utils.py

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import time
import os
import ast
import argparse


def argsparser():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--model_dir",
        type=str,
        default=None,
        help=("Directory include:'model.pdiparams', 'model.pdmodel', "
              "'infer_cfg.yml', created by tools/export_model.py."),
        required=True)
    parser.add_argument(
        "--image_file", type=str, default=None, help="Path of image file.")
    parser.add_argument(
        "--image_dir",
        type=str,
        default=None,
        help="Dir of image file, `image_file` has a higher priority.")
    parser.add_argument(
        "--video_file",
        type=str,
        default=None,
        help="Path of video file, `video_file` or `camera_id` has a highest priority."
    )
    parser.add_argument(
        "--camera_id",
        type=int,
        default=-1,
        help="device id of camera to predict.")
    parser.add_argument(
        "--threshold", type=float, default=0.5, help="Threshold of score.")
    parser.add_argument(
        "--output_dir",
        type=str,
        default="output",
        help="Directory of output visualization files.")
    parser.add_argument(
        "--run_mode",
        type=str,
        default='fluid',
        help="mode of running(fluid/trt_fp32/trt_fp16/trt_int8)")
    parser.add_argument(
        "--use_gpu",
        type=ast.literal_eval,
        default=False,
        help="Whether to predict with GPU.")
    parser.add_argument(
        "--run_benchmark",
        type=ast.literal_eval,
        default=False,
        help="Whether to predict a image_file repeatedly for benchmark")
    parser.add_argument(
        "--enable_mkldnn",
        type=ast.literal_eval,
        default=False,
        help="Whether use mkldnn with CPU.")
    parser.add_argument(
        "--cpu_threads", type=int, default=1, help="Num of threads with CPU.")
    parser.add_argument(
        "--use_dynamic_shape",
        type=ast.literal_eval,
        default=False,
        help="Dynamic_shape for TensorRT.")
    parser.add_argument(
        "--trt_min_shape", type=int, default=1, help="min_shape for TensorRT.")
    parser.add_argument(
        "--trt_max_shape",
        type=int,
        default=1280,
        help="max_shape for TensorRT.")
    parser.add_argument(
        "--trt_opt_shape",
        type=int,
        default=640,
        help="opt_shape for TensorRT.")
    parser.add_argument(
        "--trt_calib_mode",
        type=bool,
        default=False,
        help="If the model is produced by TRT offline quantitative "
        "calibration, trt_calib_mode need to set True.")

    return parser


class Times(object):
    def __init__(self):
        self.time = 0.
        # start time
        self.st = 0.
        # end time
        self.et = 0.

    def start(self):
        self.st = time.time()

    def end(self, repeats=1, accumulative=True):
        self.et = time.time()
        if accumulative:
            self.time += (self.et - self.st) / repeats
        else:
            self.time = (self.et - self.st) / repeats

    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.preprocess_time = Times()
        self.inference_time = Times()
        self.postprocess_time = Times()
        self.img_num = 0

    def info(self, average=False):
        total_time = self.preprocess_time.value() + self.inference_time.value(
        ) + self.postprocess_time.value()
        total_time = round(total_time, 4)
        print("------------------ Inference Time Info ----------------------")
        print("total_time(ms): {}, img_num: {}".format(total_time * 1000,
                                                       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 = total_time / self.img_num
        print("average latency time(ms): {:.2f}, QPS: {:2f}".format(
            average_latency * 1000, 1 / average_latency))
        print(
            "preprocess_time(ms): {:.2f}, inference_time(ms): {:.2f}, postprocess_time(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
        total_time = self.preprocess_time.value() + self.inference_time.value(
        ) + self.postprocess_time.value()
        dic['total_time'] = round(total_time, 4)
        return dic


def get_current_memory_mb():
    """
    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
    gpu_id = os.environ.get('CUDA_VISIBLE_DEVICES', 0)

    pid = os.getpid()
    p = psutil.Process(pid)
    info = p.memory_full_info()
    cpu_mem = info.uss / 1024. / 1024.
    gpu_mem = 0
    gpu_percent = 0
    gpus = GPUtil.getGPUs()
    if gpu_id is not None and len(gpus) > 0:
        gpu_percent = gpus[gpu_id].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)


class LoggerHelper(object):
    def __init__(self, args, times, mem_info=None):
        """
        args: utility.parse_args()
        times: The Timer class
        """
        self.args = args
        self.times = times
        self.model_name = args.model_dir.strip('/').split('/')[-1]
        self.batch_size = 1
        self.shape = "dynamic shape"
        if args.run_mode == 'fluid':
            self.precision = "fp32"
            self.use_tensorrt = False
        else:
            self.precision = args.run_mode.split('_')[-1]
            self.use_tensorrt = True

        self.device = "gpu" if args.use_gpu else "cpu"
        self.preprocess_time = round(times['preprocess_time'], 4)
        self.inference_time = round(times['inference_time'], 4)
        self.postprocess_time = round(times['postprocess_time'], 4)
        self.data_num = times['img_num']
        self.total_time = round(times['total_time'], 4)
        self.mem_info = {"cpu_rss": 0, "gpu_rss": 0, "gpu_util": 0}
        if mem_info is not None:
            self.mem_info = mem_info

    def report(self):
        print("\n")
        print("----------------------- Config info -----------------------")
        print("runtime_device:", self.device)
        print("ir_optim:", True)
        print("enable_memory_optim:", True)
        print("enable_tensorrt:", self.use_tensorrt)
        print("precision:", self.precision)
        print("enable_mkldnn:", self.args.enable_mkldnn)
        print("cpu_math_library_num_threads:", self.args.cpu_threads)

        print("----------------------- Model info ----------------------")
        print("model_name:", self.model_name)

        print("------------------------ Data info ----------------------")
        print("batch_size:", self.batch_size)
        print("input_shape:", self.shape)

        print("----------------------- Perf info -----------------------")
        print("[cpu_rss(MB): {} gpu_rss(MB): {}, gpu_util: {}%".format(
            round(self.mem_info['cpu_rss'], 4),
            round(self.mem_info['gpu_rss'], 4),
            round(self.mem_info['gpu_util'], 2)))
        print("total number of predicted data: {} and total time spent(s): {}".
              format(self.data_num, self.total_time))
        print(
            "preproce_time(ms): {}, inference_time(ms): {}, postprocess_time(ms): {}".
            format(self.preprocess_time * 1000, self.inference_time * 1000,
                   self.postprocess_time * 1000))