def infer_with_predictor(input):
"""
infer a image with predictor, for example
:param input: image base64
:return: pred out
"""
start_time = time.time()
period = time.time() - start_time
logger.info("warp_input cost time: {}".format("%2.2f sec" % period))
predictor = predictor_manager.get_predictor()
if predictor is not None:
try:
start_time = time.time()
outputs = predictor.run(input)
period = time.time() - start_time
logger.info("predictor infer cost time: {}".format("%2.2f sec" %
period))
if len(outputs) > 0:
output = outputs[0]
print(output.name)
output_data = np.array(output.data.float_data())
output_data = output_data.reshape(-1, 6)
# logger.debug("predictor infer result {}".format(output_data))
return output_data.tolist()
return "infer nothing"
finally:
predictor_manager.return_predictor(predictor)
else:
return None
def set_task_queue(self, data, timeout=0.5):
"""
用于生产者向队列中提交任务，并返回对应的索引位置
:param data: 需要提交的数据
:param timeout: 最长等待时间
:return: 生产者到时候需要在结果数组中获取结果的索引
"""
index = None
# 这里如果查询放在锁外面有概率造成，请求提前结束等待，然后返回请求超时,所以查询的时候也需要加锁，
# 如果目前队列满，那么先解锁，然后等待，然后加锁再次尝试添加数据，如果还是不行，就超时
self.q_lock.acquire()
if not self.q_full.is_set():
self.q_lock.release()
self.q_full.wait(timeout)
self.q_lock.acquire()
if self.q_full.is_set():
index = self.task_queue_index
self.task_queue[index] = data
self.task_queue_index += 1
# 如果消费者还没被唤醒，那就唤醒他
if not self.c_lock.is_set():
self.c_lock.set()
# 如果任务队列满了，夯住自己
if self.task_queue_index == self.task_num:
self.q_full.clear()
else:
logger.info("set:task_queue is full, try a moment later")
self.q_lock.release()
return index
def read_image(image_bytes, target_size):
"""
read image
:param image_base64: input image, base64 encode string
:param target_size: image resize target
:return: origin image and resized image
"""
start_time = time.time()
# image_bytes = base64.b64decode(image_base64)
img_array = np.fromstring(image_bytes, np.uint8)
img = cv2.imdecode(img_array, cv2.COLOR_RGB2BGR)
origin = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
img = resize_img(origin, target_size)
if img.mode != 'RGB':
img = img.convert('RGB')
# img = np.array(img).astype(np.float32).transpose((2, 0, 1)) # HWC to CHW
# img -= 127.5
# img *= 0.007843
mean_rgb = [127.5, 127.5, 127.5]
img = crop_image(img, target_size)
img = np.array(img).astype('float32')
img -= mean_rgb
img = img.transpose((2, 0, 1)) # HWC to CHW
img *= 0.007843
img = img[np.newaxis,:]
period = time.time() - start_time
logger.info("read image base64 and resize cost time: {}".format("%2.2f sec" % period))
return origin, img
def __init__(self):
"""
create predictor manager
"""
self.get_predictor_timeout = float(
config.get('get.predictor.timeout', default_value=0.5))
predictor_count = 0
enable_mkl = False
gpu_memory = 200
gpu_device_ids = []
model_dir = config.get('model.dir')
device_type = config.get('device.type')
if device_type == PredictorManager.CPU_DEVICE:
cpu_predictor_count = int(
config.getint('cpu.predictor.count', default_value=0))
predictor_count = cpu_predictor_count
enable_mkl = config.getboolean('cpu.enable_mkl',
default_value=False)
elif device_type == PredictorManager.GPU_DEVICE:
gpu_predictor_count = int(
config.getint('gpu.predictor.count', default_value=0))
predictor_count = gpu_predictor_count
gpu_memory = config.getint('gpu.predictor.memory',
default_value=200)
gpu_device_ids = config.get('gpu.predictor.device.id').split(',')
gpu_device_ids = map(int, gpu_device_ids)
if PYTHON_VERSION == 3:
gpu_device_ids = list(gpu_device_ids)
assert len(
gpu_device_ids
) == gpu_predictor_count, "gpu predictor count doesn't match device count"
else:
raise Exception("no device to run predictor!")
