def inference_net():
input_shape = (91, 862, 862)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = unet.unet(
raw_bc,
12,
6,
[[1, 3, 3], [1, 3, 3], [3, 3, 3]],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(10, 1, 1),
fov=(10, 1, 1),
)
logits_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=2,
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = logits_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:] # strip the channel dimension
probabilities = tf.reshape(tf.nn.softmax(logits_bc, dim=1)[0], output_shape_c)
predictions = tf.argmax(probabilities, axis=0)
print(probabilities.name)
tf.train.export_meta_graph(filename="unet_inference.meta")
def inference_net():
input_shape = (91, 862, 862)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = autoencoder.autoencoder(
raw_bc,
12,
6,
[[1, 3, 3], [1, 3, 3], [3, 3, 3]],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(10, 1, 1),
fov=(10, 1, 1),
)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=1,
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:]
output_shape = output_shape_c[1:]
dist = tf.reshape(dist_bc, output_shape)
tf.train.export_meta_graph(filename="autoencoder_inference.meta")
def inference_net(labels):
input_shape = (340, 340, 340)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = unet.unet(
raw_bc,
[12, 12 * 6, 12 * 6 * 6, 12 * 6 * 6 * 6],
[48, 12 * 6, 12 * 6 * 6, 12 * 6 * 6 * 6],
[[2, 2, 2], [2, 2, 2], [3, 3, 3]],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(1, 1, 1),
fov=(1, 1, 1),
)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=len(labels),
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:]
dist_c = tf.reshape(dist_bc, output_shape_c)
network_outputs = tf.unstack(dist_c, len(labels), axis=0)
tf.train.export_meta_graph(filename="unet_inference.meta")
def inference_net():
input_shape = (400, 400, 400)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = unet.unet(
raw_bc,
12,
6,
[[2, 2, 2], [2, 2, 2], [3, 3, 3]],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(1, 1, 1),
fov=(1, 1, 1),
)
pred_raw_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=1,
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = pred_raw_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:]
output_shape = output_shape_c[1:]
pred_raw = tf.reshape(pred_raw_bc, output_shape)
tf.train.export_meta_graph(filename="unet_inference.meta")
def train_net():
# z [1, 1, 1]: 66 -> 38 -> 10
# y, x [2, 2, 2]: 228 -> 140 -> 52
shape_0 = (220, ) * 3
shape_1 = (132, ) * 3
shape_2 = (44, ) * 3
affs_0_bc = tf.ones((1, 3) + shape_0) * 0.5
with tf.variable_scope("autocontext") as scope:
# phase 1
raw_0 = tf.placeholder(tf.float32, shape=shape_0)
raw_0_bc = tf.reshape(raw_0, (1, 1) + shape_0)
input_0 = tf.concat([raw_0_bc, affs_0_bc], 1)
out_bc, fov, anisotropy = unet.unet(
input_0,
24,
3,
[[2, 2, 2], [2, 2, 2], [2, 2, 2]],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
)
affs_1_bc, fov = ops3d.conv_pass(
out_bc,
kernel_size=[[1, 1, 1]],
num_fmaps=3,
activation="sigmoid",
fov=fov,
voxel_size=anisotropy,
)
affs_1_c = tf.reshape(affs_1_bc, (3, ) + shape_1)
gt_affs_1_c = tf.placeholder(tf.float32, shape=(3, ) + shape_1)
loss_weights_1_c = tf.placeholder(tf.float32, shape=(3, ) + shape_1)
loss_1 = tf.losses.mean_squared_error(gt_affs_1_c, affs_1_c,
loss_weights_1_c)
# phase 2
tf.summary.scalar("loss_pred0", loss_1)
scope.reuse_variables()
raw_1 = ops3d.center_crop(raw_0, shape_1)
raw_1_bc = tf.reshape(raw_1, (1, 1) + shape_1)
input_1 = tf.concat([raw_1_bc, affs_1_bc], 1)
out_bc, fov, anisotropy = unet.unet(
input_1,
24,
3,
[[2, 2, 2], [2, 2, 2], [2, 2, 2]],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
fov=fov,
voxel_size=anisotropy,
)
affs_2_bc, fov = ops3d.conv_pass(
out_bc,
kernel_size=[[1, 1, 1]],
num_fmaps=3,
activation="sigmoid",
fov=fov,
voxel_size=anisotropy,
)
