def create_cnn(data, arch, pretrained=False, is_mono_input=True, **kwargs):
meta = cnn_config(arch)
body = create_body(arch, pretrained)
# sum up the weights of in_channels axis, to reduce to single input channel
# Suggestion by David Gutman
# https://forums.fast.ai/t/black-and-white-images-on-vgg16/2479/2
if is_mono_input:
first_conv_layer = body[0]
first_conv_weights = first_conv_layer.state_dict()['weight']
assert first_conv_weights.size(1) == 3 # RGB channels dim
summed_weights = torch.sum(first_conv_weights, dim=1, keepdim=True)
first_conv_layer.weight.data = summed_weights
first_conv_layer.in_channels = 1
else:
# In this case, the input is a stereo
first_conv_layer = body[0]
first_conv_weights = first_conv_layer.state_dict()['weight']
assert first_conv_weights.size(1) == 3 # RGB channels dim
summed_weights = torch.sum(first_conv_weights, dim=1, keepdim=True)
first_conv_layer.weight.data = first_conv_weights[:, :
2, :, :] # Keep only 2 channels for the weights
first_conv_layer.in_channels = 2
nf = num_features_model(body) * 2
head = create_head(nf, data.c, None, 0.5)
model = nn.Sequential(body, head)
learn = Learner(data, model, **kwargs)
learn.split(meta['split'])
if pretrained:
learn.freeze()
apply_init(model[1], nn.init.kaiming_normal_)
return learn
def __init__(self,
bu,
instructor,
td_c=1,
bu_c=0,
lateral=laterals.ConvAddLateral,
td_out_lateral=None,
embedding=fv.embedding,
img_size: Tuple[int, int] = (256, 256)):
super().__init__()
self.ifn, self.bu_body, self.td, self.td_head, self.laterals, channels = create_bu_td(
bu, td_head=td_c, lateral=lateral, img_size=img_size)
if td_out_lateral:
self.laterls[-1].remove()
hm_lat = td_out_lateral(self.td[-1], self.bu_body[0], td_c,
channels[0])
self.laterls[-1] = hm_lat
self.emb = embedding(instructor.n_inst,
channels[-1]) if embedding else None
self.bu_head = fv.create_head(channels[-1] * 2, bu_c) if bu_c else None
self.instructor = instructor
self.instructor.on_init_end(self)
def __init__(self, encoder=models.resnet18, s_out=512):
# TODO warn is s_out is to large
super().__init__()
self.body = create_body(encoder, cut=-2)
self.head = create_head(1024, 1, [s_out])[:5]
def resnet(classes):
base_model = models.resnet50(pretrained=True)
body = nn.Sequential(*list(base_model.children())[:-2])
nf = num_features_model(body) * 2
head = create_head(nf, classes, None, ps=0.5, bn_final=False)
return nn.Sequential(body, head, nn.LogSoftmax())