assert predictor_count > 0, "no device to predict"
logger.info(
"device type:{} predictor count:{} model dir:{} get predictor timeout:{}s"
.format(device_type, predictor_count, model_dir,
self.get_predictor_timeout))
self.predictor_queue = Queue(maxsize=predictor_count)
for i in range(predictor_count):
# Set config
predictor_config = AnalysisConfig(model_dir)
# predictor_config.specify_input_name()
if device_type == PredictorManager.CPU_DEVICE:
predictor_config.disable_gpu()
if enable_mkl:
predictor_config.enable_mkldnn()
else:
device_id = gpu_device_ids[i]
predictor_config.enable_use_gpu(gpu_memory, device_id)
# Create PaddlePredictor
predictor = create_paddle_predictor(predictor_config)
self.predictor_queue.put(predictor)
def init_evn(cls, current_evn):
"""
init config environment, and only once
:param current_evn:
:return:
"""
if cls.env is None:
cls.env = current_evn
logger.info("current config section %s",
cls.env,
extra={'logid': 'react-main-loop-logid'})
def recognize():
"""
处理post请求,并返回处理后的结果
"""
# print(request.headers)
# print(request.form)
file = request.files['imgfile']
# print(file, file.filename)
# with open('receiveimg.jpg', 'wb') as f:
# f.write(file.stream.read())
if not request.form:
abort(400)
if 'log_id' in request.form:
log_id = request.form['log_id']
logger.set_logid(str(log_id))
else:
logger.set_auto_logid()
log_id = logger.get_logid()
log_id = int(log_id)
result = ApiResult(log_id=log_id)
start_time = time.time()
try:
image_bytes = file.stream.read()
task_data = process_data(image_bytes)
# print(task_data.shape)
start_time = time.time()
max_request_time = float(config.get('max_request_time'))
index = pivot.set_task_queue(task_data, max_request_time)
if index is not None:
pred = pivot.get_result_list(index)
pred['log_id'] = log_id
result.success(data=pred)
# 历史记录入库
label, score, user_id = result.top3[0][0], result.top3[0][1], 1
sql = "insert into record(img,label,score,user_id) values (%s,{},{},{});".format(
label, score, user_id)
print(result.top3[0][0], result.top3[0][1])
cursor.execute(sql, args=(image_bytes))
database.commit()
else:
result.error(message="busy, wait then retry")
except Exception as err:
logger.error("infer exception: {}".format(err))
result.error(message=str(err))
period = time.time() - start_time
logger.info("request cost:{}".format("%2.2f sec" % period))
return json.dumps(result, default=lambda o: o.__dict__)
def process_data(image_bytes):
"""
根据infer_type来处理数据，并返回, 可根据业务需要自行修改
:param inputs: 原始的输入数据
:return:
"""
# 两种方式 处理数据公共的部分
input_size = config.get('input.size').split(',')
input_size = map(int, input_size)
if PYTHON_VERSION == 3:
input_size = list(input_size)
start_time = time.time()
origin, image_data = data_util.read_image(image_bytes, input_size[1:])
# 如果是predictor，需要进一步的把数据处理为PaddleTensor
if PREDICTOR == infer_type:
image_data = data_util.warp_input(image_data, input_size)
period = time.time() - start_time
logger.info("prepare input cost time: {}".format("%2.2f sec" % period))
return image_data
def business_process(inputs):
"""
不同业务，数据处理的方式可能不一样，统一在这里修改和控制
为了充分利用多线程并行的方式来处理数据，所以首先由生产者调用process_data来处理数据，并提交至任务队列
此处直接从任务队列中获取处理好的数据即可，因此该部分应该和process_data一起修改
:param inputs: 从任务队列中获取的预处理后的数据
:return:
"""
if PREDICTOR == infer_type:
outputs = infer_with_predictor(inputs)
return outputs
elif EXECUTOR == infer_type:
# 重组数据
image_data = [data[0] for data in inputs]
image_data = np.array(image_data)
start_time = time.time()
# 通过不同的model名称来调用模型级联任务中 不同阶段需要用到的模型
outputs = infer_with_executor("SE_ResNeXt", image_data)
period = time.time() - start_time
infer_num = image_data.shape[0]
