dlg-home-content

setup environment

• conda env create -f environment.yml

• install detectron2 from source

• cpu version

> conda install pytorch torchvision cpuonly -c pytorch
> python -m pip install detectron2 -f \
  https://dl.fbaipublicfiles.com/detectron2/wheels/cpu/torch1.6/index.html
> update environment `conda env update --file environment.yml`

• for other gpu versions, use this

CLI commands available

• convert labelme2coco

labelme2coco --labelme_json_location 'data/processed_tv_annotations_v1/' --labels_loc "assets/keypoints.yml" --save_json "data/keypoints/" --train_ratio 0.9 --seed 50

• train using custom dataset

We need to define three config files

• base cfg file name available on detectron. check detectron/configs for examples.
• cfg file which contains modified params . check configs folder for specific examples
• data_cfg which has dataset and keypoints related params. For example assets/datasets.yml

# normal instance segmentation
custom_train --base_cfg 'COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml' --cfg 'configs/mask_only_exp1.yml' --data_cfg "assets/datasets.yml"

# instance segmentation with keypoints
custom_train --base_cfg 'COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml' --cfg 'configs/keypoint_mask_on_exp1.yml' --data_cfg "assets/datasets.yml"

Inference

LOGO Detection

Download latest inference file from here

from dlg_home_content.tv_detection import InferLogo
config = '../assets/e2e_infer.yml
model = InferLogo(config)
model.predict(img_loc, visualize=True)

Inference for Keypoint Detetion

Download weight files and config files from [here] (https://fractalanalytic-my.sharepoint.com/:u:/g/personal/sindhura_k_fractal_ai/EXCaFSHWv3hMo99lvfP4zKIBLBO8dlnWzY7iUAFWYiXHKA?e=23XheZ)

#for inner keyoint detection
from dlg_home_content.inference_pipeline import KeypointInference

config = '../assets/e2e_infer.yml'
#kp_type in ['kp_inner_edge','kp_outer_edge','kp_sticky_note']
model_inner = KeypointInference(config, kp_type='kp_inner_edge')
predicted_keyoints = model_inner.predict_keypoints(img_loc, visualize=True)

End-to-End Inference pipeline

from dlg_home_content.e2e_inference import E2EInference
config = '../assets/e2e_infer.yml'
final_pipeline = E2EInference(config)
result = final_pipeline.infer(img_loc, 8, 8, True)