问题咨询

[Toms-BigData]

您好，
我在复现您的代码时，出现了一些bug，但我不知道如何去调整，望您从百忙之中，抽空看一下，感谢您。
我在执行Training memory bank model/train_singlenet_phase_1fc.py时出现的bug如图所示：

[YuemingJin]

你好，是因为你的data读取路径不对

[HannahHaensen]

That is actually not the case.

The problem is with the preprocessing, if the preprocessing is done as described, we have e.g. 2346 images for folder data/cutMargin/5

but the get labels file tries the following

for phase_line in phase_file:
        phase_split = phase_line.split()
        if first_line:
            first_line = False
            continue
        if int(phase_split[0]) % downsample_rate == 0:
            info_each = []
            img_file_each_path = os.path.join(img_dir_paths2[j], phase_split[0] + '.jpg')

it reads out the phase annotation, the specific frame but actually we downscaled the 25fps so 1 fps to file 54000 can not exist.

the phase annotation files is:

Frame Phase
0  Preparation
...
54000 GallbladderRetraction

an approach could be to use an index instead starting at 1, as the images also start with 1 and not zero

[gxr12]

Hi, I have the same problem as you, have you solved it?

[HannahHaensen]

@gxr12 you can check out my fork

https://github.com/HannahHaensen/TMRNet/blob/main/code/Training%20memory%20bank%20model/get_paths_labels.py

i think this was the file I solved it. But I can check later and post a sufficient reply currently just available via phone

[YuemingJin]

Thanks @HannahHaensen for the reply!
We can also adjust the names of downsampled frames to ensure alignment with the GT labels.