Timing may not be accurate

[YiqunChen1999]

Timing in Tracker may not be accurate enough.

As pointed out in Pytorch Forums post, CUDA is asynchronous so you will need some tools to measure time.

I then give a simple example which may be helpful:

from typing import Union
import torch

class Tracker(object):
    # original code here.
    @property
    def is_using_cuda(self):
        return self.device != torch.device("cpu")

    def _init_timer(self):
        if self.is_using_cuda:
            self._timer_start = torch.cuda.Event(enable_timing=True)
            self._timer_stop = torch.cuda.Event(enable_timing=True)
        self._timestamp = None

    def _start_timing(self) -> Union[float, None]:
        if self.is_using_cuda:
            self._timer_start.record()
            timestamp = None
        else:
            timestamp = time.time()
            self._timestamp = timestamp
        return timestamp

    def _stop_timing(self) -> float:
        if self.is_using_cuda:
            self._timer_stop.record()
            torch.cuda.synchronize()
            # cuda event record return duration in milliseconds.
            duration = self._timer_start.elapsed_time(
                self._timer_stop
            )
            duration /= 1000.0
        else:
            duration = time.time() - self._timestamp
        return duration

    def track(...): # args are omitted here.
        self._init_timer()
        # code for initialization is omitted.
        for f, img_file in enumerate(img_files):
            # code for data preparation is omitted.
            self._start_timing()
            if f == 0: 
                self.init(image, box)
                times[f] = self._stop_timing()
            if fail_count >= 10 and method == 'restart' and f in restart_flag:
                # the tracker will be restarted when the cumulative number of failures reaches 10
                print('init again in %s' % f)                
                init_positions.append(f)
                self.init(image, anno[f,:])
                fail_count = 0
            else:
                frame_box = self.update(image) 
                frame_box = np.rint(frame_box)
                times[f] = self._stop_timing()
            # the rest code is omitted.

If you find this code snippet is helpful, you may consider update your toolkit.

[huuuuusy]

Thanks a lot; I have updated the codes based on your suggestion.