dnn: EfficientDet

pull/17384/head
Dmitry Kurtaev 5 years ago
parent d5e8792f55
commit d9bada9867
  1. 11
      modules/dnn/perf/perf_net.cpp
  2. 28
      modules/dnn/src/tensorflow/tf_importer.cpp
  3. 33
      modules/dnn/test/test_tf_importer.cpp
  4. 2
      samples/dnn/tf_text_graph_common.py
  5. 236
      samples/dnn/tf_text_graph_efficientdet.py

@ -235,6 +235,17 @@ PERF_TEST_P_(DNNTestNetwork, Inception_v2_Faster_RCNN)
Mat(cv::Size(800, 600), CV_32FC3));
}
PERF_TEST_P_(DNNTestNetwork, EfficientDet)
{
if (backend == DNN_BACKEND_HALIDE || target != DNN_TARGET_CPU)
throw SkipTestException("");
Mat sample = imread(findDataFile("dnn/dog416.png"));
resize(sample, sample, Size(512, 512));
Mat inp;
sample.convertTo(inp, CV_32FC3, 1.0/255);
processNet("dnn/efficientdet-d0.pb", "dnn/efficientdet-d0.pbtxt", "", inp);
}
INSTANTIATE_TEST_CASE_P(/*nothing*/, DNNTestNetwork, dnnBackendsAndTargets());
} // namespace

@ -1542,22 +1542,32 @@ void TFImporter::populateNet(Net dstNet)
connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
}
else if (type == "Mul")
else if (type == "Mul" || type == "RealDiv")
{
bool haveConst = false;
for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)
int constId = -1;
for(int ii = 0; ii < layer.input_size(); ++ii)
{
Pin input = parsePin(layer.input(ii));
haveConst = value_id.find(input.name) != value_id.end();
if (value_id.find(input.name) != value_id.end())
{
constId = ii;
break;
}
CV_Assert(!haveConst || layer.input_size() == 2);
}
CV_Assert((constId != -1) || (layer.input_size() == 2));
if (haveConst)
if (constId != -1)
{
// Multiplication by constant.
CV_Assert(layer.input_size() == 2);
Mat scaleMat = getTensorContent(getConstBlob(layer, value_id));
CV_Assert(scaleMat.type() == CV_32FC1);
if (type == "RealDiv")
{
if (constId == 0)
CV_Error(Error::StsNotImplemented, "Division of constant over variable");
scaleMat = 1.0f / scaleMat;
}
int id;
if (scaleMat.total() == 1) // is a scalar.
@ -1659,11 +1669,15 @@ void TFImporter::populateNet(Net dstNet)
int id;
if (equalInpShapes || netInputShapes.empty())
{
layerParams.set("operation", "prod");
layerParams.set("operation", type == "RealDiv" ? "div" : "prod");
id = dstNet.addLayer(name, "Eltwise", layerParams);
}
else
{
if (type == "RealDiv")
CV_Error(Error::StsNotImplemented, "Division of non equal tensors");
id = dstNet.addLayer(name, "Scale", layerParams);
}
layer_id[name] = id;

@ -1123,4 +1123,37 @@ TEST_P(Test_TensorFlow_nets, Mask_RCNN)
expectNoFallbacks(net);
}
TEST_P(Test_TensorFlow_nets, EfficientDet)
{
if (target != DNN_TARGET_CPU)
{
if (target == DNN_TARGET_OPENCL_FP16) applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL_FP16);
if (target == DNN_TARGET_OPENCL) applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL);
if (target == DNN_TARGET_MYRIAD) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD);
}
checkBackend();
std::string proto = findDataFile("dnn/efficientdet-d0.pbtxt");
std::string model = findDataFile("dnn/efficientdet-d0.pb");
Net net = readNetFromTensorflow(model, proto);
Mat img = imread(findDataFile("dnn/dog416.png"));
Mat blob = blobFromImage(img, 1.0/255, Size(512, 512), Scalar(123.675, 116.28, 103.53));
net.setPreferableBackend(backend);
net.setPreferableTarget(target);
net.setInput(blob);
// Output has shape 1x1xNx7 where N - number of detections.
// An every detection is a vector of values [id, classId, confidence, left, top, right, bottom]
Mat out = net.forward();
// References are from test for TensorFlow model.
Mat ref = (Mat_<float>(3, 7) << 0, 1, 0.8437444, 0.153996080160141, 0.20534580945968628, 0.7463544607162476, 0.7414066195487976,
0, 17, 0.8245924, 0.16657517850399017, 0.3996818959712982, 0.4111558794975281, 0.9306337833404541,
0, 7, 0.8039304, 0.6118435263633728, 0.13175517320632935, 0.9065558314323425, 0.2943994700908661);
double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 4e-3 : 1e-5;
double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 2e-3 : 1e-4;
normAssertDetections(ref, out, "", 0.5, scoreDiff, iouDiff);
expectNoFallbacksFromIE(net);
}
}

