def unet_learner(
data: DataBunch,
arch: Callable,
pretrained: bool = True,
blur_final: bool = True,
norm_type: Optional[NormType] = NormType,
split_on: Optional[SplitFuncOrIdxList] = None,
blur: bool = False,
self_attention: bool = False,
y_range: Optional[Tuple[float, float]] = None,
last_cross: bool = True,
bottle: bool = False,
cut: Union[int, Callable] = None,
hypercolumns=True,
**learn_kwargs: Any,
) -> Learner:
"Build Unet learner from `data` and `arch`."
meta = cnn_config(arch)
body = create_body(arch, pretrained, cut)
M = DynamicUnet_Hcolumns if hypercolumns else DynamicUnet
model = to_device(
M(
body,
n_classes=data.c,
blur=blur,
blur_final=blur_final,
self_attention=self_attention,
y_range=y_range,
norm_type=norm_type,
last_cross=last_cross,
bottle=bottle,
),
data.device,
)
learn = Learner(data, model, **learn_kwargs)
learn.split(ifnone(split_on, meta["split"]))
if pretrained:
learn.freeze()
apply_init(model[2], nn.init.kaiming_normal_)
return learn
def unet_learner_wide(data: DataBunch,
arch: Callable,
pretrained: bool = True,
blur_final: bool = True,
norm_type: Optional[NormType] = NormType,
split_on: Optional[SplitFuncOrIdxList] = None,
blur: bool = False,
self_attention: bool = False,
y_range: Optional[Tuple[float, float]] = None,
last_cross: bool = True,
bottle: bool = False,
nf_factor: int = 1,
**kwargs: Any) -> Learner:
"Build Unet learner from `data` and `arch`."
meta = cnn_config(arch)
body = create_body(arch, pretrained)
# can tell to go to another gpu
model = to_device(
DynamicUnetWide(
body,
n_classes=data.c,
blur=blur,
blur_final=blur_final,
self_attention=self_attention,
y_range=y_range,
norm_type=norm_type,
last_cross=last_cross,
bottle=bottle,
nf_factor=nf_factor,
),
data.device,
)
learn = Learner(data, model, **kwargs)
learn.split(ifnone(split_on, meta['split']))
if pretrained:
learn.freeze()
apply_init(model[2], nn.init.kaiming_normal_)
return learn