Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

Merge pull request #13614 from dkurt:dnn_tf_ssd_fpn

Browse Source
pull/13539/head
Alexander Alekhin 6 years ago
parent 78bd55c8df 4ae5df5509
commit 6e39856623
4 changed files with 132 additions and 42 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 2
      samples/dnn/tf_text_graph_common.py
  2. 2
      samples/dnn/tf_text_graph_faster_rcnn.py
  3. 2
      samples/dnn/tf_text_graph_mask_rcnn.py
  4. 158
      samples/dnn/tf_text_graph_ssd.py

2
samples/dnn/tf_text_graph_common.py
Unescape View File

@ -289,7 +289,7 @@ def removeUnusedNodesAndAttrs(to_remove, graph_def):
op = graph_def.node[i].op
name = graph_def.node[i].name
if op == 'Const' or to_remove(name, op):
if to_remove(name, op):
if op != 'Const':
removedNodes.append(name)

2
samples/dnn/tf_text_graph_faster_rcnn.py
Unescape View File

@ -49,7 +49,7 @@ def createFasterRCNNGraph(modelPath, configPath, outputPath):
removeIdentity(graph_def)
def to_remove(name, op):
return name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep) or \
return op == 'Const' or name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep) or \
(name.startswith('CropAndResize') and op != 'CropAndResize')
removeUnusedNodesAndAttrs(to_remove, graph_def)

2
samples/dnn/tf_text_graph_mask_rcnn.py
Unescape View File

@ -55,7 +55,7 @@ graph_def = parseTextGraph(args.output)
removeIdentity(graph_def)
def to_remove(name, op):
return name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep) or \
return op == 'Const' or name.startswith(scopesToIgnore) or not name.startswith(scopesToKeep) or \
(name.startswith('CropAndResize') and op != 'CropAndResize')
removeUnusedNodesAndAttrs(to_remove, graph_def)

