提交 b8834933 编写于 作者: L LDOUBLEV

delete benchmark

上级 bc999986
...@@ -48,8 +48,6 @@ class TextClassifier(object): ...@@ -48,8 +48,6 @@ class TextClassifier(object):
self.predictor, self.input_tensor, self.output_tensors, _ = \ self.predictor, self.input_tensor, self.output_tensors, _ = \
utility.create_predictor(args, 'cls', logger) utility.create_predictor(args, 'cls', logger)
self.cls_times = utility.Timer()
def resize_norm_img(self, img): def resize_norm_img(self, img):
imgC, imgH, imgW = self.cls_image_shape imgC, imgH, imgW = self.cls_image_shape
h = img.shape[0] h = img.shape[0]
...@@ -85,35 +83,28 @@ class TextClassifier(object): ...@@ -85,35 +83,28 @@ class TextClassifier(object):
cls_res = [['', 0.0]] * img_num cls_res = [['', 0.0]] * img_num
batch_num = self.cls_batch_num batch_num = self.cls_batch_num
elapse = 0 elapse = 0
self.cls_times.total_time.start()
for beg_img_no in range(0, img_num, batch_num): for beg_img_no in range(0, img_num, batch_num):
end_img_no = min(img_num, beg_img_no + batch_num) end_img_no = min(img_num, beg_img_no + batch_num)
norm_img_batch = [] norm_img_batch = []
max_wh_ratio = 0 max_wh_ratio = 0
starttime = time.time()
for ino in range(beg_img_no, end_img_no): for ino in range(beg_img_no, end_img_no):
h, w = img_list[indices[ino]].shape[0:2] h, w = img_list[indices[ino]].shape[0:2]
wh_ratio = w * 1.0 / h wh_ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, wh_ratio) max_wh_ratio = max(max_wh_ratio, wh_ratio)
self.cls_times.preprocess_time.start()
for ino in range(beg_img_no, end_img_no): for ino in range(beg_img_no, end_img_no):
norm_img = self.resize_norm_img(img_list[indices[ino]]) norm_img = self.resize_norm_img(img_list[indices[ino]])
norm_img = norm_img[np.newaxis, :] norm_img = norm_img[np.newaxis, :]
norm_img_batch.append(norm_img) norm_img_batch.append(norm_img)
norm_img_batch = np.concatenate(norm_img_batch) norm_img_batch = np.concatenate(norm_img_batch)
norm_img_batch = norm_img_batch.copy() norm_img_batch = norm_img_batch.copy()
starttime = time.time()
self.cls_times.preprocess_time.end()
self.cls_times.inference_time.start()
self.input_tensor.copy_from_cpu(norm_img_batch) self.input_tensor.copy_from_cpu(norm_img_batch)
self.predictor.run() self.predictor.run()
prob_out = self.output_tensors[0].copy_to_cpu() prob_out = self.output_tensors[0].copy_to_cpu()
self.cls_times.inference_time.end()
self.cls_times.postprocess_time.start()
self.predictor.try_shrink_memory() self.predictor.try_shrink_memory()
cls_result = self.postprocess_op(prob_out) cls_result = self.postprocess_op(prob_out)
self.cls_times.postprocess_time.end()
elapse += time.time() - starttime elapse += time.time() - starttime
for rno in range(len(cls_result)): for rno in range(len(cls_result)):
label, score = cls_result[rno] label, score = cls_result[rno]
...@@ -121,9 +112,6 @@ class TextClassifier(object): ...@@ -121,9 +112,6 @@ class TextClassifier(object):
if '180' in label and score > self.cls_thresh: if '180' in label and score > self.cls_thresh:
img_list[indices[beg_img_no + rno]] = cv2.rotate( img_list[indices[beg_img_no + rno]] = cv2.rotate(
img_list[indices[beg_img_no + rno]], 1) img_list[indices[beg_img_no + rno]], 1)
self.cls_times.total_time.end()
self.cls_times.img_num += img_num
elapse = self.cls_times.total_time.value()
return img_list, cls_res, elapse return img_list, cls_res, elapse
......
