def simple_mil(extractor, channels, freeze=True):
cnn_to_bw(extractor, IMGNET_MEAN, IMGNET_STD)
if freeze:
for param in extractor.parameters():
param.requires_grad = False
mil_scores = nn.Sequential(nn.Conv2d(channels, 1, 1, 1, 0))
return PretrainedMIL(extractor, mil_scores)
def simple_pretrained(extractor, channels, features_shape, freeze=True):
cnn_to_bw(extractor, IMGNET_MEAN, IMGNET_STD)
if freeze:
for param in extractor.parameters():
param.requires_grad = False
predict = nn.Sequential(nn.Conv2d(channels, 1, features_shape, 1, 0),
nn.Sigmoid())
return Pretrained(extractor, predict)
def simple_slicerMIL(extractor, feature_dims, freeze = True):
cnn_to_bw(extractor, IMGNET_MEAN, IMGNET_STD)
if freeze:
for param in extractor.parameters():
param.requires_grad = False
mil_scores = nn.Sequential(
nn.Conv3d(feature_dims, 1, 1, 1, 0))
return PretrainedSlicerMIL(extractor, mil_scores)
def __init__(self, extractor):
cnn_to_bw(extractor, IMGNET_MEAN, IMGNET_STD)
for param in extractor.parameters():
param.requires_grad = False
self.extractor = extractor
if torch.cuda.is_available():
self.extractor = self.extractor.cuda()
self.tree = None
def simple_slicer(extractor, feature_dims, feature_size, freeze = True):
cnn_to_bw(extractor, IMGNET_MEAN, IMGNET_STD)
if freeze:
for param in extractor.parameters():
param.requires_grad = False
predict = nn.Sequential(
nn.Conv3d(feature_dims, 1, feature_size, 1, 0),
nn.Sigmoid())
return PretrainedSlicer(extractor, predict)