kangaroo-detection/environment

Environment Setups

Note

Experiments are done on window laptop with GTX 1060.

Quick start

Duplicate my environment with conda using environment config (.yml files)

    conda env create -f <config-filename>.yml

• base-env.yml: bare bone environment that can use GPU computation
• full-env.yml: full environment used in this project

Self replicate

If you wanna install everything yourself. Follow instructions below

Important packages

For GPU computation and basic needs

Our environment will be named 'kangaroo'

1. Python: 3.8

   conda create -n kangaroo python=3.8
   

2. Activate: conda activate kangaroo

3. Check python version

   python --version
   > Python 3.8.15
   

4. ⭐ Tensorflow-GPU:

   conda install tensorflow-gpu==2.5.0
   

   (2.3.0) does not work !! at all This install also included numpy (and keras)

5. (Optional) Check and sync/downgrade tensorflow-base version To match with tensorflow version

   conda install tensorflow-base==2.5.0
   

6. Verify installation packages recognizes your GPU

    python src/scripts/verify_gpu.py 
   

7. Continue install all other packages

8. Clone submodule and install Mask-RCNN

    # clone
    git submodule init
    git submodule update
   
    # install
    cd src/submodules/akTwelve_MaskRCNN
    python setup.py install
   

Other packages

Exception:

1. Mask_RCNN: Link

Other used packages:

  conda install -c conda-forge [packages]

1. Pillow: Image loading/processing (also the base of Keras Image)
2. scikit-image: (! version: 0.16.2)
3. ipykernel
4. pycocotools: Needed for training on windows (pip)
5. dvc

Errors and Fixes

• Error: When installing pycocotools: Missing C++ stuffs

  • Suggested fix: Install C++ Build Tool 2015

• Pitfall: Wrong installation of build tools / installation with mismatching C++/SDK version / missing components

  • Fix: Install Dev package for C++ with win 10 SDK (Reference)
  • Specifications: Exported to .vsconfig

• Error: module 'keras.engine' has no attribute 'Layer'

  • Discussion: Link
  • Fix: Use this Repo and make sure to install pycocotools

• Keras conflict with scikit-image in RCNN code

  • Error output: Input image dtype is bool. Interpolation is not defined with bool data type
  • Ref/Suggestions: here
  • Resolve: down grade scikit-image from 0.19.3 to 0.16.2

• Getting stuck with Error No module named 'tensorflow_core.estimator'

  • Reasons/Discussion: here
  • Resolve: downgrade tensorflow-estimator to same version as tensorflow

• Tensorflow 2.1 conflict with h5py

  • Ref
  • Solve: use h5py ver 2.10.0

Other versions tested

1:

• python 3.7
• tensorflow-gpu 2.1.0