affs_2_c = tf.reshape(affs_2_bc, (3, ) + shape_2)
gt_affs_2_c = ops3d.center_crop(gt_affs_1_c, (3, ) + shape_2)
loss_weights_2_c = ops3d.center_crop(loss_weights_1_c, (3, ) + shape_2)
loss_2 = tf.losses.mean_squared_error(gt_affs_2_c, affs_2_c,
loss_weights_2_c)
tf.summary.scalar("loss_pred1", loss_2)
loss = loss_1 + loss_2
tf.summary.scalar("loss_total", loss)
tf.summary.scalar("loss_diff", loss_1 - loss_2)
for trainable in tf.trainable_variables():
custom_ops.tf_var_summary(trainable)
merged = tf.summary.merge_all()
opt = tf.train.AdamOptimizer(learning_rate=0.5e-4,
beta1=0.95,
beta2=0.999,
epsilon=1e-8)
optimizer = opt.minimize(loss)
tf.train.export_meta_graph(filename="wnet.meta")
names = {
"raw": raw_0.name,
"affs_1": affs_1_c.name,
"affs_2": affs_2_c.name,
"gt_affs": gt_affs_1_c.name,
"loss_weights": loss_weights_1_c.name,
"loss": loss.name,
"optimizer": optimizer.name,
"summary": merged.name,
}
with open("net_io_names.json", "w") as f:
json.dump(names, f)
def make_net(labels, added_steps, mode="train", loss_name="loss_total"):
unet0 = scale_net.SerialUNet(
[12, 12 * 6, 12 * 6**2],
[48, 12 * 6, 12 * 6**2],
[(3, 3, 3), (3, 3, 3)],
[[(3, 3, 3), (3, 3, 3)], [(3, 3, 3), (3, 3, 3)], [(3, 3, 3),
(3, 3, 3)]],
[[(3, 3, 3), (3, 3, 3)], [(3, 3, 3), (3, 3, 3)]],
input_voxel_size=(4, 4, 4),
)
unet1 = scale_net.SerialUNet(
[12, 12 * 6, 12 * 6**2],
[12 * 6**2, 12 * 6**2, 12 * 6**2],
[(3, 3, 3), (3, 3, 3)],
[[(3, 3, 3), (3, 3, 3)], [(3, 3, 3), (3, 3, 3)], [(3, 3, 3),
(3, 3, 3)]],
[[(3, 3, 3), (3, 3, 3)], [(3, 3, 3), (3, 3, 3)]],
input_voxel_size=(36, 36, 36),
)
# input_voxel_size=(
# 36,36,36))
input_size = unet0.min_input_shape
input_size_actual = input_size + added_steps * unet0.step_valid_shape
scnet = scale_net.ScaleNet([unet0, unet1],
input_size_actual,
name="scnet_" + mode)
inputs = []
names = dict()
for k, (inp, vs) in enumerate(zip(scnet.input_shapes, scnet.voxel_sizes)):
raw = tf.placeholder(tf.float32, shape=inp)
raw_bc = tf.reshape(raw, (1, 1) + tuple(inp.astype(np.int)))
inputs.append(raw_bc)
names["raw_{0:}".format(vs[0])] = raw.name
last_fmap, fov, anisotropy = scnet.build(inputs)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[(1, 1, 1)],
num_fmaps=len(labels),
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:]
output_shape = output_shape_c[1:]
dist_c = tf.reshape(dist_bc, output_shape_c)
names["dist"] = dist_c.name
network_outputs = tf.unstack(dist_c, len(labels), axis=0)
if mode.lower() == "train" or mode.lower() == "training":
# mask = tf.placeholder(tf.float32, shape=output_shape)
# names['mask'] = mask.name
# ribo_mask = tf.placeholder(tf.float32, shape=output_shape)
# names['ribo_mask'] = ribo_mask.name
gt = []
w = []
cw = []
masks = []
for l in labels:
masks.append(tf.placeholder(tf.float32, shape=output_shape))
gt.append(tf.placeholder(tf.float32, shape=output_shape))
w.append(tf.placeholder(tf.float32, shape=output_shape))
cw.append(l.class_weight)
lb = []
lub = []
for output_it, gt_it, w_it, m_it, l in zip(network_outputs, gt, w,
masks, labels):
lb.append(
tf.losses.mean_squared_error(gt_it, output_it, w_it * m_it))
lub.append(tf.losses.mean_squared_error(gt_it, output_it, m_it))
# if l.labelname != 'ribosomes':
# lub.append(tf.losses.mean_squared_error(gt_it, output_it, mask))
# else:
# lub.append(tf.losses.mean_squared_error(gt_it, output_it, ribo_mask))
names[l.labelname] = output_it.name
names["gt_" + l.labelname] = gt_it.name
names["w_" + l.labelname] = w_it.name
names["mask_" + l.labelname] = m_it.name
for l, lb_it, lub_it in zip(labels, lb, lub):
tf.summary.scalar("lb_" + l.labelname, lb_it)
tf.summary.scalar("lub_" + l.labelname, lub_it)
names["lb_" + l.labelname] = lb_it.name
names["lub_" + l.labelname] = lub_it.name
loss_total = tf.add_n(lb)
loss_total_unbalanced = tf.add_n(lub)
loss_total_classweighted = tf.tensordot(lb, cw, axes=1)
