def create_model():
from elektronn2 import neuromancer as nm
in_sh = (None, 1, 22, 140, 140)
inp = nm.Input(in_sh, 'b,f,z,x,y', name='raw') # high res
# Convolution, downsampling of intermediate features
conv0 = nm.Conv(inp, 20, (1, 3, 3))
conv1 = nm.Conv(conv0, 20, (1, 3, 3))
down0 = nm.Pool(conv1, (1, 2, 2), mode='max') # mid res
conv2 = nm.Conv(down0, 30, (1, 3, 3))
conv3 = nm.Conv(conv2, 30, (1, 3, 3))
down1 = nm.Pool(conv3, (1, 2, 2), mode='max') # low res
conv4 = nm.Conv(down1, 35, (1, 3, 3))
conv5 = nm.Conv(conv4, 35, (1, 3, 3))
down2 = nm.Pool(conv5, (1, 2, 2), mode='max') # very low res
conv6 = nm.Conv(down2, 42, (3, 3, 3))
down2b = nm.Pool(conv6, (1, 2, 2), mode='max') # very low res, even lower
conv7 = nm.Conv(down2b, 42, (3, 3, 3))
# Merging very low-res features with low-res features
mrg0 = nm.UpConvMerge(conv5, conv7, 45)
mconv0 = nm.Conv(mrg0, 42, (1, 3, 3))
mconv1 = nm.Conv(mconv0, 42, (1, 3, 3))
# Merging low-res with mid-res features
mrg1 = nm.UpConvMerge(conv3, mconv1, 42)
mconv2 = nm.Conv(mrg1, 35, (3, 3, 3))
mconv3 = nm.Conv(mconv2, 35, (3, 3, 3))
# Merging mid-res with high-res features
mrg2 = nm.UpConvMerge(conv1, mconv3, 30)
mconv4 = nm.Conv(mrg2, 20, (3, 3, 3))
mconv5 = nm.Conv(mconv4, 20, (3, 3, 3))
barr = nm.Conv(mconv5, 2, (1, 1, 1), activation_func='lin', name='barr')
probs = nm.Softmax(barr)
target = nm.Input_like(mconv5, override_f=1, name='target')
loss_pix = nm.MultinoulliNLL(probs,
target,
target_is_sparse=True,
name='nll_barr')
loss = nm.AggregateLoss(loss_pix, name='loss')
errors = nm.Errors(probs, target, target_is_sparse=True)
model = nm.model_manager.getmodel()
model.designate_nodes(input_node=inp,
target_node=target,
loss_node=loss,
prediction_node=probs,
prediction_ext=[loss, errors, probs])
return model
#data is in format: # raw data=(1,16,128,128) as NP array # affinities=(3,16,128,128) as NP array ################### # BUILD GRAPH # ################### in_sh = (None, 1, 16, 128, 128) inp = nm.Input(in_sh, 'b,f,z,x,y', name='raw') # high res # Convolution, downsampling of intermediate features conv0 = nm.Conv(inp, 8, (15, 15, 15), (1, 1, 1), "same", name="c0") down0 = nm.Pool(conv0, (1, 4, 4), mode="max", name="d0") #full conv1 = nm.Conv(down0, 32, (15, 15, 15), (1, 1, 1), "same", name="c1") down1 = nm.Pool(conv1, (1, 2, 2), mode="max", name="d1") #high conv2 = nm.Conv(down1, 64, (15, 15, 15), (1, 1, 1), "same", name="c2") #mid down2 = nm.Pool(conv2, (1, 2, 2), mode="max", name="d2") #high conv3 = nm.Conv(down2, 64, (15, 15, 15), (1, 1, 1), "same", name="c3") #low #merge C2 and C3 and convolve mrg0 = nm.UpConvMerge(conv2, conv3, 64, name="m0", merge_mode="add") mconv0 = nm.Conv(mrg0, 32, (15, 15, 15), (1, 1, 1), "same", name="mc0") #merge mc0 and c1 and convolve mrg1 = nm.UpConvMerge(mconv0, conv1, 32, name="m1", merge_mode="add")
def create_model():
from elektronn2 import neuromancer
import theano.tensor as T
import numpy as np
in_sh = (None,1,572-32*9,572-32*9)
img = neuromancer.Input(in_sh, 'b,f,x,y', name='raw')
out0 = neuromancer.Conv(img, 64, (3,3), (1,1))