...@@ -66,8 +66,6 @@ class TextRecognizer(object): ...@@ -66,8 +66,6 @@ class TextRecognizer(object):
self.predictor, self.input_tensor, self.output_tensors, self.config = \ self.predictor, self.input_tensor, self.output_tensors, self.config = \
utility.create_predictor(args, 'rec', logger) utility.create_predictor(args, 'rec', logger)
self.rec_times = utility.Timer()
def resize_norm_img(self, img, max_wh_ratio): def resize_norm_img(self, img, max_wh_ratio):
imgC, imgH, imgW = self.rec_image_shape imgC, imgH, imgW = self.rec_image_shape
assert imgC == img.shape[2] assert imgC == img.shape[2]
...@@ -168,14 +166,13 @@ class TextRecognizer(object): ...@@ -168,14 +166,13 @@ class TextRecognizer(object):
width_list.append(img.shape[1] / float(img.shape[0])) width_list.append(img.shape[1] / float(img.shape[0]))
# Sorting can speed up the recognition process # Sorting can speed up the recognition process
indices = np.argsort(np.array(width_list)) indices = np.argsort(np.array(width_list))
self.rec_times.total_time.start()
rec_res = [['', 0.0]] * img_num rec_res = [['', 0.0]] * img_num
batch_num = self.rec_batch_num batch_num = self.rec_batch_num
st = time.time()
for beg_img_no in range(0, img_num, batch_num): for beg_img_no in range(0, img_num, batch_num):
end_img_no = min(img_num, beg_img_no + batch_num) end_img_no = min(img_num, beg_img_no + batch_num)
norm_img_batch = [] norm_img_batch = []
max_wh_ratio = 0 max_wh_ratio = 0
self.rec_times.preprocess_time.start()
for ino in range(beg_img_no, end_img_no): for ino in range(beg_img_no, end_img_no):
h, w = img_list[indices[ino]].shape[0:2] h, w = img_list[indices[ino]].shape[0:2]
wh_ratio = w * 1.0 / h wh_ratio = w * 1.0 / h
...@@ -216,8 +213,6 @@ class TextRecognizer(object): ...@@ -216,8 +213,6 @@ class TextRecognizer(object):
gsrm_slf_attn_bias1_list, gsrm_slf_attn_bias1_list,
gsrm_slf_attn_bias2_list, gsrm_slf_attn_bias2_list,
] ]
self.rec_times.preprocess_time.end()
self.rec_times.inference_time.start()
input_names = self.predictor.get_input_names() input_names = self.predictor.get_input_names()
for i in range(len(input_names)): for i in range(len(input_names)):
input_tensor = self.predictor.get_input_handle(input_names[ input_tensor = self.predictor.get_input_handle(input_names[
...@@ -241,15 +236,13 @@ class TextRecognizer(object): ...@@ -241,15 +236,13 @@ class TextRecognizer(object):
output = output_tensor.copy_to_cpu() output = output_tensor.copy_to_cpu()
outputs.append(output) outputs.append(output)
preds = outputs[0] preds = outputs[0]
self.rec_times.inference_time.end()
self.rec_times.postprocess_time.start()
rec_result = self.postprocess_op(preds) rec_result = self.postprocess_op(preds)
for rno in range(len(rec_result)): for rno in range(len(rec_result)):
rec_res[indices[beg_img_no + rno]] = rec_result[rno] rec_res[indices[beg_img_no + rno]] = rec_result[rno]
self.rec_times.postprocess_time.end() self.rec_times.postprocess_time.end()
self.rec_times.img_num += int(norm_img_batch.shape[0]) self.rec_times.img_num += int(norm_img_batch.shape[0])
self.rec_times.total_time.end()
return rec_res, self.rec_times.total_time.value() return rec_res, time.time() - st
def main(args): def main(args):
...@@ -278,12 +271,6 @@ def main(args): ...@@ -278,12 +271,6 @@ def main(args):
img_list.append(img) img_list.append(img)
try: try:
rec_res, _ = text_recognizer(img_list) rec_res, _ = text_recognizer(img_list)
if args.benchmark:
cm, gm, gu = utility.get_current_memory_mb(0)
cpu_mem += cm
gpu_mem += gm
gpu_util += gu
count += 1
except Exception as E: except Exception as E:
logger.info(traceback.format_exc()) logger.info(traceback.format_exc())
...@@ -292,38 +279,6 @@ def main(args): ...@@ -292,38 +279,6 @@ def main(args):
for ino in range(len(img_list)): for ino in range(len(img_list)):
logger.info("Predicts of {}:{}".format(valid_image_file_list[ino], logger.info("Predicts of {}:{}".format(valid_image_file_list[ino],
rec_res[ino])) rec_res[ino]))
if args.benchmark:
mems = {
'cpu_rss_mb': cpu_mem / count,
'gpu_rss_mb': gpu_mem / count,
'gpu_util': gpu_util * 100 / count
}
else:
mems = None
logger.info("The predict time about recognizer module is as follows: ")
rec_time_dict = text_recognizer.rec_times.report(average=True)
rec_model_name = args.rec_model_dir
if args.benchmark:
# construct log information
model_info = {
'model_name': args.rec_model_dir.split('/')[-1],
'precision': args.precision
}
data_info = {
'batch_size': args.rec_batch_num,
'shape': 'dynamic_shape',
'data_num': rec_time_dict['img_num']
}
perf_info = {
'preprocess_time_s': rec_time_dict['preprocess_time'],
'inference_time_s': rec_time_dict['inference_time'],
'postprocess_time_s': rec_time_dict['postprocess_time'],
'total_time_s': rec_time_dict['total_time']
}
benchmark_log = benchmark_utils.PaddleInferBenchmark(
text_recognizer.config, model_info, data_info, perf_info, mems)
benchmark_log("Rec")
if __name__ == "__main__": if __name__ == "__main__":
......
