/Capstone-Project---Dogs-vs-Cats-by-COCO

Primary LanguageJupyter Notebook

Coco Detectron2 Dogs 🐶 and Cats 🐱

Train a detectron model on a custom dataset (Coco formated).

Introduction

  • We use catvsdogs dataset which only has two classes: cats and dogs. We'll train a segmentation model from an existing model pre-trained on COCO dataset, available in detectron2's model zoo.
  • Following the steps below you can create your dataset and train your instance segmentation model on detectron2.

How To Create Custom Dataset

  • First, you need to get a labelling tool such as Labelme. You can insall labelme using the link.
  • Every individual object in an image must be classified and identified each instance of an object within an image. More information : Image Segmentation
  • Dataset folders must be like in the following format. Example Dataset
.
├── train/
│   ├── images
└── evaluation/
        ├── images

How To Convert Labelme Annotation Files to COCO Format

Run the labelme2coco.py script to generate a COCO data format file for both train and validation dataset as below.

python labelme2coco.py {path_to_train_folder} --output {path_to_train_folder/via_region_data.json}

python labelme2coco.py {path_to_validation_folder} --output {path_to_validation_folder/via_region_data.json}
  • Folder structure after conversion.
.
├── train/
│   ├── images
│   ├── via_region_data.json
└── evaluation/
        ├── images   
        ├── via_region_data.json

How To Run Model Train

  • Zip your dataset containing folder formatted like above.
  • Upload the jupyter notebook to Google Colab.
  • Upload your dataset to Google Colab.
  • Update the code snippet below according to the dataset you've been created.
from detectron2.data.datasets import register_coco_instances
register_coco_instances("YourTrainDatasetName", {},"path to train.json", "path to train image folder")
register_coco_instances("YourTestDatasetName", {}, "path to test.json", "path to test image folder")
  • An example code:
from detectron2.data.datasets import register_coco_instances
register_coco_instances("train_dataset", {}, "/content/catsvsdogs/train/via_region_data.json", "/content/catsvsdogs/train")
register_coco_instances("val_dataset", {}, "/content/catsvsdogs/validation/via_region_data.json", "/content/catsvsdogs/validation")
  • Before you starting the training, please make sure to update value of cfg.MODEL.ROI_HEADS.NUM_CLASSES = [Update Here] is the same as number of classes in the dataset.
from detectron2.engine import DefaultTrainer

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.DATASETS.TRAIN = ("train_dataset",)
cfg.DATASETS.TEST = ("val_dataset",)
cfg.DATALOADER.NUM_WORKERS = 2
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")  # Let training initialize from model zoo
cfg.SOLVER.IMS_PER_BATCH = 2  # This is the real "batch size" commonly known to deep learning people
cfg.SOLVER.BASE_LR = 0.00025  # pick a good LR
cfg.SOLVER.MAX_ITER = 300    # 300 iterations seems good enough for this toy dataset; you will need to train longer for a practical dataset
cfg.SOLVER.STEPS = []        # do not decay learning rate
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 128   # The "RoIHead batch size". 128 is faster, and good enough for this toy dataset (default: 512)
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 2  # only has two class (cats and dogs). (see https://detectron2.readthedocs.io/tutorials/datasets.html#update-the-config-for-new-datasets)
# NOTE: this config means the number of classes, but a few popular unofficial tutorials incorrect uses num_classes+1 here.
  • Run the notebook.