Using stardist-conic without pip install

[ajinkya-kulkarni]

Hello team, I was trying to try out the conic model (https://github.com/stardist/stardist/tree/conic-2022/examples/conic-2022) on my custom dataset. I tried several times installing the conic branch using pip install. I was wondering if I can simply use the conic branch without a pip install. Like locally importing the necessary libraries after downloading the code.

Thanks!

[uschmidt83]

Unfortunately, I don't think this will work since we also changed the C++ code in the conic-2022 branch. I.e. if you first install the "normal" StarDist (pip install stardist) and then "replace" StarDist's Python source code files with those from the conic-2022 branch, the post-processing will throw an error (or not work correctly).

Did you read the troubleshooting section?

[ajinkya-kulkarni]

Yes, I read the troubleshooting section as well. This is how I tried to install the conic branch:

conda create stardist-conic python=3.9 followed by brew install gcc and then CC=gcc-10 CXX=g++-10 pip install git+https://github.com/stardist/stardist.git@conic-2022.
But this gives an error and the installation stops. Am I doing anything wrong? I am using a macbook, but I get the same error on a Ubuntu machine.

[uschmidt83]

conda create stardist-conic python=3.9 followed by brew install gcc and then CC=gcc-10 CXX=g++-10 pip install git+https://github.com/stardist/stardist.git@conic-2022.

On my Mac, brew install gcc tries to install gcc version 13, i.e. gcc-10 / g++-10 have to be changed to gcc-13 / g++-13. You can install a specific gcc version like this: brew install gcc@13. (The same applies to linux.)

This might work:

brew install gcc@13
conda create stardist-conic python=3.9
# ...
conda activate stardist-conic
CC=gcc-13 CXX=g++-13 pip install git+https://github.com/stardist/stardist.git@conic-2022

But this gives an error and the installation stops.

It would have been helpful to see the error message.

I am using a macbook

Despite what I wrote above, compiling on Mac might also work without installing gcc first. I.e. try this:

brew install libomp  # optional to speed up some computations
conda create stardist-conic python=3.9
# ...
conda activate stardist-conic
pip install git+https://github.com/stardist/stardist.git@conic-2022

[ajinkya-kulkarni]

Thanks, I will try your approach now. In the meantime, is there a way you can share the conic trained model, so it will be easier for me to test it on custom HE image?

[ajinkya-kulkarni]

Hello @uschmidt83, I was able to pip install the conic branch on a Ubuntu machine, but only the CPU version. The training for 1000 epochs as mentioned in the train notebook would roughly take one month to complete on my CPU workstation.
It would therefore be great if you could share/upload the trained model as well, as I can test it out on my colon HE images, instead of waiting for a month for the training to finish.

Best, Ajinkya

[maweigert]

The training for 1000 epochs as mentioned in the train notebook would roughly take one month to complete on my CPU workstation.

There seems to be a problem with tensorflow not recognizing your GPU. You can check this with

tf.test.is_gpu_available()