Example code returns flat image without any segment.

[kosuke1701]

First of all, thank you very much for sharing your interesting project!

I tried to setup an environment with Docker, and the following example code ran without error and output three images as expected.

python segment.py ./emilia.jpg

However, the output looks like somewhat different from what is shown on README. (The following image is current_skelton.png which is output by the sample code.)

Dockerfile

Note: I've changed numba version according to this issue.

FROM nvidia/cuda:9.0-cudnn7-runtime-ubuntu16.04

RUN apt-get update\
    && apt-get install -y --no-install-recommends \
    wget gcc make zlib1g-dev libssl-dev libopencv-dev \
    && apt-get -y clean \
    && rm -rf /var/lib/apt/lists/*

RUN apt-get remove -y --allow-change-held-packages libcudnn7


RUN apt-get update && apt-get install -y --no-install-recommends \
    libcudnn7=7.0.5.15-1+cuda9.0 \
    && apt-mark hold libcudnn7 && \
    rm -rf /var/lib/apt/lists/*

RUN wget https://www.python.org/ftp/python/3.6.11/Python-3.6.11.tgz \
    && tar zxvf Python-3.6.11.tgz \
    && cd Python-3.6.11 \
    && ./configure \
    && make && make install \
    && ln -s /usr/local/bin/python3.6 /usr/local/bin/python \
    && ln -s /usr/local/bin/pip3.6 /usr/local/bin/pip \
    && cd .. \
    && rm -r Python-3.6.11*

RUN pip install -U pip \
    && pip install tensorflow-gpu==1.5.0 \
    && pip install keras==2.2.4 \
    && pip install opencv-python==3.4.2.17 \
    && pip install numpy==1.15.4 \
    && pip install numba==0.49.0 \
    && pip install scipy==1.1.0 \
    && pip install scikit-image==0.13.0 \
    && pip install scikit-learn==0.22.2 \
    && pip install -U h5py==2.10.0 \
    && pip cache purge

What I did

# Working directory is a directory where I created the Dockerfile.
docker build -t sample:0.1 .

# Mount project directory and execute sample code.
sudo docker run -it --rm --gpus all -v /path/to/DanbooRegion:/DanbooRegion --name debug sample:0.1 /bin/bash
$ cd /DanbooRegion/code
$ python segment.py ./emilia.jpg

Environment

• OS: Ubuntu 20.04 LTS
• Docker version: 20.10.5, build 55c4c88
• nvidia-docker2 version: 2.6.0-1
• GPU: GeForce RTX 3090
• Nvidia driver version: 460.73.01

[lllyasviel]

Do you mean that you get a blank white output image and cannot get any meaningful outputs? You have uploaded a blank white image, and that image is what you get?

[kosuke1701]

Yes. The uploaded blank white image is what I got as a skelton map. The other two images (current_flatten.png, current_region.png) are also blank images with different colors.

[lllyasviel]

If no python errors are reported, it is likely that the models are not properly loaded. Have you downloaded the pretrained model and put it in the correct places?

[kosuke1701]

I used the following pretrained models which are uploaded to the github repo.

https://github.com/lllyasviel/DanbooRegion/blob/master/code/DanbooRegion2020UNet.net
https://github.com/lllyasviel/DanbooRegion/blob/master/code/srcnn.net

Currently, I'm training a new model from scratch using a provided training code to see whether it works or not.

EDITED:

I tested the new model which is trained from scratch, but it still returned blank images.

[kosuke1701]

It seems that Ampere architecture of RTX 3090 is not supported by CuDNN 7, which is required by tensorflow version 1.

https://docs.nvidia.com/deeplearning/cudnn/support-matrix/index.html

Since it seems to be an issue at lower layer related to hardware, I close this issue. Thank you for your responses.

[lllyasviel]

Thank you for your report! Another thing is that I have seen many of my clients with 3090/30XX using tf 1.4 without any trouble. But the official document does say they do not support these versions. Have you done anything unique to your environment?

[kosuke1701]

I think I have followed standard procedures to setup my environment. After a bit of research, I found that there is a way to use tensorflow 1.15 with 3090 GPU.

https://www.pugetsystems.com/labs/hpc/How-To-Install-TensorFlow-1-15-for-NVIDIA-RTX30-GPUs-without-docker-or-CUDA-install-2005/

After I set up new tensorflow 1.15 environment with the above procedure, I got different output as follows, and it still seems to be not working :( I will try same environment setup on my different PC with RTX 1070 Ti to see if something different happens when I return home.