mednet.config.models.densenet_pretrained

DenseNet, to be fine-tuned. Pre-trained on ImageNet.

This configuration contains a version of DenseNet (c.f. TorchVision’s page <alexnet_pytorch_>), modified for a variable number of outputs (defaults to 1).

N.B.: The output layer is always initialized from scratch.

from torch.nn import BCEWithLogitsLoss
from torch.optim import Adam

from mednet.data.augmentations import ElasticDeformation
from mednet.models.densenet import Densenet

model = Densenet(
    train_loss=BCEWithLogitsLoss(),
    validation_loss=BCEWithLogitsLoss(),
    optimizer_type=Adam,
    optimizer_arguments=dict(lr=0.0001),
    augmentation_transforms=[ElasticDeformation(p=0.2)],
    pretrained=True,
    dropout=0.1,
)