UNet_Decoder_CRF

Setup (Compatible with NERSC Perlmutter)

1. Create a conda environment

Conda is a great tool to import tools and create environments for projects. To avoid any conflicts with existing work or other versions of required dependencies, use conda to create a virtual environment. Make sure to activate the environment before proceeding to the next steps.

• Python Version: 3.8 conda create --name CRF_CPU_Env python=3.8 conda activate CRF_CPU_Env

2. Clone the repository

• git clone https://github.com/mavaylon1/UNet_Decoder_CRF.git
• pip install -e .
• pip install protobuf==3.20.*

3. Install cudatoolkit and cudnn via conda

In the current state, if users want to use GPU for training their models, users need to install the following cudatoolkit and cudnn versions.

• conda install cudatoolkit==11.3.1 cudnn==8.2.1

4. Install tensorflow v. 2.6

In the current state, tensorflow 2.6.0 is confirmed as the latest version to work with the cuda and cudnn versions above on Perlmutter GPU.

• pip install tensorflow==2.6

5. Downgrade keras

As of August 2022, the keras installed with tensorflow v. 2.6 is not compatible. The fix we found was to downgrade keras to 2.6:

• pip install keras==2.6.*

4. CRF Setup

From the cloned directory go to the src folder.

• cd src/UNet_Decoder_CRF/crf/cpp
• make

If you run into issues with the CRF, refer to https://github.com/sadeepj/crfasrnn_keras and follow the demo. Depending on the configuration, one may need to specify the absolute path to high_dim_filter.so in high_dim_filter_loader.py for tensorflow module load.