def Inference(self, x, y):
x = x.reshape(
1024, 1024).copy() #make sure memory order is the same as np order
y = y.reshape(
1024, 3072).copy() #make sure memory order is the same as np order
sizes = [1024 * 1024, 1024 * 3072]
input_tensor_list = []
input_tensor_list.append(
hiai.NNTensor(x,
height=1024,
width=1024,
channel=1,
name='x',
data_type=DataType.FLOAT32_T,
size=sizes[0]))
input_tensor_list.append(
hiai.NNTensor(y,
height=1024,
width=3072,
channel=1,
name='y',
data_type=DataType.FLOAT32_T,
size=sizes[1]))
nntensorList = hiai.NNTensorList(
[input_tensor_list[0], input_tensor_list[1]])
if not nntensorList:
print(" matmul nntensorList is null")
resultList = self.model.Inference(self.graph, nntensorList)
# pdb.set_tracce()
result = resultList[0]
return result
def main():
inferenceModel = hiai.AIModelDescription('ShadowGAN', omModelName)
myGraph = CreateGraph(inferenceModel)
if myGraph is None:
exit(0)
if not osp.exists(outputDir):
os.makedirs(outputDir)
# Read images and corresponding masks
inputImage = ReadImage(osp.join(testDataDir, 'noshadow', 'demo.jpg'),
channels=3)
if inputImage is None:
print("[ERROR] No input image.")
DestroyGraph()
exit(0)
inputMask = ReadImage(osp.join(testDataDir, 'mask', 'demo.jpg'),
channels=1)
if inputMask is None:
print("[ERROR] No input mask.")
DestroyGraph()
exit(0)
# Normalize to [-1.0, 1.0]
inputImage = inputImage / 127.5 - 1.0
inputMask = 1.0 - inputMask / 127.5
# Convert HWC format to CHW
# Note: hiai.NNTensor() only takes NCHW data as input.
# [https://ascend.huawei.com/doc/Atlas200DK/1.3.0.0/zh/zh-cn_topic_0161025273.html]
inputImage = inputImage.transpose([2, 0, 1]).copy()
inputMask = inputMask.transpose([2, 0, 1]).copy()
inputImageTensor = hiai.NNTensor(inputImage, inputWidth, inputHeight, 3,
'input_image', DataType.FLOAT32_T,
inputWidth * inputHeight * 3)
inputMaskTensor = hiai.NNTensor(inputMask, inputWidth, inputHeight, 1,
'input_mask', DataType.FLOAT32_T,
inputWidth * inputHeight * 1)
nntensorList = hiai.NNTensorList([inputImageTensor, inputMaskTensor])
print("Inference start...")
startTime = time.time()
resultList = GraphInference(myGraph, nntensorList)
if resultList is None:
print("[ERROR] Inference failed.")
DestroyGraph()
exit(0)
endTime = time.time()
inferenceTime = endTime - startTime
SaveResult(resultList, 'demo.jpg')
DestroyGraph()
print("Inference finished. Inference time: %.3fms" %
(inferenceTime * 1000))
def main():
inferenceModel = hiai.AIModelDescription('faster-rcnn', omFileName)
print omFileName
print inferenceModel
# we will resize the jpeg to 896*608 to meet faster-rcnn requirement via opencv,
# so DVPP resizing is not needed
myGraph = CreateGraphWithoutDVPP(inferenceModel)
if myGraph is None:
print "CreateGraph failed"
return None
# in this sample demo, the faster-rcnn model requires 896*608 images
dvppInWidth = 896
dvppInHeight = 608
start = time.time()
index = 0
pathDir = os.listdir(srcFileDir)
for allDir in pathDir:
child = os.path.join('%s%s' % (srcFileDir, allDir))
if (not jpegHandler.is_img(child)):
print '[info] file : ' + child + ' is not image !'
