kangaroo-detection/environment

Note

Experiments are done on window laptop with GTX 1060.

Scripts

in `/scripts

• verify_gpu.py: check if gpu exists

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: 3.7

   conda create -n kangaroo python=3.7
   

   I also recommend python 3.7 for compatibility purposes (from personal experience)

2. Activate: conda activate kangaroo

3. Check python version

   python --version
   > Python 3.7.15
   

4. Ipykernel: conda install ipykernel

5. ⭐ Tensorflow-GPU:

   conda install tensorflow-gpu==2.1.0
   

   current version as of now (2.3.0) does not work for my system This install also included numpy (and keras)

6. Verify install with /notebooks/test-gpu.ipynb

Other packages

Other used packages

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

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