Check if Everything is Installed Correctly – Run a Pre-trained Detectron2 Model
Before you start training, it’s a good idea to check that everything is working properly. The best way to do this is to perform inference using a pre-trained model.
First, we read the image & display the image on which we are going to perform inferencing
!wget http://images.cocodataset.org/val2017/000000439715.jpg -q -O input.jpg
image = cv2.imread("./input.jpg")
cv2_imshow(image)
Visualize Predictions
Get a image with Bounding Boxes, Masks Applied & Class predictions
visualizer = Visualizer(image[:, :, ::-1], MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), scale=1.2)
out = visualizer.draw_instance_predictions(outputs["instances"].to("cpu"))
cv2_imshow(out.get_image()[:, :, ::-1])
Check if Pre-trained Detectron2 Model is Working
We are going to check if the downloaded Pre-trained Detectron2 Model is working. Before training the model on our dataset we will do a quick check by performing inferencing on a image & see if the model predicts the masks, bounding boxes & classes correctly
Below we download a image from Coco dataset & display it
!wget http://images.cocodataset.org/val2017/000000439715.jpg -q -O input.jpg
image = cv2.imread("./input.jpg")
cv2_imshow(image)
In the code snippet below we do the following:
- cfg – Instantiate a configuration object, this holds the model parameters used for training
- merge_from_file – Populate the configuration object from a yaml file which holds the parameters for the model
- ROI_HEADS.SCORE_THRESH_TEST – Set this to 0.5 or 50%, which tells the model to only consider predictions of 0.5 & above threshold
- WEIGHTS – Check the trained checkpoint i.e., weights to be used for inferencing
- DefaultPredictor – Instantiate a predictor object by passing the configuration object to predictor
- predictor(image) – Pass the image to predictor & get the inferencing outputs
cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")
predictor = DefaultPredictor(cfg)
outputs = predictor(image)
Verify if prediction is done correctly by printing tensors which contain the outputs for Predicted Classes & Bounding boxes
print(outputs["instances"].pred_classes)
print(outputs["instances"].pred_boxes)
Visualize Training Dataset Image
Lets take a look at one of the training images along with the mask, bounding box & class displayed on the image
metadata = MetadataCatalog.get(TRAIN_DATA_SET_NAME)
dataset_train = DatasetCatalog.get(TRAIN_DATA_SET_NAME)
dataset_entry = dataset_train[0]
image = cv2.imread(dataset_entry[“file_name”])
visualizer = Visualizer(
image[:, :, ::-1],
metadata=metadata,
scale=0.8,
instance_mode=ColorMode.IMAGE_BW
)
out = visualizer.draw_dataset_dict(dataset_entry)
cv2_imshow(out.get_image()[:, :, ::-1])
Visualize & Evaluate the Trained Model
In the below code we loop through each image file in validation dataset, pass the image to model to make prediction & then visualize the predicted image to check if the predictions are accurate
dataset_valid = DatasetCatalog.get(VALID_DATA_SET_NAME)
for d in dataset_valid:
img = cv2.imread(d[“file_name”])
outputs = predictor(img)
visualizer = Visualizer(
img[:, :, ::-1],
metadata=metadata,
scale=0.8,
instance_mode=ColorMode.IMAGE_BW
)
out = visualizer.draw_instance_predictions(outputs[“instances”].to(“cpu”))
cv2_imshow(out.get_image()[:, :, ::-1])
