DongSky/lbba_boosted_wsod

how to use the requirement

Closed this issue · 8 comments

fsted commented

I try to install environment by pip install -r requirement，but it's error.

DongSky commented

Hi fsted.
This requirement file is exported from the anaconda virtual environment. Therefore you can use the command below to create the corresponding environment via anaconda / miniconda:

conda create --name <env> --file <this file>

Similarly, you can simply create the training environment following the two reference project mentioned in readme.txt, and make sure that the version of pytorch is 1.1.0 or 1.2.0.

HMCCMH commented

I used the anaconda. However, it made an error.

ERROR: Invalid requirement: '_libgcc_mutex=0.1=main' (from line 4 of lbba_requirements.txt)
Hint: = is not a valid operator. Did you mean == ?
DongSky commented

I used the anaconda. However, it made an error.

ERROR: Invalid requirement: '_libgcc_mutex=0.1=main' (from line 4 of lbba_requirements.txt)
Hint: = is not a valid operator. Did you mean == ?

Hi HMCCMH
We provide two solutions.

  1. install pytorch, numpy, easydict and opencv manually with conda install. Since most of packages in the config file will be installed automatically, you can focus on these core packages.
  2. I uploaded a new config file, and the line including libgcc is deleted.
HMCCMH commented

Sorry, I see the config file updated 7 months ago. The line including libgcc is existing.

HMCCMH commented

And PyTorch 1.2.0 now is only used for CUDA10.0 in the web

DongSky commented

Sorry, I see the config file updated 7 months ago. The line including libgcc is existing.

Hi HMCCMH

  1. the new config file is here https://github.com/DongSky/lbba_boosted_wsod/blob/master/lbba.yml
  2. About CUDA version, I found that my workstation installed both CUDA 10.0 and 10.1, but I wrongly stated CUDA 10.1 in the readme. Sorry for the confusion. By the way, you can adapt it to newer version of pytorch if necessary.
HMCCMH commented

Hi DongSky,
I find that your lbba.yaml use the scipy1.4.1, but the network.py use from scipy.misc import imresize . However, the version 1.4.1 can't support the imresize function. I see that the imresize function was completely abandoned in the 1.3.0 version of scipy. So what actually you use in the environment?

DongSky commented

We use scipy == 1.1.0.
Sorry I'm not sure why anaconda exported a newer version into the lbba.yml. A possible reason is that the packages installed by pip may cover the originally installed packages from anaconda, but cannot be detected by conda package manager. Sincerely sorry for the confusion.