IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[Python API] setting and getting preprocessing info through InferRequest (#1009)

Browse Source
This commit is contained in:
Anastasia Kuporosova 2020-07-08 11:25:54 +03:00 committed by GitHub
parent 8ff7e3381d
commit b553b6ea17
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
9 changed files with 135 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
inference-engine/ie_bridges/python/src/openvino/inference_engine/__init__.py
View File

@ -1,4 +1,4 @@
from .ie_api import *
__all__ = ['IENetwork', "TensorDesc", "IECore", "Blob", "get_version"]
__all__ = ['IENetwork', "TensorDesc", "IECore", "Blob", "PreProcessInfo", "get_version"]
__version__ = get_version()

1
inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api.pxd
View File

@ -75,6 +75,7 @@ cdef class InputInfoCPtr:
cdef class PreProcessInfo:
cdef CPreProcessInfo* _ptr
cpdef object _user_data
cdef class PreProcessChannel:
cdef CPreProcessChannel.Ptr _ptr

40
inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api.pyx
View File

@ -503,6 +503,14 @@ cdef class PreProcessChannel:
## This class stores pre-process information for the input
cdef class PreProcessInfo:
def __cinit__(self):
self._ptr = new CPreProcessInfo()
self._user_data = True
def __dealloc__(self):
if self._user_data:
del self._ptr
def __getitem__(self, size_t index):
cdef CPreProcessChannel.Ptr c_channel = deref(self._ptr)[index]
channel = PreProcessChannel()
@ -544,10 +552,18 @@ cdef class PreProcessInfo:
## Resize Algorithm to be applied for input before inference if needed.
#
# \note It's need to set your input via the set_blob method.
#
# Usage example:\n
# ```python
# net = ie_core.read_network(model=path_to_xml_file, weights=path_to_bin_file)
# net.input_info['data'].preprocess_info.resize_algorithm = ResizeAlgorithm.RESIZE_BILINEAR
# exec_net = ie_core.load_network(net, 'CPU')
# tensor_desc = ie.TensorDesc("FP32", [1, 3, image.shape[2], image.shape[3]], "NCHW")
# img_blob = ie.Blob(tensor_desc, image)
# request = exec_net.requests[0]
# request.set_blob('data', img_blob)
# request.infer()
# ```
@property
def resize_algorithm(self):
@ -615,6 +631,8 @@ cdef class InputInfoPtr:
def preprocess_info(self):
cdef CPreProcessInfo* c_preprocess_info = &deref(self._ptr).getPreProcess()
preprocess_info = PreProcessInfo()
del preprocess_info._ptr
preprocess_info._user_data = False
preprocess_info._ptr = c_preprocess_info
return preprocess_info
@ -1028,9 +1046,24 @@ cdef class InferRequest:
output_blobs[output] = deepcopy(blob)
return output_blobs
## Dictionary that maps input layer names to corresponding preprocessing information
@property
def preprocess_info(self):
preprocess_info = {}
cdef const CPreProcessInfo** c_preprocess_info
for input_blob in self.input_blobs.keys():
preprocess = PreProcessInfo()
del preprocess._ptr
preprocess._user_data = False
c_preprocess_info = <const CPreProcessInfo**>(&preprocess._ptr)
deref(self.impl).getPreProcess(input_blob.encode(), c_preprocess_info)
preprocess_info[input_blob] = preprocess
return preprocess_info
## Sets user defined Blob for the infer request
# @param blob_name: A name of input blob
# @param blob: Blob object to set for the infer request
# @param preprocess_info: PreProcessInfo object to set for the infer request.
# @return None
#
# Usage example:\n
@ -1043,8 +1076,11 @@ cdef class InferRequest:
# blob = Blob(td, blob_data)
# exec_net.requests[0].set_blob(blob_name="input_blob_name", blob=blob),
# ```
def set_blob(self, blob_name : str, blob : Blob):
deref(self.impl).setBlob(blob_name.encode(), blob._ptr)
def set_blob(self, blob_name : str, blob : Blob, preprocess_info: PreProcessInfo = None):
if preprocess_info:
deref(self.impl).setBlob(blob_name.encode(), blob._ptr, deref(preprocess_info._ptr))
else:
deref(self.impl).setBlob(blob_name.encode(), blob._ptr)
self._user_blobs[blob_name] = blob
## Starts synchronous inference of the infer request and fill outputs array
#

