arch

The arch config group determines the model architecture used by mml. Currently there are two backbone libraries: timm and segmentation-models-pytorch. The former supports classification and regression tasks while the latter supports classification and segmentation tasks.

Default config option is arch=timm. Call arch=smp for segmentation models.

timm

name

default: resnet34

• the identifier of the model architecture in the timm library - to be passed to timm.create_model(model_name=…)

• full list of models can be found here: https://huggingface.co/docs/timm/models

• also note the timm.list_models() option (https://huggingface.co/docs/timm/reference/models#timm.list_models)

pretrained

default: true

• boolean to indicate whether the model should be initialized with pretraiend weights (timm.create_model(pretrained=…)

• results files for pretrained weights: https://github.com/huggingface/pytorch-image-models/tree/main/results

drop_rate

default: 0.

• dropout rate, set to some value within [0, 1), will be used to initialize a torch.nn.Dropout layer at the beginning of each classifier head

smp

arch

default: unet

• segmentation architecture, passed to smp.create_model(arch=…)

• options are [unet, unetplusplus, manet, linknet, fpn, pspnet, deeplabv3, deeplapv3plus, pan]

encoder

default: tu-resnet34

• the encoder backbone to use, passed to smp.create_model(encoder_name=…)

• for a full list of supported backbones, see https://smp.readthedocs.io/en/latest/encoders.html

• the <<tu->> prefix automatically tries to use a timm encoder, see https://smp.readthedocs.io/en/latest/encoders_timm.html

weights

default: imagenet

• weight initialization of the backbone, passed to smp.create_model(encoder_weights=…)

• supported weights per backbone are found here: https://smp.readthedocs.io/en/latest/encoders.html