loss_total_unbalanced_classweighted = tf.tensordot(lub, cw, axes=1)
tf.summary.scalar("loss_total", loss_total)
names["loss_total"] = loss_total.name
tf.summary.scalar("loss_total_unbalanced", loss_total_unbalanced)
names["loss_total_unbalanced"] = loss_total_unbalanced.name
tf.summary.scalar("loss_total_classweighted", loss_total_classweighted)
names["loss_total_classweighted"] = loss_total_classweighted.name
tf.summary.scalar("loss_total_unbalanced_classweighted",
loss_total_unbalanced_classweighted)
names[
"loss_total_unbalanced_classweighted"] = loss_total_unbalanced_classweighted.name
opt = tf.train.AdamOptimizer(learning_rate=0.5e-4,
beta1=0.95,
beta2=0.999,
epsilon=1e-8)
if loss_name == "loss_total":
optimizer = opt.minimize(loss_total)
elif loss_name == "loss_total_unbalanced":
optimizer = opt.minimize(loss_total_unbalanced)
elif loss_name == "loss_total_unbalanced_classweighted":
optimizer = opt.minimize(loss_total_unbalanced_classweighted)
elif loss_name == "loss_total_classweighted":
optimizer = opt.minimize(loss_total_classweighted)
else:
raise ValueError(loss_name + " not defined")
names["optimizer"] = optimizer.name
merged = tf.summary.merge_all()
names["summary"] = merged.name
with open("net_io_names.json", "w") as f:
json.dump(names, f)
elif mode.lower() == "inference" or mode.lower() == "prediction":
pass
else:
raise ValueError(
"unknown mode for network construction: {0:}".format(mode))
tf.train.export_meta_graph(filename=scnet.name + ".meta")
return scnet
def train_net():
input_shape = (43, 430, 430)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = autoencoder.autoencoder(
raw_bc,
12,
6,
[[1, 3, 3], [1, 3, 3], [3, 3, 3]],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(10, 1, 1),
fov=(10, 1, 1),
)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=1,
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:] # strip the channel dimension
dist = tf.reshape(dist_bc, output_shape)
gt_dist = tf.placeholder(tf.float32, shape=output_shape)
loss_weights = tf.placeholder(tf.float32, shape=output_shape)
mask = tf.placeholder(tf.float32, shape=output_shape)
loss_balanced = tf.losses.mean_squared_error(gt_dist, dist, loss_weights)
tf.summary.scalar("loss_balanced_syn", loss_balanced)
loss_unbalanced = tf.losses.mean_squared_error(gt_dist, dist, mask)
tf.summary.scalar("loss_unbalanced_syn", loss_unbalanced)
opt = tf.train.AdamOptimizer(learning_rate=0.5e-4,
beta1=0.95,
beta2=0.999,
epsilon=1e-8)
optimizer = opt.minimize(loss_balanced)
merged = tf.summary.merge_all()
tf.train.export_meta_graph(filename="autoencoder.meta")
names = {
"raw": raw.name,
"dist": dist.name,
"gt_dist": gt_dist.name,
"loss_balanced_syn": loss_balanced.name,
"loss_unbalanced_syn": loss_unbalanced.name,
"loss_weights": loss_weights.name,
"mask": mask.name,
"optimizer": optimizer.name,
"summary": merged.name,
}
with open("net_io_names.json", "w") as f:
json.dump(names, f)
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
)
affs_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=3,
activation="sigmoid",
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = affs_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
affs_c = tf.reshape(affs_bc, output_shape_c)
gt_affs_c = tf.placeholder(tf.float32, shape=output_shape_c)
loss_weights_c = tf.placeholder(tf.float32, shape=output_shape_c)
loss = tf.losses.mean_squared_error(gt_affs_c, affs_c, loss_weights_c)
def train_net():
input_shape = (196, 196, 196)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = unet.unet(
raw_bc,
12,
6,
[[2, 2, 2], [2, 2, 2], [3, 3, 3]],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(1, 1, 1),
fov=(1, 1, 1),
)
pred_raw_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=1,
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = pred_raw_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:]
