten.op = 'Flatten'
-flatten.input.append('FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd')
-graph_def.node.extend([flatten])
+addFlatten('FirstStageBoxPredictor/ClassPredictor/softmax',
+           'FirstStageBoxPredictor/ClassPredictor/softmax/flatten')
+
+# Compare with FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd
+addFlatten('FirstStageBoxPredictor/BoxEncodingPredictor/BiasAdd',
+           'FirstStageBoxPredictor/BoxEncodingPredictor/flatten')
 
 proposals = NodeDef()
 proposals.name = 'proposals'  # Compare with ClipToWindow/Gather/Gather (NOTE: normalized)
@@ -194,7 +189,7 @@ detectionOut.name = 'detection_out'
 detectionOut.op = 'DetectionOutput'
 
 detectionOut.input.append('FirstStageBoxPredictor/BoxEncodingPredictor/flatten')
-detectionOut.input.append('FirstStageBoxPredictor/ClassPredictor/softmax')
+detectionOut.input.append('FirstStageBoxPredictor/ClassPredictor/softmax/flatten')
 detectionOut.input.append('proposals')
 
 text_format.Merge('i: 2', detectionOut.attr['num_classes'])
@@ -204,11 +199,21 @@ text_format.Merge('f: 0.7', detectionOut.attr['nms_threshold'])
 text_format.Merge('i: 6000', detectionOut.attr['top_k'])
 text_format.Merge('s: "CENTER_SIZE"', detectionOut.attr['code_type'])
 text_format.Merge('i: 100', detectionOut.attr['keep_top_k'])
-text_format.Merge('b: true', detectionOut.attr['clip'])
-text_format.Merge('b: true', detectionOut.attr['loc_pred_transposed'])
+text_format.Merge('b: false', detectionOut.attr['clip'])
 
 graph_def.node.extend([detectionOut])
 
+addConstNode('clip_by_value/lower', [0.0], graph_def)
+addConstNode('clip_by_value/upper', [1.0], graph_def)
+
+clipByValueNode = NodeDef()
+clipByValueNode.name = 'detection_out/clip_by_value'
+clipByValueNode.op = 'ClipByValue'
+clipByValueNode.input.append('detection_out')
+clipByValueNode.input.append('clip_by_value/lower')
+clipByValueNode.input.append('clip_by_value/upper')
+graph_def.node.extend([clipByValueNode])
+
 # Save as text.
 for node in reversed(topNodes):
     graph_def.node.extend([node])
@@ -225,17 +230,13 @@ addReshape('SecondStageBoxPredictor/Reshape_1/slice',
 # Replace Flatten subgraph onto a single node.
 for i in reversed(range(len(graph_def.node))):
     if graph_def.node[i].op == 'CropAndResize':
-        graph_def.node[i].input.insert(1, 'detection_out')
+        graph_def.node[i].input.insert(1, 'detection_out/clip_by_value')
 
     if graph_def.node[i].name == 'SecondStageBoxPredictor/Reshape':
-        shapeNode = NodeDef()
-        shapeNode.name = 'SecondStageBoxPredictor/Reshape/shape2'
-        shapeNode.op = 'Const'
-        text_format.Merge(tensorMsg([1, -1, 4]), shapeNode.attr["value"])
-        graph_def.node.extend([shapeNode])
+        addConstNode('SecondStageBoxPredictor/Reshape/shape2', [1, -1, 4], graph_def)
 
         graph_def.node[i].input.pop()
-        graph_def.node[i].input.append(shapeNode.name)
+        graph_def.node[i].input.append('SecondStageBoxPredictor/Reshape/shape2')
 
     if graph_def.node[i].name in ['SecondStageBoxPredictor/Flatten/flatten/Shape',
                                   'SecondStageBoxPredictor/Flatten/flatten/strided_slice',
@@ -246,12 +247,15 @@ for node in graph_def.node:
     if node.name == 'SecondStageBoxPredictor/Flatten/flatten/Reshape':
         node.op = 'Flatten'
         node.input.pop()
-        break
+
+    if node.name in ['FirstStageBoxPredictor/BoxEncodingPredictor/Conv2D',
+                     'SecondStageBoxPredictor/BoxEncodingPredictor/MatMul']:
+        text_format.Merge('b: true', node.attr["loc_pred_transposed"])
 
 ################################################################################
 ### Postprocessing
 ################################################################################
-addSlice('detection_out', 'detection_out/slice', [0, 0, 0, 3], [-1, -1, -1, 4])
+addSlice('detection_out/clip_by_value', 'detection_out/slice', [0, 0, 0, 3], [-1, -1, -1, 4])
 
 variance = NodeDef()
 variance.name = 'proposals/variance'
@@ -268,12 +272,13 @@ text_format.Merge('i: 2', varianceEncoder.attr["axis"])
 graph_def.node.extend([varianceEncoder])
 
 addReshape('detection_out/slice', 'detection_out/slice/reshape', [1, 1, -1])
+addFlatten('variance_encoded', 'variance_encoded/flatten')
 
 detectionOut = NodeDef()
 detectionOut.name = 'detection_out_final'
 detectionOut.op = 'DetectionOutput'
 
-detectionOut.input.append('variance_encoded')
+detectionOut.input.append('variance_encoded/flatten')
 detectionOut.input.append('SecondStageBoxPredictor/Reshape_1/Reshape')
 detectionOut.input.append('detection_out/slice/reshape')
 
@@ -283,7 +288,6 @@ text_format.Merge('i: %d' % (args.num_classes + 1), detectionOut.attr['backgroun
 text_format.Merge('f: 0.6', detectionOut.attr['nms_threshold'])
 text_format.Merge('s: "CENTER_SIZE"', detectionOut.attr['code_type'])
 text_format.Merge('i: 100', detectionOut.attr['keep_top_k'])
-text_format.Merge('b: true', detectionOut.attr['loc_pred_transposed'])
 text_format.Merge('b: true', detectionOut.attr['clip'])
 text_format.Merge('b: true', detectionOut.attr['variance_encoded_in_target'])
 graph_def.node.extend([detectionOut])

samples/dnn/tf_text_graph_ssd.py

@@ -15,6 +15,7 @@ from math import sqrt
 from tensorflow.core.framework.node_def_pb2 import NodeDef
 from tensorflow.tools.graph_transforms import TransformGraph
 from google.protobuf import text_format
+from tf_text_graph_common import tensorMsg, addConstNode
 
 parser = argparse.ArgumentParser(description='Run this script to get a text graph of '
                                              'SSD model from TensorFlow Object Detection API. '
@@ -160,28 +161,6 @@ graph_def.node[1].input.append(weights)
 # Create SSD postprocessing head ###############################################
 
 # Concatenate predictions of classes, predictions of bounding boxes and proposals.
-def tensorMsg(values):
-    if all([isinstance(v, float) for v in values]):
-        dtype = 'DT_FLOAT'
-        field = 'float_val'
-    elif all([isinstance(v, int) for v in values]):
-        dtype = 'DT_INT32'
-        field = 'int_val'
-    else:
-        raise Exception('Wrong values types')
-
-    msg = 'tensor { dtype: ' + dtype + ' tensor_shape { dim { size: %d } }' % len(values)
-    for value in values:
-        msg += '%s: %s ' % (field, str(value))
-    return msg + '}'
-
-def addConstNode(name, values):
-    node = NodeDef()
-    node.name = name
-    node.op = 'Const'
-    text_format.Merge(tensorMsg(values), node.attr["value"])
-    graph_def.node.extend([node])
-
 def addConcatNode(name, inputs, axisNodeName):
     concat = NodeDef()
     concat.name = name