13
inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api_impl.cpp
View File

@ -362,6 +362,19 @@ void InferenceEnginePython::InferRequestWrap::setBlob(const std::string &blob_na
IE_CHECK_CALL(request_ptr->SetBlob(blob_name.c_str(), blob_ptr, &response));
}
void InferenceEnginePython::InferRequestWrap::setBlob(const std::string &blob_name,
const InferenceEngine::Blob::Ptr &blob_ptr,
const InferenceEngine::PreProcessInfo& info) {
InferenceEngine::ResponseDesc response;
IE_CHECK_CALL(request_ptr->SetBlob(blob_name.c_str(), blob_ptr, info, &response));
}
void InferenceEnginePython::InferRequestWrap::getPreProcess(const std::string& blob_name,
const InferenceEngine::PreProcessInfo** info) {
InferenceEngine::ResponseDesc response;
IE_CHECK_CALL(request_ptr->GetPreProcess(blob_name.c_str(), info, &response));
}
void InferenceEnginePython::InferRequestWrap::getBlobPtr(const std::string &blob_name,
InferenceEngine::Blob::Ptr &blob_ptr) {
InferenceEngine::ResponseDesc response;

5
inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api_impl.hpp
View File

@ -114,8 +114,13 @@ struct InferRequestWrap {
void setBlob(const std::string &blob_name, const InferenceEngine::Blob::Ptr &blob_ptr);
void setBlob(const std::string &name, const InferenceEngine::Blob::Ptr &data,
const InferenceEngine::PreProcessInfo& info);
void setBatch(int size);
void getPreProcess(const std::string& blob_name, const InferenceEngine::PreProcessInfo** info);
std::map<std::string, InferenceEnginePython::ProfileInfo> getPerformanceCounts();
};

2
inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api_impl_defs.pxd
View File

@ -200,6 +200,8 @@ cdef extern from "ie_api_impl.hpp" namespace "InferenceEnginePython":
int index;
void getBlobPtr(const string & blob_name, CBlob.Ptr & blob_ptr) except +
void setBlob(const string & blob_name, const CBlob.Ptr & blob_ptr) except +
void setBlob(const string &blob_name, const CBlob.Ptr &blob_ptr, CPreProcessInfo& info) except +
void getPreProcess(const string& blob_name, const CPreProcessInfo** info) except +
map[string, ProfileInfo] getPerformanceCounts() except +
void infer() except +
void infer_async() except +

4
inference-engine/ie_bridges/python/tests/test_Blob.py
View File

@ -122,11 +122,11 @@ def test_incompatible_array_and_td():
def test_incompatible_input_precision():
import cv2
n, c, h, w = (1, 3, 32, 32)
image = cv2.imread(path_to_image) / 255
image = cv2.imread(path_to_image)
if image is None:
raise FileNotFoundError("Input image not found")
image = cv2.resize(image, (h, w))
image = cv2.resize(image, (h, w)) / 255
image = image.transpose((2, 0, 1))
image = image.reshape((n, c, h, w))
tensor_desc = TensorDesc("FP32", [1, 3, 32, 32], "NCHW")