@ -269,7 +269,7 @@ def parseTextGraph(filePath):
def removeIdentity(graph_def):
identities = {}
for node in graph_def.node:
if node.op == 'Identity':
if node.op == 'Identity' or node.op == 'IdentityN':
identities[node.name] = node.input[0]
graph_def.node.remove(node)

@ -0,0 +1,236 @@
# This file is a part of OpenCV project.
# It is a subject to the license terms in the LICENSE file found in the top-level directory
# of this distribution and at http://opencv.org/license.html.
#
# Copyright (C) 2020, Intel Corporation, all rights reserved.
# Third party copyrights are property of their respective owners.
#
# Use this script to get the text graph representation (.pbtxt) of EfficientDet
# deep learning network trained in https://github.com/google/automl.
# 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 AnchorGenerator:
def __init__(self, min_level, aspect_ratios, num_scales, anchor_scale):
self.min_level = min_level
self.aspect_ratios = aspect_ratios
self.anchor_scale = anchor_scale
self.scales = [2**(float(s) / num_scales) for s in range(num_scales)]
def get(self, layer_id):
widths = []
heights = []
for s in self.scales:
for a in self.aspect_ratios:
base_anchor_size = 2**(self.min_level + layer_id) * self.anchor_scale
heights.append(base_anchor_size * s * a[1])
widths.append(base_anchor_size * s * a[0])
return widths, heights
def createGraph(modelPath, outputPath, min_level, aspect_ratios, num_scales,
anchor_scale, num_classes, image_width, image_height):
print('Min level: %d' % min_level)
print('Anchor scale: %f' % anchor_scale)
print('Num scales: %d' % num_scales)
print('Aspect ratios: %s' % str(aspect_ratios))
print('Number of classes: %d' % num_classes)
print('Input image size: %dx%d' % (image_width, image_height))
# Read the graph.
_inpNames = ['image_arrays']
outNames = ['detections']
writeTextGraph(modelPath, outputPath, outNames)
graph_def = parseTextGraph(outputPath)
def getUnconnectedNodes():
unconnected = []
for node in graph_def.node:
if node.op == 'Const':
continue
unconnected.append(node.name)
for inp in node.input:
if inp in unconnected:
unconnected.remove(inp)
return unconnected
nodesToKeep = ['truediv'] # Keep preprocessing nodes
removeIdentity(graph_def)
scopesToKeep = ('image_arrays', 'efficientnet', 'resample_p6', 'resample_p7',
'fpn_cells', 'class_net', 'box_net', 'Reshape', 'concat')
addConstNode('scale_w', [2.0], graph_def)
addConstNode('scale_h', [2.0], graph_def)
nodesToKeep += ['scale_w', 'scale_h']
for node in graph_def.node:
if re.match('efficientnet-(.*)/blocks_\d+/se/mul_1', node.name):
node.input[0], node.input[1] = node.input[1], node.input[0]
if re.match('fpn_cells/cell_\d+/fnode\d+/resample(.*)/nearest_upsampling/Reshape_1$', node.name):
node.op = 'ResizeNearestNeighbor'
node.input[1] = 'scale_w'
node.input.append('scale_h')
for inpNode in graph_def.node:
if inpNode.name == node.name[:node.name.rfind('_')]:
node.input[0] = inpNode.input[0]
if re.match('box_net/box-predict(_\d)*/separable_conv2d$', node.name):
node.addAttr('loc_pred_transposed', True)
# Replace RealDiv to Mul with inversed scale for compatibility
if node.op == 'RealDiv':
for inpNode in graph_def.node:
if inpNode.name != node.input[1] or not 'value' in inpNode.attr:
continue
tensor = inpNode.attr['value']['tensor'][0]
if not 'float_val' in tensor:
continue
scale = float(inpNode.attr['value']['tensor'][0]['float_val'][0])
addConstNode(inpNode.name + '/inv', [1.0 / scale], graph_def)
nodesToKeep.append(inpNode.name + '/inv')
node.input[1] = inpNode.name + '/inv'
node.op = 'Mul'
break
def to_remove(name, op):
if name in nodesToKeep:
return False
return op == 'Const' or not name.startswith(scopesToKeep)
