DeepOPG (Off-the-shelf)

This repository implements DeepOPG with mostly off-the-shelf components.

Prerequisites

Anaconda

Installation

$ git clone --recurse-submodules https://github.com/stmharry/deepopg-off-the-shelf.git && cd deepopg-off-the-shelf
$ conda env create -f environment.yaml && conda activate deepopg
$ make install

Usage

Before running any commands, please set up the follow environment variables:

$ export ROOT_DIR=/path/to/root

The typical workflow is as follows:

Train a model or identify a model to use.
Run inference on the test set.
Postprocess the results.
Visualize the results.

Training

To run training, one need to specify the following variables:

ARCH: Either detectron2 or maskdino.
CONFIG_FILE: Path to the config file.

The model will be saved to $(ROOT_DIR)/models/$(MODEL_NAME) by default where MODEL_NAME is the name of an auto-generated model directory.

Example

$ make train ARCH=detectron2 CONFIG_NAME=mask_rcnn_mvitv2_t_3x.py
> PYTHONPATH=./detectron2:. PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:32 \
    python scripts/main.py \
        --main-app tools.lazyconfig_train_net:main \
        --config-file ./configs/mask_rcnn_mvitv2_t_3x.py \
        --data-dir /mnt/hdd/PANO/data \
        train.output_dir=/mnt/hdd/PANO/models/2023-07-02-070303

Inference

To run inference, one need to specify the following variables:

ARCH: Either detectron2 or maskdino.
DATASET_NAME: Name of the dataset.
MODEL_NAME: Name of the model directory.

The inference results will be saved to $(ROOT_DIR)/results/$(RESULT_NAME) by default where RESULT_NAME is the name of an auto-generated result directory.

Example

$ make test ARCH=detectron2 DATASET_NAME=pano_debug MODEL_NAME=2023-05-30-175110
> PYTHONPATH=./detectron2:. PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:32 \
    python scripts/main.py \
        --main-app tools.lazyconfig_train_net:main \
        --config-file /mnt/hdd/PANO/models/2023-05-30-175110/config.yaml \
        --data-dir /mnt/hdd/PANO/data \
        --eval-only \
        train.init_checkpoint=/mnt/hdd/PANO/models/2023-05-30-175110/model_final.pth \
        train.output_dir=/mnt/hdd/PANO/results/2023-07-02-070543 \
        dataloader.test.dataset.names=pano_debug \
        dataloader.evaluator.output_dir=/mnt/hdd/PANO/results/2023-07-02-070543 \
        model.roi_heads.box_predictor.test_score_thresh=0.0 \
        model.roi_heads.box_predictor.test_nms_thresh=0.0 \
        model.roi_heads.box_predictor.test_topk_per_image=500

Postprocess

To run postprocess, one need to specify the following variables:

ARCH: Either detectron2 or maskdino.
DATASET_NAME: Name of the dataset.
RESULT_NAME: Name of the result directory.

Example

$ make postprocess ARCH=detectron2 DATASET_NAME=pano_debug RESULT_NAME=2023-07-01-075136
> PYTHONPATH=./detectron2:. PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:32 \
    python scripts/commands.py \
        --data_dir /mnt/hdd/PANO/data \
        --result_dir /mnt/hdd/PANO/results/2023-07-01-075136 \
        --dataset_name pano_debug \
        --do_postprocess \
        --output_prediction_name instances_predictions.postprocessed.pth

Visualization

To run visualization, one need to specify the following variables:

ARCH: Either detectron2 or maskdino.
DATASET_NAME: Name of the dataset.
RESULT_NAME: Name of the result directory.

Example

$ make visualize ARCH=detectron2 DATASET_NAME=pano_debug RESULT_NAME=2023-07-01-075136
> PYTHONPATH=./detectron2:. PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:32 \
    python scripts/commands.py \
        --data_dir /mnt/hdd/PANO/data \
        --result_dir /mnt/hdd/PANO/results/2023-07-01-075136 \
        --dataset_name pano_debug  \
        --prediction_name instances_predictions.postprocessed.pth \
        --do_visualize \
        --visualizer_dir visualize.postprocessed \
        --nodo_coco \
        --coco_annotator_url http://192.168.0.79:5000/api