continue
# read the jpeg file and resize it to required w&h, than change it to YUV format.
input_image = jpegHandler.jpeg2yuv(child, dvppInWidth, dvppInHeight)
inputImageTensor = hiai.NNTensor(input_image, dvppInWidth,
dvppInHeight, 3, 'testImage',
DataType.UINT8_T,
dvppInWidth * dvppInHeight * 3 / 2)
imageinfo = np.array([896, 608, 3]).astype(np.float32)
imageinfo = np.reshape(imageinfo, (1, 3))
infoTensor = hiai.NNTensor(imageinfo, 1, 3, 1, 'testinfo',
DataType.FLOAT32_T, imageinfo.size)
datalist = [inputImageTensor, infoTensor]
nntensorList = hiai.NNTensorList(datalist)
if not nntensorList:
print "nntensorList is null"
break
resultList = GraphInference(myGraph, nntensorList)
if resultList is None:
print "graph inference failed"
continue
print resultList[1].shape
PostProcess(resultList, dstFileDir, child)
end = time.time()
print 'cost time ' + str((end - start) * 1000) + 'ms'
hiai.hiai._global_default_graph_stack.get_default_graph().destroy()
print '-------------------end'
def Inference(self,input_image):
inputImageTensor = hiai.NNTensor(input_image, self.width, self.height, 3, 'testImage', DataType.UINT8_T,self.width * self.height * 3)
imageinfo=np.array([800,600,3]).astype(np.float32)
imageinfo=np.reshape(imageinfo,(1,3))
infoTensor = hiai.NNTensor(imageinfo, 1,3,1,'testinfo',DataType.FLOAT32_T, imageinfo.size)
#print('inputImageTensor is :', inputImageTensor)
datalist=[inputImageTensor, infoTensor]
nntensorList = hiai.NNTensorList(datalist)
#print('nntensorList', nntensorList)
resultList = self.model.Inference(self.graph, nntensorList)
#print('inference over')
return resultList
def make_input_tensor(wav_features):
input_tensor_height = wav_features.shape[2]
input_tensor_width = wav_features.shape[3]
#wav_features = np.reshape(wav_features, (0,3,1,2))
#print 555
#print wav_features.shape[0]
#print wav_features.shape[1]
#print wav_features.shape[2]
#print wav_features.shape[3]
#print wav_features.shape
input_tensor = hiai.NNTensor(wav_features, input_tensor_height, input_tensor_width, 1, "wav_features", DataType.FLOAT32_T, wav_features.size)
width0 = input_tensor.width
height0 = input_tensor.height
channel0 = input_tensor.channel
#print str(height0)
#print str(width0)
#print str(channel0)
nntensorList = hiai.NNTensorList(input_tensor)
return nntensorList
#return wav_features
def Inference(self, input_image):
inputImageTensor = hiai.NNTensor(input_image, self.height, self.width,
3, 'testImage', DataType.UINT8_T,
self.height * self.width * 3)
nntensorList = hiai.NNTensorList(inputImageTensor)
resultList = self.model.Inference(self.graph, nntensorList)
if not resultList:
print("Inference fail")
return None
#返回推理结果
return resultList
def Inference(self, img, mask):
input_tensor_list = []
sizes = [3 * 512 * 512, 512 * 512]
input_tensor_list.append(
hiai.NNTensor(img,
height=512,
width=512,
channel=3,
name='image',
data_type=DataType.FLOAT32_T,
size=sizes[0]))
input_tensor_list.append(
hiai.NNTensor(mask,
height=512,
width=512,
channel=1,
name='mask',
data_type=DataType.FLOAT32_T,
size=sizes[1]))
nntensorList = hiai.NNTensorList(
[input_tensor_list[0], input_tensor_list[1]])
if not nntensorList:
print("nntensorList is null")
resultList = self.model.Inference(self.graph, nntensorList)
inpainted_512 = resultList[0]
attention = resultList[1]
mask_512_new = resultList[2]
# output in memory is NCHW, but in numpy it is displayed as NHWC, need to make numpy consistent with memory
inpainted_512.shape = (1, 3, 512, 512)
attention.shape = (1, 1024, 32, 32)
mask_512_new.shape = (1, 1, 512, 512)
