hubconf.py

dependencies = ["efficientnet_pytorch", "pretrainedmodels",
                "timm", "torch", "torchvision"]
import torch
from utils.utils import Params
from backbone import HybridNetsBackbone
from pathlib import Path
import os


def hybridnets(pretrained=True, backbone=None, compound_coef=3, device='cpu'):
    """Creates a HybridNets model

    Arguments:
        pretrained (bool): load pretrained weights into the model
        backbone (str): use timm to create another backbone replacing efficientnet
        compound_coef (int): compound coefficient of efficientnet backbone
        device (str): 'cuda:0' or 'cpu'

    Returns:
        HybridNets model
    """
    params = Params(os.path.join(Path(__file__).resolve().parent, "projects/bdd100k.yml"))
    model = HybridNetsBackbone(num_classes=len(params.obj_list), compound_coef=compound_coef,
                               ratios=eval(params.anchors_ratios), scales=eval(params.anchors_scales),
                               seg_classes=len(params.seg_list), backbone_name=backbone)
    if pretrained and not backbone and compound_coef == 3:
        weight_url = 'https://github.com/datvuthanh/HybridNets/releases/download/v1.0/hybridnets.pth'
        model.load_state_dict(torch.hub.load_state_dict_from_url(weight_url, map_location=device))
    model = model.to(device)
    return model


if __name__ == "__main__":
    model = hybridnets(device='cpu')
    img = torch.rand(1, 3, 384, 640)

    result = model(img)
    print(result)