Loading a config from .yaml file locally

[InexperiencedMe]

I want to load a config file, that has been saved right after training because I have separate scripts for training and inference. I saved the config file according to a tutorial with following lines:

import yaml
    # Save the configuration to a config.yaml file
    config_yaml_path = "output/myConfig.yaml"
    with open(config_yaml_path, 'w') as file:
        yaml.dump(cfg, file)

However, in the tutorial the data was never loaded back. All other tutorials on the internet use a python cloud notebook, so it's all done in one instance and they never need to save or load config for inference, since its instantiated and modified by the training accordingly.

I have tried almost all possible combinations with cfg.merge_from_file(), load_yaml_with_base(), CfgNode casting, opening a file and all that. Thanks in advance.

[augusts-bit]

You have tried

cfg = get_cfg()

cfg.merge_from_file(config_path)

?

I do not know if that answers anything, but else I am doing something similar using SAHI:

%pip install -U sahi

from sahi.utils.detectron2 import Detectron2TestConstants
from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction, predict, get_prediction
from sahi.utils.file import download_from_url
from sahi.utils.cv import read_image
from sahi.utils.detectron2 import export_cfg_as_yaml

Maybe it also how you are saving your config file. I also had trouble with dump. I am saving my yaml file now as follows:

# After training detectron2 model, saving yaml file
export_cfg_as_yaml(cfg, export_path=os.path.join(cfg.OUTPUT_DIR, "model_cfg.yaml"))

Then

# Load detection model from trained Detectron2 model
detection_model = AutoDetectionModel.from_pretrained(
    model_type='detectron2',
    model_path="outputs/model/model_weights.pt", 
    config_path="outputs/model/model_cfg.yaml", 
    confidence_threshold=0.3,
    image_size=640,
    device= #'cuda:0' # "cpu"
)

results = []

for img in imgs: # imgs is a list of image paths I want to do predictions on

  # Takes very long if large image
  result = get_sliced_prediction(
      img, # path to image
      detection_model,
      slice_height = 250, # If you don't want to slice, make height and width as large as your total image
      slice_width = 250, # i.d.
      overlap_height_ratio = 0.1, # Can be 0
      overlap_width_ratio = 0.1,
  )

  results.append(result)