out1 = neuromancer.Conv(out0, 64, (3,3), (1,1))
out2 = neuromancer.Pool(out1, (2,2))
out3 = neuromancer.Conv(out2, 128, (3,3), (1,1))
out4 = neuromancer.Conv(out3, 128, (3,3), (1,1))
out5 = neuromancer.Pool(out4, (2,2))
out6 = neuromancer.Conv(out5, 256, (3,3), (1,1))
out7 = neuromancer.Conv(out6, 256, (3,3), (1,1))
out8 = neuromancer.Pool(out7, (2,2))
out9 = neuromancer.Conv(out8, 512, (3,3), (1,1))
out10 = neuromancer.Conv(out9, 512, (3,3), (1,1))
out11 = neuromancer.Pool(out10, (2,2))
out12 = neuromancer.Conv(out11, 1024, (3,3), (1,1))
out13 = neuromancer.Conv(out12, 1024, (3,3), (1,1))
out14 = neuromancer.Pool(out13, (2,2))
####
up0 = neuromancer.UpConvMerge(out10, out14, 1024)
up1 = neuromancer.Conv(up0, 512, (3,3), (1,1))
up2 = neuromancer.Conv(up1, 512, (3,3), (1,1))
up3 = neuromancer.UpConvMerge(out7, up2, 512)
up4 = neuromancer.Conv(up3, 256, (3,3), (1,1))
up5 = neuromancer.Conv(up4, 256, (3,3), (1,1))
up6 = neuromancer.UpConvMerge(out4, up5, 256)
up7 = neuromancer.Conv(up6, 128, (3,3), (1,1))
up8 = neuromancer.Conv(up7, 128, (3,3), (1,1))
up9 = neuromancer.UpConvMerge(out1, up8, 128)
up10 = neuromancer.Conv(up9, 64, (3,3), (1,1))
top_feat = neuromancer.Conv(up10, 64, (3,3), (1,1))
# Target outputs
barr_out = neuromancer.Conv(top_feat, 3, (1,1), (1,1), activation_func='lin', name='barr')
obj_out = neuromancer.Conv(top_feat, 4, (1,1), (1,1), activation_func='lin', name='obj')
my_out = neuromancer.Conv(top_feat, 3, (1,1), (1,1), activation_func='lin', name='my')
barr_out = neuromancer.Softmax(barr_out)
obj_out = neuromancer.Softmax(obj_out)
my_out = neuromancer.Softmax(my_out)
target = neuromancer.Input_like(top_feat, dtype='int16', override_f=3, name='target')
barr, obj, my = neuromancer.split(target, 'f', n_out=3, name=['barr_t', 'obj_t', 'my_t'])
# Target loss
barr_loss_pix = neuromancer.MultinoulliNLL(barr_out, barr, target_is_sparse=True,name='nll_barr')
obj_loss_pix = neuromancer.MultinoulliNLL(obj_out, obj, target_is_sparse=True, name='nll_obj')
my_loss_pix = neuromancer.MultinoulliNLL(my_out, my, target_is_sparse=True, name='nll_my')
pred = neuromancer.Concat([barr_out, obj_out, my_out], axis='f')
pred.feature_names = ['barrier_bg', 'barr_mem', 'barr_ecs', 'obj_bg',
'obj_mito', 'obj_ves', 'obj_syn', 'my_bg', 'my_out', 'my_in']
# Objective
weights = np.array([2.154, 0.42, 0.42])
weights *= len(weights) / weights.sum()
loss = neuromancer.AggregateLoss([barr_loss_pix,
obj_loss_pix,
my_loss_pix],
mixing_weights=weights)
# Monitoring / Debug outputs
nll_barr = neuromancer.ApplyFunc(barr_loss_pix, T.mean, name='mnll_barr')
nll_obj = neuromancer.ApplyFunc(obj_loss_pix, T.mean, name='mnll_obj')
nll_my = neuromancer.ApplyFunc(my_loss_pix, T.mean, name='mnll_my')
errors = neuromancer.Errors(barr_out, barr, target_is_sparse=True)
model = neuromancer.model_manager.getmodel()
model.designate_nodes(input_node=img, target_node=target, loss_node=loss,
prediction_node=pred,
prediction_ext=[loss, errors, pred],
debug_outputs =[nll_barr, errors, nll_obj, nll_my])
