Memory leak in AugmentationSequence

[bnawras]

There seems to be a memory leak in augraphy.AugmentationSequence

My environment

Dockerfile

FROM nvidia/cuda:11.1.1-runtime-ubuntu20.04

ENV DEBIAN_FRONTEND noninteracrive
RUN apt-get update
RUN apt-get upgrade -y
RUN apt-get install -y \
        build-essential git python3 python3-pip \
        ffmpeg libsm6 libxext6 libxrender1 libglib2.0-0

WORKDIR /app
COPY requirements.txt .
RUN pip install --ignore-installed -r requirements.txt

CMD jupyter lab --ip 0.0.0.0 --port 1110 --allow-root

Requirements

torch==1.13.1
torchvision==0.14.1
albumentations==1.3.0
augraphy==8.2.4
opencv-python==4.8.1.78

Сode to reproduce the leak

num_workers = 8
n_epoch = 10
batch_size = 4

ink_phase = [ ]
post_phase = [ ]
paper_phase = [
    agr.AugmentationSequence(
        [
            agr.NoiseTexturize(
                sigma_range=(2, 3),
                turbulence_range=(2, 3),
                p=1
            ),
            agr.BrightnessTexturize(
                texturize_range=(0.999, 0.9999),
                deviation=0.01,
                p=1
            ),
        ],
    p=0.1),
]

pipeline = agr.AugraphyPipeline(ink_phase, paper_phase, post_phase)
def paperfy(image, **params):
    return pipeline(image)

synth_augmentations = A.Compose([
    A.RandomCrop(width=900, height=900),
    A.Lambda(image=paperfy, p=0.8),
    ToTensorV2()
])

train_set = # any dataset with synth_augmentations
trainloader = data.DataLoader(train_set, batch_size=batch_size, num_workers=num_workers, shuffle=True, pin_memory=True, persistent_workers=True)

for epoch in range(n_epoch):
    for i, samples in enumerate(trainloader):
        pass

:fire: workaround: don't use AugmentationSequence or persistent_workers=False

[kwcckw]

I can see a similar issue here: https://github.com/pytorch/pytorch/issues/62066

Could you cross check again and see if it is really related to AugmentationSequence?

[bnawras]

@kwcckw I have not reviewed the details of the implementation of AugmentationSequence, but this is directly or indirectly related to this module. This is what will happen if we run the same code without it:

• persistent_workers=True
• removed AugmentationSequence

  paper_phase = [
        agr.NoiseTexturize(
            sigma_range=(2, 3),
            turbulence_range=(2, 3),
            p=1
        ),
        agr.BrightnessTexturize(
            texturize_range=(0.999, 0.9999),
            deviation=0.01,
            p=1
        ),
  ]

memory usage:

[kwcckw]

Thanks, i was able to reproduce the problem and fix it with this workaround from my end:

https://github.com/sparkfish/augraphy/pull/418

Could you reinstall it from the repo and try again?

[bnawras]

fixed!