# transpose to NHWC
inpainted_512 = inpainted_512.transpose(0, 2, 3, 1)
attention = attention.transpose(0, 2, 3, 1)
mask_512_new = mask_512_new.transpose(0, 2, 3, 1)
return inpainted_512, attention, mask_512_new
def CreateNntensorList(self, input_data):
'''
Create NNTensorList instance with input_data
Args:
input_data: a numpy array, the data user wants to process
Returns:
nntensorList: NNTensorList instance that can be used by graph.proc
'''
inputImageTensor = hiai.NNTensor(input_data)
nntensorList = hiai.NNTensorList(inputImageTensor)
return nntensorList
def Inference(self, input_image):
if isinstance(input_image, np.ndarray) is None:
return False
strattime = time.time()
# Image PreProcess
resized_image = cv2.resize(input_image, (self.width, self.height))
print('[0] resize cost: ' + str((time.time() - strattime) * 1000) +
'ms')
strattime = time.time()
inputImageTensor = hiai.NNTensor(resized_image)
print('[1] input image cost: ' +
str((time.time() - strattime) * 1000) + 'ms')
strattime = time.time()
nntensorList = hiai.NNTensorList(inputImageTensor)
print('[2] get list cost: ' + str((time.time() - strattime) * 1000) +
'ms')
strattime = time.time()
# 调用推理接口
resultList = self.model.Inference(self.graph, nntensorList)
print('[3] get result list cost: ' +
str((time.time() - strattime) * 1000) + 'ms')
strattime = time.time()
if resultList is not None:
bboxes = utils.get_result(resultList, self.width,
self.height) # 获取检测结果
# print("bboxes:", bboxes)
print('[4] get box cost: ' +
str((time.time() - strattime) * 1000) + 'ms')
strattime = time.time()
# Yolov_resnet18 Inference
output_image = utils.draw_boxes(resized_image, bboxes) # 在图像上画框
print('[5] draw box cost: ' +
str((time.time() - strattime) * 1000) + 'ms')
strattime = time.time()
output_image = cv2.resize(
output_image, (input_image.shape[1], input_image.shape[0]))
print('[6] resize cost: ' + str((time.time() - strattime) * 1000) +
'ms')
strattime = time.time()
img_name = datetime.datetime.now().strftime("%Y-%m-%d%H-%M-%S-%f")
cv2.imwrite('output_image/' + str(img_name) + '.jpg', output_image)
print('[7] write cost: ' + str((time.time() - strattime) * 1000) +
'ms')
strattime = time.time()
else:
print('no person in this frame.')
return False
return True
def Inference(self, nparryList, boxList):
if not (isinstance(nparryList, list) and isinstance(boxList, list)):
return False
# 将上一帧图片包含每一个人脸的头部姿势描述清空
del (self.resultList[:])
# 将上一帧图片包含每一个人脸的68个关键点清空
del (self.facepointList[:])
for i in range(len(nparryList)):
if isinstance(nparryList[i], np.ndarray) is None:
print("please check your input format.")
return False
else:
box_width = boxList[i][2] - boxList[i][0]
box_height = boxList[i][3] - boxList[i][1]
resized_image = cv2.resize(nparryList[i],
(self.width, self.height))
inputImageTensor = hiai.NNTensor(resized_image, self.height,
self.width, 3, 'testImage',
DataType.UINT8_T,
self.height * self.width * 3)
nntensorList = hiai.NNTensorList(inputImageTensor)
# 调用推理接口
resultList = self.model.Inference(self.graph, nntensorList)
if resultList is not None:
# pitch yaw roll
# *50
# 每次循环完后都要保存结果到self.resultList
self.resultList.append([
resultList[1][0][0][0][0] * 50,
resultList[1][0][0][0][1] * 50,
resultList[1][0][0][0][2] * 50
])
HeadPosePoint = []
for j in range(136):
if j % 2 == 0:
HeadPosePoint.append(
(1 + resultList[0][0][0][0][j]) / 2 *
box_width + boxList[i][0])
else:
HeadPosePoint.append(
(1 + resultList[0][0][0][0][j]) / 2 *
box_height + boxList[i][1])
self.facepointList.append(HeadPosePoint)
else:
print('not inference head pose in this frame.')