return model
def create_model():
from elektronn2 import neuromancer as nm
trainee = trainee_dict["create_model"]()
inp = trainee.input_node
trainee_gt = trainee.target_node
trainee_out = trainee.prediction_node
trainee_loss = trainee.loss_node
adv_input, adv_target = nm.advmerge(trainee_out, trainee_gt)
# raw data
diff = np.array(inp.shape.spatial_shape, dtype=np.int32) - np.array(
trainee_out.shape.spatial_shape, dtype=np.int32)
assert not np.any(diff % 2)
raw_inp = nm.Crop(inp, list((diff // 2)))
conv0 = nm.Conv(raw_inp,
5, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
conv1 = nm.Conv(conv0,
10, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
down0 = nm.Pool(conv1, (1, 2, 2), mode='max') # mid res
conv2 = nm.Conv(down0,
10, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
conv3 = nm.Conv(conv2,
15, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
# Merging low-res with mid-res features
mrg1 = nm.UpConvMerge(conv1, conv3, 20, name="upconv_adv")
mconv2 = nm.Conv(mrg1,
15, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
raw_out = nm.Conv(mconv2,
15, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
# segmentation
conv0 = nm.Conv(adv_input,
5, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
conv1 = nm.Conv(conv0,
10, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
down0 = nm.Pool(conv1, (1, 2, 2), mode='max') # mid res
conv2 = nm.Conv(down0,
10, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
conv3 = nm.Conv(conv2,
15, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
# Merging low-res with mid-res features
mrg1 = nm.UpConvMerge(conv1, conv3, 20, name="upconv_adv")
mconv2 = nm.Conv(mrg1,
15, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
seg_out = nm.Conv(mconv2,
15, (2, 3, 3),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
out = nm.Concat([raw_out, seg_out], axis="f")
out = nm.Conv(out,
20, (2, 5, 5),
pool_shape=(1, 2, 2),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
out = nm.Conv(out,
40, (2, 3, 3),
pool_shape=(1, 2, 2),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
out = nm.Conv(out,
60, (2, 2, 2),
pool_shape=(1, 2, 2),
activation_func=act,
dropout_rate=dr,
name="conv_adv")
dec = nm.Perceptron(out,
2,
flatten=True,
activation_func='lin',
name="perc_adv")
adv_out = nm.Softmax(dec)
# as in orig. GAN paper of Godfellow et al. only the positive label is taken
# into account (i.e. adversarial network predicts input is ground truth)
# for the adv. prediction loss when training the trainee
# if training the adv. network only the binary cross-entropy of the adv.
# prediction is backpropagated.
adv_loss = nm.MultinoulliNLL(adv_out,
adv_target,
target_is_sparse=True,
name='nll_adversarial',
class_weights=None)
loss = nm.AggregateLoss([adv_loss, trainee_loss],
mixing_weights=None,
name='loss_adversarial')
model = nm.model_manager.getmodel()
model.designate_nodes(
input_node=inp,
target_node=trainee.target_node,
loss_node=loss,
prediction_node=trainee.prediction_node,
prediction_ext=trainee.prediction_ext,
)
return model