NJUVISION / rho-vision

ρ-Vision: Efficient Visual Computing with Camera RAW Snapshots
https://njuvision.github.io/rho-vision
MIT License
39 stars 3 forks source link

Incomplete installation instructions #8

Open maartenvds opened 10 months ago

maartenvds commented 10 months ago

Hi,

I'm trying to install this package using the requirements.txt file, but it seems the be incomplete. It is at least missing torch and torchvision. I can install them manually, but I don't know what exact version works well with mmcv-full 1.3.8 and mmgen 0.5.0.

Could you please provide the used pytorch version and also the exact python version.

Thanks

lizhihao6 commented 10 months ago

Hi, maartenvds

Pytorch 1.7.0 with Python 3.7 should be compatible and work well.

maartenvds commented 10 months ago

Hi lizhihao6,

The installation instructions I currently use are (based on mmgen installation instructions, your mmcv-full version, mmgen version and trial and error):

First pytorch

pip install torch==1.7.0 torchvision

Then mmcv-full

pip install mmcv-full==1.3.8 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.7.0/index.html

Then the other packages

pip install mmcls mmgen==0.5.0 matplotlib==3.5.1 rawpy

Installation completes without errors.

Then I run:

python inference.py --ckpt checkpoints/bdd100k_rgb_to_iphone_raw.pth --rgb resources/bdd100k.jpg

which gives:

File "/home/maarten/.virtualenvs/p37_rho_vision/lib/python3.7/site-packages/mmgen/datasets/pipelines/augmentation.py", line 4, in <module>
    from mmcls.datasets import PIPELINES as CLS_PIPELINE
  File "/home/maarten/.virtualenvs/p37_rho_vision/lib/python3.7/site-packages/mmcls/__init__.py", line 57, in <module>
    f'MMCV=={mmcv.__version__} is used but incompatible. ' \
AssertionError: MMCV==1.3.8 is used but incompatible. Please install mmcv>=1.3.16, <=1.5.0.

What happens here is that mmgen 0.5.0 requires mmcls 0.18.0, which in turn requires mmcv>=1.3.16, <=1.5.0. But you apparently using mmcv-full 1.3.8, so I'm not there yet.

This is quite the versioning nightmare. I got around it by installing mmcls 0.17.0, which seems to work but it gives pip dependency errors.

Could you, for the sake of easily reproducing your installation, do a pip freeze or something so I can get a hint on all of your package versions?