removeUnusedNodesAndAttrs(to_remove, graph_def)
# Attach unconnected preprocessing
assert(graph_def.node[1].name == 'truediv' and graph_def.node[1].op == 'RealDiv')
graph_def.node[1].input.insert(0, 'image_arrays')
graph_def.node[2].input.insert(0, 'truediv')
priors_generator = AnchorGenerator(min_level, aspect_ratios, num_scales, anchor_scale)
priorBoxes = []
for i in range(5):
inpName = ''
for node in graph_def.node:
if node.name == 'Reshape_%d' % (i * 2 + 1):
inpName = node.input[0]
break
priorBox = NodeDef()
priorBox.name = 'PriorBox_%d' % i
priorBox.op = 'PriorBox'
priorBox.input.append(inpName)
priorBox.input.append(graph_def.node[0].name) # image_tensor
priorBox.addAttr('flip', False)
priorBox.addAttr('clip', False)
widths, heights = priors_generator.get(i)
priorBox.addAttr('width', widths)
priorBox.addAttr('height', heights)
priorBox.addAttr('variance', [1.0, 1.0, 1.0, 1.0])
graph_def.node.extend([priorBox])
priorBoxes.append(priorBox.name)
addConstNode('concat/axis_flatten', [-1], graph_def)
def addConcatNode(name, inputs, axisNodeName):
concat = NodeDef()
concat.name = name
concat.op = 'ConcatV2'
for inp in inputs:
concat.input.append(inp)
concat.input.append(axisNodeName)
graph_def.node.extend([concat])
addConcatNode('PriorBox/concat', priorBoxes, 'concat/axis_flatten')
sigmoid = NodeDef()
sigmoid.name = 'concat/sigmoid'
sigmoid.op = 'Sigmoid'
sigmoid.input.append('concat')
graph_def.node.extend([sigmoid])
addFlatten(sigmoid.name, sigmoid.name + '/Flatten', graph_def)
addFlatten('concat_1', 'concat_1/Flatten', graph_def)
detectionOut = NodeDef()
detectionOut.name = 'detection_out'
detectionOut.op = 'DetectionOutput'
detectionOut.input.append('concat_1/Flatten')
detectionOut.input.append(sigmoid.name + '/Flatten')
detectionOut.input.append('PriorBox/concat')
detectionOut.addAttr('num_classes', num_classes)
detectionOut.addAttr('share_location', True)
detectionOut.addAttr('background_label_id', num_classes + 1)
detectionOut.addAttr('nms_threshold', 0.6)
detectionOut.addAttr('confidence_threshold', 0.2)
detectionOut.addAttr('top_k', 100)
detectionOut.addAttr('keep_top_k', 100)
detectionOut.addAttr('code_type', "CENTER_SIZE")
graph_def.node.extend([detectionOut])
graph_def.node[0].attr['shape'] = {
'shape': {
'dim': [
{'size': -1},
{'size': image_height},
{'size': image_width},
{'size': 3}
]
}
}
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)
if __name__ == "__main__":
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.')
parser.add_argument('--input', required=True, help='Path to frozen TensorFlow graph.')
parser.add_argument('--output', required=True, help='Path to output text graph.')
parser.add_argument('--min_level', default=3, type=int, help='Parameter from training config')
parser.add_argument('--num_scales', default=3, type=int, help='Parameter from training config')
parser.add_argument('--anchor_scale', default=4.0, type=float, help='Parameter from training config')
parser.add_argument('--aspect_ratios', default=[1.0, 1.0, 1.4, 0.7, 0.7, 1.4],
nargs='+', type=float, help='Parameter from training config')
parser.add_argument('--num_classes', default=90, type=int, help='Number of classes to detect')
parser.add_argument('--width', default=512, type=int, help='Network input width')
parser.add_argument('--height', default=512, type=int, help='Network input height')
args = parser.parse_args()
ar = args.aspect_ratios
assert(len(ar) % 2 == 0)
ar = list(zip(ar[::2], ar[1::2]))
createGraph(args.input, args.output, args.min_level, ar, args.num_scales,
args.anchor_scale, args.num_classes, args.width, args.height)
Loading…
Cancel
Save