158
samples/dnn/tf_text_graph_ssd.py
Unescape View File

@ -10,14 +10,60 @@
# Then you can import it with a binary frozen graph (.pb) using readNetFromTensorflow() function.
# See details and examples on the following wiki page: https://github.com/opencv/opencv/wiki/TensorFlow-Object-Detection-API
import argparse
import re
from math import sqrt
from tf_text_graph_common import *
class SSDAnchorGenerator:
def __init__(self, min_scale, max_scale, num_layers, aspect_ratios,
reduce_boxes_in_lowest_layer, image_width, image_height):
self.min_scale = min_scale
self.aspect_ratios = aspect_ratios
self.reduce_boxes_in_lowest_layer = reduce_boxes_in_lowest_layer
self.image_width = image_width
self.image_height = image_height
self.scales = [min_scale + (max_scale - min_scale) * i / (num_layers - 1)
for i in range(num_layers)] + [1.0]
def get(self, layer_id):
if layer_id == 0 and self.reduce_boxes_in_lowest_layer:
widths = [0.1, self.min_scale * sqrt(2.0), self.min_scale * sqrt(0.5)]
heights = [0.1, self.min_scale / sqrt(2.0), self.min_scale / sqrt(0.5)]
else:
widths = [self.scales[layer_id] * sqrt(ar) for ar in self.aspect_ratios]
heights = [self.scales[layer_id] / sqrt(ar) for ar in self.aspect_ratios]
widths += [sqrt(self.scales[layer_id] * self.scales[layer_id + 1])]
heights += [sqrt(self.scales[layer_id] * self.scales[layer_id + 1])]
widths = [w * self.image_width for w in widths]
heights = [h * self.image_height for h in heights]
return widths, heights
class MultiscaleAnchorGenerator:
def __init__(self, min_level, aspect_ratios, scales_per_octave, anchor_scale):
self.min_level = min_level
self.aspect_ratios = aspect_ratios
self.anchor_scale = anchor_scale
self.scales = [2**(float(s) / scales_per_octave) for s in range(scales_per_octave)]
def get(self, layer_id):
widths = []
heights = []
for a in self.aspect_ratios:
for s in self.scales:
base_anchor_size = 2**(self.min_level + layer_id) * self.anchor_scale
ar = sqrt(a)
heights.append(base_anchor_size * s / ar)
widths.append(base_anchor_size * s * ar)
return widths, heights
def createSSDGraph(modelPath, configPath, outputPath):
# Nodes that should be kept.
keepOps = ['Conv2D', 'BiasAdd', 'Add', 'Relu6', 'Placeholder', 'FusedBatchNorm',
keepOps = ['Conv2D', 'BiasAdd', 'Add', 'Relu', 'Relu6', 'Placeholder', 'FusedBatchNorm',
'DepthwiseConv2dNative', 'ConcatV2', 'Mul', 'MaxPool', 'AvgPool', 'Identity',
'Sub']
'Sub', 'ResizeNearestNeighbor', 'Pad']
# Node with which prefixes should be removed
prefixesToRemove = ('MultipleGridAnchorGenerator/', 'Postprocessor/', 'Preprocessor/map')
@ -27,7 +73,15 @@ def createSSDGraph(modelPath, configPath, outputPath):
config = config['model'][0]['ssd'][0]
num_classes = int(config['num_classes'][0])
ssd_anchor_generator = config['anchor_generator'][0]['ssd_anchor_generator'][0]
fixed_shape_resizer = config['image_resizer'][0]['fixed_shape_resizer'][0]
image_width = int(fixed_shape_resizer['width'][0])
image_height = int(fixed_shape_resizer['height'][0])
box_predictor = 'convolutional' if 'convolutional_box_predictor' in config['box_predictor'][0] else 'weight_shared_convolutional'
anchor_generator = config['anchor_generator'][0]
if 'ssd_anchor_generator' in anchor_generator:
ssd_anchor_generator = anchor_generator['ssd_anchor_generator'][0]
min_scale = float(ssd_anchor_generator['min_scale'][0])
max_scale = float(ssd_anchor_generator['max_scale'][0])
num_layers = int(ssd_anchor_generator['num_layers'][0])
@ -35,18 +89,34 @@ def createSSDGraph(modelPath, configPath, outputPath):
reduce_boxes_in_lowest_layer = True
if 'reduce_boxes_in_lowest_layer' in ssd_anchor_generator:
reduce_boxes_in_lowest_layer = ssd_anchor_generator['reduce_boxes_in_lowest_layer'][0] == 'true'
priors_generator = SSDAnchorGenerator(min_scale, max_scale, num_layers,
aspect_ratios, reduce_boxes_in_lowest_layer,
image_width, image_height)
fixed_shape_resizer = config['image_resizer'][0]['fixed_shape_resizer'][0]
image_width = int(fixed_shape_resizer['width'][0])
image_height = int(fixed_shape_resizer['height'][0])
box_predictor = 'convolutional' if 'convolutional_box_predictor' in config['box_predictor'][0] else 'weight_shared_convolutional'
print('Number of classes: %d' % num_classes)
print('Number of layers: %d' % num_layers)
print('Scale: [%f-%f]' % (min_scale, max_scale))
print('Aspect ratios: %s' % str(aspect_ratios))
print('Reduce boxes in the lowest layer: %s' % str(reduce_boxes_in_lowest_layer))
elif 'multiscale_anchor_generator' in anchor_generator:
multiscale_anchor_generator = anchor_generator['multiscale_anchor_generator'][0]
min_level = int(multiscale_anchor_generator['min_level'][0])
max_level = int(multiscale_anchor_generator['max_level'][0])
anchor_scale = float(multiscale_anchor_generator['anchor_scale'][0])
aspect_ratios = [float(ar) for ar in multiscale_anchor_generator['aspect_ratios']]
scales_per_octave = int(multiscale_anchor_generator['scales_per_octave'][0])
num_layers = max_level - min_level + 1
priors_generator = MultiscaleAnchorGenerator(min_level, aspect_ratios,
scales_per_octave, anchor_scale)
print('Levels: [%d-%d]' % (min_level, max_level))
print('Anchor scale: %f' % anchor_scale)
print('Scales per octave: %d' % scales_per_octave)