return False
# 判断此时的头部姿势
HeadposeInference.head_status_get(self)
return True
def Inference_mult(self, *input_data):
'''
Inferece interface, process data with configured model
Args:
input_data: a numpy array, the data user wants to process
Returns:
a list, inference result
'''
result_list = []
for data in input_data:
inputImageTensor = hiai.NNTensor(data)
inputNntensorList = hiai.NNTensorList(inputImageTensor)
result = self.graph.proc(inputNntensorList)
result_list.append(result)
return result_list
def main():
inferenceModel = hiai.AIModelDescription('restnet18', resnet18OmFileName)
# we will resize the jpeg to 256*224 to meet resnet18 requirement via opencv,
# so DVPP resizing is not needed
myGraph = CreateGraphWithoutDVPP(inferenceModel)
if myGraph is None:
print "CreateGraph failed"
return None
# in this sample demo, the resnet18 model requires 256*224 images
dvppInWidth = 42
dvppInHeight = 42
start = time.time()
jpegHandler.mkdirown(dstFileDir)
pathDir = os.listdir(srcFileDir)
for allDir in pathDir:
child = os.path.join('%s%s' % (srcFileDir, allDir))
if (not jpegHandler.is_img(child)):
print '[info] file : ' + child + ' is not image !'
continue
# read the jpeg file and resize it to required w&h, than change it to YUV format.
input_image = jpegHandler.jpeg2yuv(child, dvppInWidth, dvppInHeight)
inputImageTensor = hiai.NNTensor(input_image, dvppInWidth,
dvppInHeight, 3, 'testImage',
DataType.UINT8_T,
dvppInWidth * dvppInHeight)
nntensorList = hiai.NNTensorList(inputImageTensor)
resultList = GraphInference(myGraph, nntensorList)
if resultList is None:
print "graph inference failed"
continue
Resnet18PostProcess(resultList, srcFileDir, dstFileDir, allDir)
end = time.time()
print 'cost time ' + str((end - start) * 1000) + 'ms'
hiai.hiai._global_default_graph_stack.get_default_graph().destroy()
print '-------------------end'
def ExcuteInference(self, images):
result = []
for i in range(0, len(images)):
nArray = Yuv2Array(images[i])
ssd = {"name": "face_detection", "path": self.modelPath}
nntensor = hiai.NNTensor(nArray)
tensorList = hiai.NNTensorList(nntensor)
if self.first == True:
self.graph = hiai.Graph(hiai.GraphConfig(graph_id=2001))
with self.graph.as_default():
self.engine_config = hiai.EngineConfig(
engine_name="HIAIDvppInferenceEngine",
side=hiai.HiaiPythonSide.Device,
internal_so_name='/lib/libhiai_python_device2.7.so',
engine_id=2001)
self.engine = hiai.Engine(self.engine_config)
self.ai_model_desc = hiai.AIModelDescription(
name=ssd['name'], path=ssd['path'])
self.ai_config = hiai.AIConfig(
hiai.AIConfigItem("Inference", "item_value_2"))
final_result = self.engine.inference(
input_tensor_list=tensorList,
ai_model=self.ai_model_desc,
ai_config=self.ai_config)
ret = copy.deepcopy(self.graph.create_graph())
if ret != hiai.HiaiPythonStatust.HIAI_PYTHON_OK:
print("create graph failed, ret", ret)
