metadata
license: apache-2.0
tags:
- Bounding box detection
- PyTorch
This repository contains code that you can use to train or load Faster R-CNN models in half mode easily.
Below is an example of how to load pretrained weights in half mode.
import numpy as np
from PIL import Image
import json
from frcnn.visualizing_image import SingleImageViz
from frcnn.processing_image import Preprocess
from frcnn.modeling_frcnn import GeneralizedRCNN
from frcnn.utils import Config, decode_image
max_detections = 36
frcnn_config = json.load(open("frcnn/config.jsonl"))
frcnn_config = Config(frcnn_config)
image_preprocessor= Preprocess(frcnn_config).half().cuda()
box_segmentation_model= GeneralizedRCNN.from_pretrained("unc-nlp/frcnn-vg-finetuned", frcnn_config).half().cuda()
img_url = 'image.png'
raw_image = Image.open(img_url).convert('RGB')
frcnn_output = decode_image(np.asarray(raw_image), box_segmentation_model, image_preprocessor, max_detections=max_detections)