def _load_image(req,
width,
height,
mean_value=[0.0, 0.0, 0.0],
std_value=[1.0, 1.0, 1.0]):
try:
img = load_image(req["uri"], body=req['body'])
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
if img is None:
return None, {"code": 400, "message": "cv2 load image failed"}
# convert into format (batch, RGB, width, height)
img = img.astype(float)
img = cv2.resize(img, (width, height))
img -= mean_value
img /= std_value
img = np.swapaxes(img, 0, 2)
img = np.swapaxes(img, 1, 2)
return img, None
except Exception as _e:
logger.error("_load_image error: %s",
traceback.format_exc(),
extra={"reqid": ""})
if isinstance(_e, ErrorBase):
return None, {"code": _e.code, "message": str(_e)}
return None, {"code": 400, "message": str(_e)}
def net_inference(model, args):
model = model["net"]
mod, labels = model['mod'], model['labels']
width, height = model['image_width'], model['image_height']
batch_size = model['batch_size']
mean_value = [0.0, 0.0, 0.0]
std_value = [1.0, 1.0, 1.0]
if type(model["mean_value"]) is list and len(model["mean_value"]) == 3:
mean_value = model["mean_value"]
if type(model["std_value"]) is list and len(model["std_value"]) == 3:
std_value = model["std_value"]
if len(args) > batch_size:
for i in range(cur_batchsize):
raise ErrorOutOfBatchSize(batch_size)
Batch = namedtuple('Batch', ['data'])
rets = range(len(args))
valid_imgs = []
for i, data in enumerate(args):
im, err = _load_image(data['data'], width, height, mean_value, std_value)
if err:
logger.error("_load_image error : %s", data['data']['uri'], extra={"reqid": ""})
rets[i] = err
continue
limit = 1
if "params" in data and "limit" in data["params"]:
if (type(data["params"]["limit"]) is int or data["params"]["limit"].isdigit()) and int(
data["params"]["limit"]) <= len(labels):
limit = int(data["params"]["limit"])
else:
rets[i] = {"code": 400, "message": "invalid limit params"}
continue
valid_imgs.append(dict(index=i, img=im, limits=limit))
if len(valid_imgs) == 0:
return rets, 400, None
try:
img_batch = mx.nd.array(np.zeros((batch_size, 3, width, height)))
for index, image in enumerate(valid_imgs):
img_batch[index] = mx.nd.array(image["img"])
mod.forward(Batch([img_batch]))
output_batch = mod.get_outputs()[0].asnumpy()
## post process
for i in xrange(len(valid_imgs)):
rets[valid_imgs[i]["index"]] = dict(
code=0,
message='',
result=_build_result(output_batch[i], labels, valid_imgs[i]["limits"])
)
return rets, 0, None
except Exception as _e:
logger.error("_load_image error: %s", traceback.format_exc(), extra={"reqid": ""})
return None, 400, {"code": 400, "message": str(_e)}
def create_net(configs):
'''
net init
'''
logger.info("[Python net_init] load configs: %s",
configs,
extra={"reqid": ""})
file_features = configs.get('custom_files', {}).get('features.line', None)
file_labels = configs.get('custom_files', {}).get('labels.line', None)
if file_features is None or file_labels is None:
logger.error("need file_features/file_labels", extra={"reqid": ""})
return {}, 400, "miss some custom file"
custom_values = configs.get('custom_params', {})
with open(file_features, 'r') as _f:
# {'index': 1, 'url': '', 'pts':[], 'feature':[]}
features = [json.loads(line) for line in _f.readlines()]
large_features = [
line for line in features if line.get("size", "") == "large"
]
small_features = [
line for line in features if line.get("size", "") == "small"
]
xsmall_features = [
line for line in features if line.get("size", "") == "XSmall"
]
with open(file_labels, 'r') as _f2:
labels = [line.strip() for line in _f2.readlines()]
thresholds = custom_values.get(
'threshold',
[[0.38, 0.4, 0.42], [0.38, 0.4, 0.42], [0.35, 0.375, 0.4]])
size_limits = custom_values.get("size_limit", [24, 32, 60])
if len(labels) == 0:
return {}, 400, "invliad numbers of labels "
if len(large_features) == 0 or len(small_features) == 0 or len(
xsmall_features) == 0:
return {}, 400, "invliad numbers of base datasets "
if len(thresholds) != 3 or len(thresholds[0]) != 3 or len(
thresholds[1]) != 3 or len(thresholds[2]) != 3:
return {}, 400, "invliad numbers of thresholds, should be array[3][3]"
if len(size_limits) != 3:
return {}, 400, "invliad numbers of subset size_limit , should be array[3]"