...@@ -158,7 +158,7 @@ def main(args): ...@@ -158,7 +158,7 @@ def main(args):
img = np.random.uniform(0, 255, [640, 640, 3]).astype(np.uint8) img = np.random.uniform(0, 255, [640, 640, 3]).astype(np.uint8)
for i in range(10): for i in range(10):
res = text_sys(img) res = text_sys(img)
total_time = 0 total_time = 0
cpu_mem, gpu_mem, gpu_util = 0, 0, 0 cpu_mem, gpu_mem, gpu_util = 0, 0, 0
_st = time.time() _st = time.time()
...@@ -175,12 +175,6 @@ def main(args): ...@@ -175,12 +175,6 @@ def main(args):
dt_boxes, rec_res = text_sys(img) dt_boxes, rec_res = text_sys(img)
elapse = time.time() - starttime elapse = time.time() - starttime
total_time += elapse total_time += elapse
if args.benchmark and idx % 20 == 0:
cm, gm, gu = get_current_memory_mb(0)
cpu_mem += cm
gpu_mem += gm
gpu_util += gu
count += 1
logger.info( logger.info(
str(idx) + " Predict time of %s: %.3fs" % (image_file, elapse)) str(idx) + " Predict time of %s: %.3fs" % (image_file, elapse))
...@@ -215,61 +209,6 @@ def main(args): ...@@ -215,61 +209,6 @@ def main(args):
logger.info("\nThe predict total time is {}".format(total_time)) logger.info("\nThe predict total time is {}".format(total_time))
img_num = text_sys.text_detector.det_times.img_num img_num = text_sys.text_detector.det_times.img_num
if args.benchmark:
mems = {
'cpu_rss_mb': cpu_mem / count,
'gpu_rss_mb': gpu_mem / count,
'gpu_util': gpu_util * 100 / count
}
else:
mems = None
det_time_dict = text_sys.text_detector.det_times.report(average=True)
rec_time_dict = text_sys.text_recognizer.rec_times.report(average=True)
det_model_name = args.det_model_dir
rec_model_name = args.rec_model_dir
# construct det log information
model_info = {
'model_name': args.det_model_dir.split('/')[-1],
'precision': args.precision
}
data_info = {
'batch_size': 1,
'shape': 'dynamic_shape',
'data_num': det_time_dict['img_num']
}
perf_info = {
'preprocess_time_s': det_time_dict['preprocess_time'],
'inference_time_s': det_time_dict['inference_time'],
'postprocess_time_s': det_time_dict['postprocess_time'],
'total_time_s': det_time_dict['total_time']
}
benchmark_log = benchmark_utils.PaddleInferBenchmark(
text_sys.text_detector.config, model_info, data_info, perf_info, mems,
args.save_log_path)
benchmark_log("Det")
# construct rec log information
model_info = {
'model_name': args.rec_model_dir.split('/')[-1],
'precision': args.precision
}
data_info = {
'batch_size': args.rec_batch_num,
'shape': 'dynamic_shape',
'data_num': rec_time_dict['img_num']
}
perf_info = {
'preprocess_time_s': rec_time_dict['preprocess_time'],
'inference_time_s': rec_time_dict['inference_time'],
'postprocess_time_s': rec_time_dict['postprocess_time'],
'total_time_s': rec_time_dict['total_time']
}
benchmark_log = benchmark_utils.PaddleInferBenchmark(
text_sys.text_recognizer.config, model_info, data_info, perf_info, mems,
args.save_log_path)
benchmark_log("Rec")
if __name__ == "__main__": if __name__ == "__main__":
......
