def mergecrop(self, run_shape, run_shape_b_index, bottom_a, bottom_b):
run_shape_a = run_shape[-1]
run_shape_b = run_shape[run_shape_b_index]
# Shape update rules
update = RunShapeUpdater()
update.fmaps_update = lambda x: run_shape_a.fmaps + run_shape_b.fmaps
self.update_shape(run_shape, update)
return L.MergeCrop(bottom_a, bottom_b, forward=[1, 1], backward=[1, 1])
def mergecrop(run_shape, bottom_a, bottom_b):
# Shape update rules
update = [lambda x: 0, lambda x: 0, lambda x: 2 * x]
update += [[lambda x: x, lambda x: x, lambda x: x]]
update += [[lambda x, i=i: x for i in range(0, len(run_shape[-1][4]))]]
update_shape(run_shape, update)
return L.MergeCrop(bottom_a, bottom_b, forward=[1, 1], backward=[1, 1])
def mergecrop(self, run_shape, run_shape_b_index, bottom_a, bottom_b, op = 'stack'):
run_shape_a = run_shape[-1]
run_shape_b = run_shape[run_shape_b_index]
# Shape update rules
update = RunShapeUpdater()
if (op == 'stack'):
update.fmaps_update = lambda x: run_shape_a.fmaps + run_shape_b.fmaps
else:
update.fmaps_update = lambda x: run_shape_a.fmaps
self.update_shape(run_shape, update)
return L.MergeCrop(bottom_a, bottom_b, forward=[1,1], backward=[1,1], operation=(0 if (op == 'stack') else 1))
def mergecrop(bottom_a, bottom_b, op = 'stack'):
return L.MergeCrop(bottom_a, bottom_b, forward=[1,1], backward=[1,1], operation=(0 if (op == 'stack') else 1))
def approximation(args):
print('creating approximation...')
# Start a network
net = Unet3D.start(args)
#UNET config
net_depth = 3
net_width = 2
base_filters = 24
channels_in = 1
channels_out = base_filters
# ReLU negative slope
relu_slope = 0.005
strategy = [2, 2, 2]
xs = [net.data]
if args.nstreams > 1:
xs = L.Split(net.data, ntop=args.nstreams)
print('xs:', len(xs))
skip = []
# contracting part
for d in range(net_depth):
print('upsampling...', d)
channels_out = base_filters * (3**d)
x = Unet3D.vgg_block_approximation(args, xs, channels_in,
channels_out)
skip.append(x)
# maxpool
x = L.Pooling(x,
pool=P.Pooling.MAX,
kernel_size=strategy,
stride=strategy,
pad=[0],
dilation=[1])
channels_in = channels_out
xs = [x for _ in range(args.nstreams)]
# bridge
channels_out = base_filters * (3**net_depth)
x = Unet3D.vgg_block_approximation(args, xs, channels_in, channels_out)
# expanding
channels_in = channels_out
for d in reversed(range(net_depth)):
print('downsampling...', d)
x = L.Deconvolution(x,
convolution_param=dict(
num_output=channels_out,
kernel_size=strategy,
stride=strategy,
pad=[0],
group=1,
dilation=[1],
bias_term=False,
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant')))
channels_out = base_filters * (3**(d))
x = L.Convolution(x,
kernel_size=[1],
stride=[1],
dilation=[1],
num_output=channels_out,
pad=[0],
param=[dict(lr_mult=1.0),
dict(lr_mult=2.0)],
weight_filler=dict(type='msra'),
bias_filler=dict(type='constant'))
x = L.MergeCrop(x,
skip[d],
forward=[1, 1],
backward=[1, 1],
operation=0)
xs = [x for _ in range(args.nstreams)]
x = Unet3D.vgg_block_approximation(args, xs, channels_in,
channels_out)
net = Unet3D.end(net, x)
protonet = net.to_proto()
protonet.name = 'net'
# Store the network as prototxt
with open(protonet.name + '.prototxt', 'w') as f:
print(protonet, file=f)