d_ret = graph.destroy()
SetExitFlag(True)
return HIAI_APP_ERROR, None
self.first = False
else:
with self.graph.as_default():
final_result = self.engine.inference(
input_tensor_list=tensorList,
ai_model=self.ai_model_desc,
ai_config=self.ai_config)
resTensorList = self.graph.proc(input_nntensorlist=tensorList)
print("Inference result: ", resTensorList[0].shape)
result.append(resTensorList)
return HIAI_APP_OK, result
def Inference(self,input_image):
if isinstance(input_image,np.ndarray) is None:
return False
h,w,c = input_image.shape
resized_image = cv2.resize(input_image,(self.width,self.height))
inputImageTensor = hiai.NNTensor(resized_image)
nntensorList = hiai.NNTensorList(inputImageTensor)
# 将上一帧图片推理出的人脸数据清空
del(self.boxList[:])
del(self.nparray[:])
# 调用推理接口
resultList = self.model.Inference(self.graph, nntensorList)
# 将推理结果的数据保存到临时列表
# 再从临时列表中去筛选出有效的人脸坐标,保存到self.resultList中
# 从原始图片中crop出的人脸(数据类型为numpy数组),保存到一个列表中self.nparray
if resultList is not None:
for i in range(200):
if (resultList[0][i][0][0][2] > 0.8 and resultList[0][i][0][0][2] <= 1.0):
if (resultList[0][i][0][0][3] < 0):
resultList[0][i][0][0][3] = 0
if (resultList[0][i][0][0][4] < 0):
resultList[0][i][0][0][4] = 0
rect_width = resultList[0][i][0][0][5] - resultList[0][i][0][0][3]
rect_height = resultList[0][i][0][0][6] - resultList[0][i][0][0][4]
if (resultList[0][i][0][0][3] != 0 and abs(self.M_left * rect_width) <= resultList[0][i][0][0][3]):
resultList[0][i][0][0][3] = resultList[0][i][0][0][3] + self.M_left * rect_width
if (resultList[0][i][0][0][4] != 0 and abs(self.M_top * rect_height) <= resultList[0][i][0][0][4]):
resultList[0][i][0][0][4] = resultList[0][i][0][0][4] + self.M_top * rect_height
resultList[0][i][0][0][5] = resultList[0][i][0][0][5] + self.M_right * rect_width
resultList[0][i][0][0][6] = resultList[0][i][0][0][6] + self.M_bottom * rect_height
# 人脸框的坐标
self.boxList.append([resultList[0][i][0][0][3]*w, resultList[0][i][0][0][4]*h,
resultList[0][i][0][0][5]*w, resultList[0][i][0][0][6]*h])
self.nparray.append(input_image[int(self.boxList[i][1]):int(self.boxList[i][3]),int(self.boxList[i][0]):int(self.boxList[i][2])])
else:
break
else:
print ('not inference face in this frame.')
return False
return True
def main():
inferenceModel = hiai.AIModelDescription('segmentationDes', OmFileName)
myGraph = CreateGraph(inferenceModel)
if myGraph is None:
print("CreateGraph failed")
return None
start = time.time()
jpegHandler.mkdirown(dstFileDir)
pathDir = os.listdir(srcFileDir)
label_colours = cv.imread(colours, 1).astype(np.uint8)
#print label_colours
label_colours_bgr = label_colours[..., ::-1]
for allDir in pathDir:
child = os.path.join('%s%s' % (srcFileDir, allDir))
if (not jpegHandler.is_img(child)):
print('[info] file : ' + child + ' is not image !')
continue
print('[info] file : ' + child + ' begin process !')