logger.info("executor infer num {} cost time: {}".format(
str(infer_num), "%2.2f sec" % period))
outputs = np.array(outputs[0])
result = []
for lod_id in range(infer_num):
data = {}
output_data = outputs[lod_id]
# logger.debug("executor infer result {}".format(output_data))
data["result"] = np.array(output_data).tolist()
# 首先先根据每次infer的output的lod信息，把output拆成一个图一个图的result
result.append(data)
return result
else:
logger.critical(
"must set a infer type in config, executor or predictor")
exit(1)
def __init__(self):
"""
create executor manager
"""
self.get_executor_timeout = float(
config.get('get.executor.timeout', default_value=0.5))
model_dir = config.get('model.dir')
# 将地址的json字符串转为字典，key为模型的名称，value为模型的地址，放便级联使用
model_dir = json.loads(model_dir)
executor_count = 0
enable_mkl = False
gpu_memory = 200
gpu_device_ids = []
self.places_list = []
# 模型的字典
self.model_dict = {}
for model_name in model_dir.keys():
self._load_model(model_dir[model_name], model_name)
device_type = config.get('device.type')
if device_type == ExecutorManager.CPU_DEVICE:
cpu_executor_count = int(
config.getint('cpu.executor.count', default_value=0))
executor_count = cpu_executor_count
elif device_type == ExecutorManager.GPU_DEVICE:
gpu_executor_count = int(
config.getint('gpu.executor.count', default_value=0))
executor_count = gpu_executor_count
gpu_device_ids = config.get('gpu.executor.device.id').split(',')
gpu_device_ids = map(int, gpu_device_ids)
if PYTHON_VERSION == 3:
gpu_device_ids = list(gpu_device_ids)
assert len(
gpu_device_ids
) == gpu_executor_count, "gpu executor count doesn't match device count"
else:
raise Exception("no device to run executor!")
assert executor_count > 0, "no device to predict"
logger.info(
"device type:{} executor count:{} model dir:{} get executor timeout:{}s"
.format(device_type, executor_count, model_dir,
self.get_executor_timeout))
self.executor_queue = Queue(maxsize=executor_count)
for i in range(executor_count):
# executor is thread safe,
# supports single/multiple-GPU running,
# and single/multiple-CPU running,
# if CPU device, only create one for all thread
if device_type == ExecutorManager.CPU_DEVICE:
if self.executor_queue.empty():
place = fluid.CPUPlace()
executor = fluid.Executor(place)
self._temp_executor = executor
else:
executor = self._temp_executor
else:
device_id = gpu_device_ids[i]
place = fluid.CUDAPlace(device_id)
executor = fluid.Executor(place)
self.executor_queue.put(executor)
def state():
"""
访问state
"""
logger.info("visit state page")
return '0'
def index():
"""
访问root
"""
logger.info("visit root page")
return 'Index Page'
from twisted.web.resource import Resource
from twisted.web.server import Site
from twisted.web.wsgi import WSGIResource
from utils.consumer import InferConsumer
from business_service import process_data
pivot = Pivot(int(config.get("max_task_num")))
infer_consumer = InferConsumer(pivot,
float(config.get("max_get_time")))
port = int(config.get("flask.server.port"))
work_thread_count = int(config.get('work.thread.count'))
site_root = str(config.get('server.root.site'))
reactor.suggestThreadPoolSize(int(work_thread_count))
flask_site = WSGIResource(reactor, reactor.getThreadPool(), app)
root = Resource()
root.putChild(
site_root if PYTHON_VERSION == 2 else bytes(site_root,
encoding='utf-8'),
flask_site)
logger.info(
"start app listen port:{} thread pool size:{} site root:{}".format(
port, work_thread_count, site_root))
reactor.listenTCP(port, Site(root))
# 把主线程设置为infer_consumr的守护线程，如果主线程结束了，infer_consumer线程就跟着结束
infer_consumer.setDaemon(True)
infer_consumer.start()
reactor.run()