print('Aspect ratios: %s' % str(aspect_ratios))
else:
print('Unknown anchor_generator')
exit(0)
print('Number of classes: %d' % num_classes)
print('Number of layers: %d' % num_layers)
print('box predictor: %s' % box_predictor)
print('Input image size: %dx%d' % (image_width, image_height))
@ -67,8 +137,8 @@ def createSSDGraph(modelPath, configPath, outputPath):
return unconnected
def fuse_nodes(nodesToKeep):
# Detect unfused batch normalization nodes and fuse them.
def fuse_batch_normalization():
# Add_0 <-- moving_variance, add_y
# Rsqrt <-- Add_0
# Mul_0 <-- Rsqrt, gamma
@ -77,9 +147,15 @@ def createSSDGraph(modelPath, configPath, outputPath):
# Sub_0 <-- beta, Mul_2
# Add_1 <-- Mul_1, Sub_0
nodesMap = {node.name: node for node in graph_def.node}
subgraph = ['Add',
subgraphBatchNorm = ['Add',
['Mul', 'input', ['Mul', ['Rsqrt', ['Add', 'moving_variance', 'add_y']], 'gamma']],
['Sub', 'beta', ['Mul', 'moving_mean', 'Mul_0']]]
# Detect unfused nearest neighbor resize.
subgraphResizeNN = ['Reshape',
['Mul', ['Reshape', 'input', ['Pack', 'shape_1', 'shape_2', 'shape_3', 'shape_4', 'shape_5']],
'ones'],
['Pack', ['StridedSlice', ['Shape', 'input'], 'stack', 'stack_1', 'stack_2'],
'out_height', 'out_width', 'out_channels']]
def checkSubgraph(node, targetNode, inputs, fusedNodes):
op = targetNode[0]
if node.op == op and (len(node.input) >= len(targetNode) - 1):
@ -100,7 +176,7 @@ def createSSDGraph(modelPath, configPath, outputPath):
for node in graph_def.node:
inputs = {}
fusedNodes = []
if checkSubgraph(node, subgraph, inputs, fusedNodes):
if checkSubgraph(node, subgraphBatchNorm, inputs, fusedNodes):
name = node.name
node.Clear()
node.name = name
@ -112,15 +188,41 @@ def createSSDGraph(modelPath, configPath, outputPath):
node.input.append(inputs['moving_variance'])
node.addAttr('epsilon', 0.001)
nodesToRemove += fusedNodes[1:]
inputs = {}
fusedNodes = []
if checkSubgraph(node, subgraphResizeNN, inputs, fusedNodes):
name = node.name
node.Clear()
node.name = name
node.op = 'ResizeNearestNeighbor'
node.input.append(inputs['input'])
node.input.append(name + '/output_shape')
out_height_node = nodesMap[inputs['out_height']]
out_width_node = nodesMap[inputs['out_width']]
out_height = int(out_height_node.attr['value']['tensor'][0]['int_val'][0])
out_width = int(out_width_node.attr['value']['tensor'][0]['int_val'][0])
shapeNode = NodeDef()
shapeNode.name = name + '/output_shape'
shapeNode.op = 'Const'
shapeNode.addAttr('value', [out_height, out_width])
graph_def.node.insert(graph_def.node.index(node), shapeNode)
nodesToKeep.append(shapeNode.name)
nodesToRemove += fusedNodes[1:]
for node in nodesToRemove:
graph_def.node.remove(node)
fuse_batch_normalization()
nodesToKeep = []
fuse_nodes(nodesToKeep)
removeIdentity(graph_def)
def to_remove(name, op):
return (not op in keepOps) or name.startswith(prefixesToRemove)
return (not name in nodesToKeep) and \
(op == 'Const' or (not op in keepOps) or name.startswith(prefixesToRemove))
removeUnusedNodesAndAttrs(to_remove, graph_def)
@ -169,19 +271,15 @@ def createSSDGraph(modelPath, configPath, outputPath):
graph_def.node.extend([flatten])
addConcatNode('%s/concat' % label, concatInputs, 'concat/axis_flatten')
idx = 0
num_matched_layers = 0
for node in graph_def.node:
if node.name == ('BoxPredictor_%d/BoxEncodingPredictor/Conv2D' % idx) or \
node.name == ('WeightSharedConvolutionalBoxPredictor_%d/BoxPredictor/Conv2D' % idx) or \
node.name == 'WeightSharedConvolutionalBoxPredictor/BoxPredictor/Conv2D':
if re.match('BoxPredictor_\d/BoxEncodingPredictor/Conv2D', node.name) or \
re.match('WeightSharedConvolutionalBoxPredictor(_\d)*/BoxPredictor/Conv2D', node.name):
node.addAttr('loc_pred_transposed', True)
idx += 1
assert(idx == num_layers)
num_matched_layers += 1
assert(num_matched_layers == num_layers)
# Add layers that generate anchors (bounding boxes proposals).
scales = [min_scale + (max_scale - min_scale) * i / (num_layers - 1)
for i in range(num_layers)] + [1.0]
priorBoxes = []
for i in range(num_layers):
priorBox = NodeDef()
@ -199,17 +297,8 @@ def createSSDGraph(modelPath, configPath, outputPath):
priorBox.addAttr('flip', False)
priorBox.addAttr('clip', False)
if i == 0 and reduce_boxes_in_lowest_layer:
widths = [0.1, min_scale * sqrt(2.0), min_scale * sqrt(0.5)]
heights = [0.1, min_scale / sqrt(2.0), min_scale / sqrt(0.5)]
else:
widths = [scales[i] * sqrt(ar) for ar in aspect_ratios]
heights = [scales[i] / sqrt(ar) for ar in aspect_ratios]
widths, heights = priors_generator.get(i)
widths += [sqrt(scales[i] * scales[i + 1])]
heights += [sqrt(scales[i] * scales[i + 1])]
widths = [w * image_width for w in widths]
heights = [h * image_height for h in heights]
priorBox.addAttr('width', widths)
priorBox.addAttr('height', heights)
priorBox.addAttr('variance', [0.1, 0.1, 0.2, 0.2])
@ -217,6 +306,7 @@ def createSSDGraph(modelPath, configPath, outputPath):
graph_def.node.extend([priorBox])
priorBoxes.append(priorBox.name)
# Compare this layer's output with Postprocessor/Reshape
addConcatNode('PriorBox/concat', priorBoxes, 'concat/axis_flatten')
# Sigmoid for classes predictions and DetectionOutput layer

Write Preview
Loading…
Cancel
Save