# read the jpeg file and resize it to required w&h, than change it to YUV format.
input_image = jpegHandler.jpeg2yuv(child, InWidth, InHeight)
inputImageTensor = hiai.NNTensor(input_image, InWidth, InHeight, 3,
'testImage', DataType.UINT8_T,
InWidth * InHeight * 3 / 2)
nntensorList = hiai.NNTensorList(inputImageTensor)
resultList = GraphInference(myGraph, nntensorList)
if resultList is None:
print(child + "graph inference failed")
continue
resultArray = resultList[0]
resultArray = resultArray.reshape(19, InWidth, InHeight)
prediction = resultArray.argmax(axis=0)
#print prediction
prediction = np.squeeze(prediction)
prediction = np.resize(prediction, (3, InHeight, InWidth))
prediction = prediction.transpose(1, 2, 0).astype(np.uint8)
prediction_rgb = np.zeros(prediction.shape, dtype=np.uint8)
cv.LUT(prediction, label_colours_bgr, prediction_rgb)
input_path_ext = child.split(".")[-1]
input_image_name = child.split("/")[-1:][0].replace(
'.' + input_path_ext, '')
out_path_im = dstFileDir + input_image_name + '_erfnet' + '.' + input_path_ext
cv.imwrite(out_path_im,
prediction_rgb) # color images for visualization
print(input_image_name + ' process end ')
end = time.time()
print('cost time ' + str((end - start) * 1000) + 'ms')
hiai.hiai._global_default_graph_stack.get_default_graph().destroy()
print('-------------------end')
def ExcuteInference(self, images):
result = []
for i in range(0, len(images)):
nArray = Yuv2Array(images[i])
ssd = {"name": "object_detection", "path": self.modelPath}
nntensor = hiai.NNTensor(nArray)
'''
gray = cv2.cvtColor(nArray, cv2.COLOR_YUV2GRAY_420)
blurred = cv2.GaussianBlur(gray, (5, 5), 0)
thresh = cv2.threshold(blurred, 60, 255, cv2.THRESH_BINARY)[1]
cnts = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)
cnts=cnts[0]
#print("1111111")
j=0
for c in cnts:
M = cv2.moments(c)
c = c.astype("float")
c = c.astype("int")
shape = detect(c)
# cv2.drawContours(cv_image, [c], -1, (0, 255, 0), 2)
# image=cv2.cvtColor(cv_image, cv2.COLOR_YUV2RGB_I420)
# picpath='~/sample-objectdetection-python/picpath'+str(i)+'.jpg'
# cv2.imwrite(picpath,image)
# print("111111")
#BGR=cv2.cvtColor(nArray,cv2.COLOR_YUV2RGB)
#return BGR
# print(cv123_image)
# img_np = YUVtoRGB(nArray)
# cv2.imwrite("text4.jpg",img_np)
'''
tensorList = hiai.NNTensorList(nntensor)
if self.first == True:
self.graph = hiai.Graph(hiai.GraphConfig(graph_id=2001))
with self.graph.as_default():
self.engine_config = hiai.EngineConfig(
engine_name="HIAIDvppInferenceEngine",
side=hiai.HiaiPythonSide.Device,
internal_so_name='/lib/libhiai_python_device2.7.so',
engine_id=2001)
self.engine = hiai.Engine(self.engine_config)
self.ai_model_desc = hiai.AIModelDescription(
name=ssd['name'], path=ssd['path'])
self.ai_config = hiai.AIConfig(
hiai.AIConfigItem("Inference", "item_value_2"))
final_result = self.engine.inference(
input_tensor_list=tensorList,
ai_model=self.ai_model_desc,
ai_config=self.ai_config)
ret = copy.deepcopy(self.graph.create_graph())
if ret != hiai.HiaiPythonStatust.HIAI_PYTHON_OK:
print("create graph failed, ret", ret)
d_ret = graph.destroy()
SetExitFlag(True)
return HIAI_APP_ERROR, None
self.first = False
else:
with self.graph.as_default():
final_result = self.engine.inference(
input_tensor_list=tensorList,
ai_model=self.ai_model_desc,
ai_config=self.ai_config)
resTensorList = self.graph.proc(input_nntensorlist=tensorList)
# print("Inference result: ", resTensorList[0].shape)
result.append(resTensorList)
return HIAI_APP_OK, result