...@@ -124,76 +124,6 @@ def parse_args(): ...@@ -124,76 +124,6 @@ def parse_args():
return parser.parse_args() return parser.parse_args()
class Times(object):
def __init__(self):
self.time = 0.
self.st = 0.
self.et = 0.
def start(self):
self.st = time.time()
def end(self, accumulative=True):
self.et = time.time()
if accumulative:
self.time += self.et - self.st
else:
self.time = self.et - self.st
def reset(self):
self.time = 0.
self.st = 0.
self.et = 0.
def value(self):
return round(self.time, 4)
class Timer(Times):
def __init__(self):
super(Timer, self).__init__()
self.total_time = Times()
self.preprocess_time = Times()
self.inference_time = Times()
self.postprocess_time = Times()
self.img_num = 0
def info(self, average=False):
logger.info("----------------------- Perf info -----------------------")
logger.info("total_time: {}, img_num: {}".format(self.total_time.value(
), self.img_num))
preprocess_time = round(self.preprocess_time.value() / self.img_num,
4) if average else self.preprocess_time.value()
postprocess_time = round(
self.postprocess_time.value() / self.img_num,
4) if average else self.postprocess_time.value()
inference_time = round(self.inference_time.value() / self.img_num,
4) if average else self.inference_time.value()
average_latency = self.total_time.value() / self.img_num
logger.info("average_latency(ms): {:.2f}, QPS: {:2f}".format(
average_latency * 1000, 1 / average_latency))
logger.info(
"preprocess_latency(ms): {:.2f}, inference_latency(ms): {:.2f}, postprocess_latency(ms): {:.2f}".
format(preprocess_time * 1000, inference_time * 1000,
postprocess_time * 1000))
def report(self, average=False):
dic = {}
dic['preprocess_time'] = round(
self.preprocess_time.value() / self.img_num,
4) if average else self.preprocess_time.value()
dic['postprocess_time'] = round(
self.postprocess_time.value() / self.img_num,
4) if average else self.postprocess_time.value()
dic['inference_time'] = round(
self.inference_time.value() / self.img_num,
4) if average else self.inference_time.value()
dic['img_num'] = self.img_num
dic['total_time'] = round(self.total_time.value(), 4)
return dic
def create_predictor(args, mode, logger): def create_predictor(args, mode, logger):
if mode == "det": if mode == "det":
model_dir = args.det_model_dir model_dir = args.det_model_dir
...@@ -212,11 +142,10 @@ def create_predictor(args, mode, logger): ...@@ -212,11 +142,10 @@ def create_predictor(args, mode, logger):
model_file_path = model_dir + "/inference.pdmodel" model_file_path = model_dir + "/inference.pdmodel"
params_file_path = model_dir + "/inference.pdiparams" params_file_path = model_dir + "/inference.pdiparams"
if not os.path.exists(model_file_path): if not os.path.exists(model_file_path):
logger.info("not find model file path {}".format(model_file_path)) raise ValueError("not find model file path {}".format(model_file_path))
sys.exit(0)
if not os.path.exists(params_file_path): if not os.path.exists(params_file_path):
logger.info("not find params file path {}".format(params_file_path)) raise ValueError("not find params file path {}".format(
sys.exit(0) params_file_path))
config = inference.Config(model_file_path, params_file_path) config = inference.Config(model_file_path, params_file_path)
...@@ -332,7 +261,7 @@ def create_predictor(args, mode, logger): ...@@ -332,7 +261,7 @@ def create_predictor(args, mode, logger):
config.delete_pass("conv_transpose_eltwiseadd_bn_fuse_pass") config.delete_pass("conv_transpose_eltwiseadd_bn_fuse_pass")
if mode == 'table': if mode == 'table':
config.delete_pass("fc_fuse_pass") # not supported for table config.delete_pass("fc_fuse_pass") # not supported for table
config.switch_use_feed_fetch_ops(False) config.switch_use_feed_fetch_ops(False)
config.switch_ir_optim(True) config.switch_ir_optim(True)
...@@ -597,30 +526,5 @@ def draw_boxes(image, boxes, scores=None, drop_score=0.5): ...@@ -597,30 +526,5 @@ def draw_boxes(image, boxes, scores=None, drop_score=0.5):
return image return image
def get_current_memory_mb(gpu_id=None):
"""
It is used to Obtain the memory usage of the CPU and GPU during the running of the program.
And this function Current program is time-consuming.
"""
import pynvml
import psutil
import GPUtil
pid = os.getpid()
p = psutil.Process(pid)
info = p.memory_full_info()
cpu_mem = info.uss / 1024. / 1024.
gpu_mem = 0
gpu_percent = 0
if gpu_id is not None:
GPUs = GPUtil.getGPUs()
gpu_load = GPUs[gpu_id].load
gpu_percent = gpu_load
pynvml.nvmlInit()
handle = pynvml.nvmlDeviceGetHandleByIndex(0)
meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
gpu_mem = meminfo.used / 1024. / 1024.
return round(cpu_mem, 4), round(gpu_mem, 4), round(gpu_percent, 4)
if __name__ == '__main__': if __name__ == '__main__':
pass pass
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册