cu_large_features = numpy.array(
[feature.get('feature', []) for feature in large_features])
cu_small_features = numpy.array(
[feature.get('feature', []) for feature in small_features])
cu_xsmall_features = numpy.array(
[feature.get('feature', []) for feature in xsmall_features])
return {
'large_features': large_features,
'small_features': small_features,
'xsmall_features': xsmall_features,
'labels': labels,
'cm_large_features': cu_large_features,
'cm_small_features': cu_small_features,
'cm_xsmall_features': cu_xsmall_features,
'thresholds': thresholds,
'sizelimits': size_limits
}, 0, ''
def net_inference(model, args):
'''
net inference
'''
large_features = model.get('large_features', [])
small_features = model.get('small_features', [])
xsmall_features = model.get('xsmall_features', [])
cm_large_features = model.get('cm_large_features', None)
cm_small_features = model.get('cm_small_features', None)
cm_xsmall_features = model.get('cm_xsmall_features', None)
labels = model.get('labels', [])
thresholds = model.get('thresholds', 0.0)
size_limits = model.get('sizelimits')
buf = None
limit = 1
req = args[0]
pts = req.get('data', {}).get('attribute', {}).get('pts', [])
logger.info("[Python inference]: limit:%s,threshold:%s",
limit,
thresholds,
extra={"reqid": ""})
features = large_features
cm_features = cm_large_features
threshold = thresholds[2]
if len(pts) != 0:
srt, err = short_side(pts)
logger.info("[Python inference]: shortest side %d, err: %s",
srt,
err,
extra={"reqid": ""})
if err != "":
return {}, 400, err
if srt < size_limits[0]:
return [{
'code':
0,
'message':
'pts should larger than {}x{}'.format(size_limits[0],
size_limits[0]),
'result': {
"confidences": []
}
}], 0, ''
elif srt < size_limits[1]:
features = xsmall_features
cm_features = cm_xsmall_features
threshold = thresholds[0]
elif srt < size_limits[2]:
features = small_features
cm_features = cm_small_features
threshold = thresholds[1]
#阈值映射
score_map = lambda b, m, u, f, s: b + m * (s - u) / f
try:
if os.path.exists(req.get('data', {}).get('uri', '')):
_f = open(req.get('data', {}).get('uri', ''), 'rb')
buf = _f.read()
_f.close()
elif req.get('data', {}).get('body', '') is not None:
buf = req.get('data', {}).get('body', '')
else:
return {}, 400, "failed to get feature data"
feature = struct.unpack(">" + str(len(buf) / 4) + 'f', buf)
except struct.error as _e:
return {}, 400, _e.message
if feature is None:
logger.error('read input failed.', extra={"reqid": ""})
return {}, 500, "read featue failed"
if len(feature) != (len(features[0].get('feature', []))
if len(features) > 0 else 4096):
logger.error("feature len %d %d",
len(feature),
len(features[0].get('feature', [])),
extra={"reqid": ""})
return {}, 400, "bad feature input"
target_feature = numpy.array(feature)
cosins = cm_features.dot(target_feature)
if "params" in req and "limit" in req["params"]:
if (type(req["params"]["limit"]) is int or req["params"]["limit"].isdigit()) and \
int(req["params"]["limit"]) <= len(labels) and int(req["params"]["limit"]) > 0:
limit = int(req["params"]["limit"])
cosins = cosins.squeeze()
indexs = cosins.argsort()[::-1][:limit]
ret = [{'code': 0, 'message': '', 'result': {"confidences": []}}]
for i in xrange(len(indexs)):
score = float(cosins[indexs[i]])
if score < threshold[1]:
logger.info("ret {}".format(ret), extra={"reqid": ""})
return ret, 0, None
#score map，四个区段分别映射
#if score <threshold[0]:
# score=score_map(0,thresholds[2][0],score,threshold[0],threshold[0])
if score > threshold[2]:
score = score_map(thresholds[2][2], 1 - thresholds[2][2],
threshold[2], 1 - threshold[2], score)
else:
#score >= threshold[1] and score <= threshold[2]
score = score_map(thresholds[2][1],
thresholds[2][2] - thresholds[2][1],
threshold[1], threshold[2] - threshold[1], score)
ret[0]["result"]["confidences"].append({
'index':
features[indexs[i]].get('index', 0),
'class':
labels[features[indexs[i]].get('index', 0)],
'group':
features[indexs[i]].get('group', ''),
'score':
score,
'sample': {
'url': features[indexs[i]].get('url', ''),
'pts': features[indexs[i]].get('pts', ''),
'id': features[indexs[i]].get('id', ''),
}
})
return ret, 0, None