64
inference-engine/ie_bridges/python/tests/test_InferRequest.py
View File

@ -16,11 +16,11 @@ path_to_img = image_path()
def read_image():
import cv2
n, c, h, w = (1, 3, 32, 32)
image = cv2.imread(path_to_img) / 255
image = cv2.imread(path_to_img)
if image is None:
raise FileNotFoundError("Input image not found")
image = cv2.resize(image, (h, w))
image = cv2.resize(image, (h, w)) / 255
image = image.transpose((2, 0, 1)).astype(np.float32)
image = image.reshape((n, c, h, w))
return image
@ -452,3 +452,63 @@ def test_blob_setter(device):
request.infer()
res_2 = np.sort(request.output_blobs['fc_out'].buffer)
assert np.allclose(res_1, res_2, atol=1e-2, rtol=1e-2)
def test_blob_setter_with_preprocess(device):
ie_core = ie.IECore()
net = ie_core.read_network(test_net_xml, test_net_bin)
exec_net = ie_core.load_network(network=net, device_name=device, num_requests=1)
img = read_image()
tensor_desc = ie.TensorDesc("FP32", [1, 3, 32, 32], "NCHW")
img_blob = ie.Blob(tensor_desc, img)
preprocess_info = ie.PreProcessInfo()
preprocess_info.mean_variant = ie.MeanVariant.MEAN_IMAGE
request = exec_net.requests[0]
request.set_blob('data', img_blob, preprocess_info)
pp = request.preprocess_info["data"]
assert pp.mean_variant == ie.MeanVariant.MEAN_IMAGE
def test_getting_preprocess(device):
ie_core = ie.IECore()
net = ie_core.read_network(test_net_xml, test_net_bin)
exec_net = ie_core.load_network(network=net, device_name=device, num_requests=1)
request = exec_net.requests[0]
preprocess_info = request.preprocess_info["data"]
assert isinstance(preprocess_info, ie.PreProcessInfo)
assert preprocess_info.mean_variant == ie.MeanVariant.NONE
def test_resize_algorithm_work(device):
ie_core = ie.IECore()
net = ie_core.read_network(test_net_xml, test_net_bin)
exec_net_1 = ie_core.load_network(network=net, device_name=device, num_requests=1)
img = read_image()
res_1 = np.sort(exec_net_1.infer({"data": img})['fc_out'])
net.input_info['data'].preprocess_info.resize_algorithm = ie.ResizeAlgorithm.RESIZE_BILINEAR
exec_net_2 = ie_core.load_network(net, 'CPU')
import cv2
image = cv2.imread(path_to_img)
if image is None:
raise FileNotFoundError("Input image not found")
image = image / 255
image = image.transpose((2, 0, 1)).astype(np.float32)
image = np.expand_dims(image, 0)
tensor_desc = ie.TensorDesc("FP32", [1, 3, image.shape[2], image.shape[3]], "NCHW")
img_blob = ie.Blob(tensor_desc, image)
request = exec_net_2.requests[0]
assert request.preprocess_info["data"].resize_algorithm == ie.ResizeAlgorithm.RESIZE_BILINEAR
request.set_blob('data', img_blob)
request.infer()
res_2 = np.sort(request.output_blobs['fc_out'].buffer)
assert np.allclose(res_1, res_2, atol=1e-2, rtol=1e-2)

11
inference-engine/ie_bridges/python/tests/test_PreProcessInfo.py
View File

@ -103,3 +103,14 @@ def test_set_mean_image_for_channel():
assert isinstance(pre_process_channel.mean_data, Blob)
assert pre_process_channel.mean_data.tensor_desc.dims == [127, 127]
assert preprocess_info.mean_variant == MeanVariant.MEAN_IMAGE
def test_resize_algorithm_set(device):
ie_core = IECore()
net = ie_core.read_network(model=test_net_xml, weights=test_net_bin)
preprocess_info = net.input_info["data"].preprocess_info
preprocess_info.resize_algorithm = ResizeAlgorithm.RESIZE_BILINEAR
exec_net = ie_core.load_network(network=net, device_name=device, num_requests=1)
request = exec_net.requests[0]
pp = request.preprocess_info["data"]
assert pp.resize_algorithm == ResizeAlgorithm.RESIZE_BILINEAR