output_shape = output_shape_c[1:]
pred_raw = tf.reshape(pred_raw_bc, output_shape)
gt_raw_bc = ops3d.crop_zyx(raw_bc, output_shape_bc)
gt_raw = tf.reshape(gt_raw_bc, output_shape)
loss = tf.losses.mean_squared_error(gt_raw, pred_raw)
tf.summary.scalar("loss", loss)
opt = tf.train.AdamOptimizer(learning_rate=0.5e-4,
beta1=0.95,
beta2=0.999,
epsilon=1e-8)
optimizer = opt.minimize(loss)
merged = tf.summary.merge_all()
tf.train.export_meta_graph(filename="unet.meta")
names = {
"raw": raw.name,
"pred_raw": pred_raw.name,
"optimizer": optimizer.name,
"summary": merged.name,
"loss": loss.name,
}
with open("net_io_names.json", "w") as f:
json.dump(names, f)
def make_net(unet, labels, added_steps, loss_name="loss_total", mode="train"):
names = dict()
input_size = unet.min_input_shape
input_size_actual = (input_size + added_steps * unet.step_valid_shape).astype(
np.int
)
raw = tf.placeholder(tf.float32, shape=tuple(input_size_actual))
names["raw"] = raw.name
raw_bc = tf.reshape(raw, (1, 1) + tuple(input_size_actual))
last_fmap, fov, anisotropy = unet.build(raw_bc)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=len(labels),
activation=None,
padding=unet.padding,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:]
dist_c = tf.reshape(dist_bc, output_shape_c)
names["dist"] = dist_c.name
network_outputs = tf.unstack(dist_c, len(labels), axis=0)
if mode.lower() == "train" or mode.lower() == "training":
mask = tf.placeholder(tf.float32, shape=output_shape)
names["mask"] = mask.name
# ribo_mask = tf.placeholder(tf.float32, shape=output_shape)
gt = []
w = []
# cw = []
masks = []
for l in labels:
masks.append(tf.placeholder(tf.float32, shape=output_shape))
gt.append(tf.placeholder(tf.float32, shape=output_shape))
w.append(tf.placeholder(tf.float32, shape=output_shape))
#cw.append(l.class_weight)
lb = []
lub = []
for output_it, gt_it, w_it, m_it, label in zip(
network_outputs, gt, w, masks, labels
):
lb.append(tf.losses.mean_squared_error(gt_it, output_it, w_it * m_it * mask))
lub.append(tf.losses.mean_squared_error(gt_it, output_it, m_it * mask))
names[label.labelname] = output_it.name
names["gt_" + label.labelname] = gt_it.name
names["w_" + label.labelname] = w_it.name
names["mask_" + label.labelname] = m_it.name
for label, lb_it, lub_it in zip(labels, lb, lub):
tf.summary.scalar("lb_" + label.labelname, lb_it)
tf.summary.scalar("lub_" + label.labelname, lub_it)
names["lb_" + label.labelname] = lb_it.name
names["lub_" + label.labelname] = lub_it.name
loss_total = tf.add_n(lb)
loss_total_unbalanced = tf.add_n(lub)
# loss_total_classweighted = tf.tensordot(lb, cw, axes=1)
# loss_total_unbalanced_classweighted = tf.tensordot(lub, cw, axes=1)
tf.summary.scalar("loss_total", loss_total)
names["loss_total"] = loss_total.name
tf.summary.scalar("loss_total_unbalanced", loss_total_unbalanced)
names["loss_total_unbalanced"] = loss_total_unbalanced.name
# tf.summary.scalar("loss_total_classweighted", loss_total_classweighted)
# names["loss_total_classweighted"] = loss_total_classweighted.name
# tf.summary.scalar(
# "loss_total_unbalanced_classweighted", loss_total_unbalanced_classweighted
# )
# names[
# "loss_total_unbalanced_classweighted"
# ] = loss_total_unbalanced_classweighted.name
#
opt = tf.train.AdamOptimizer(
learning_rate=0.5e-4, beta1=0.95, beta2=0.999, epsilon=1e-8
)
if loss_name == "loss_total":
optimizer = opt.minimize(loss_total)
elif loss_name == "loss_total_unbalanced":
optimizer = opt.minimize(loss_total_unbalanced)
# elif loss_name == "loss_total_unbalanced_classweighted":
# optimizer = opt.minimize(loss_total_unbalanced_classweighted)
# elif loss_name == "loss_total_classweighted":
# optimizer = opt.minimize(loss_total_classweighted)
else:
raise ValueError(loss_name + " not defined")
names["optimizer"] = optimizer.name
merged = tf.summary.merge_all()
names["summary"] = merged.name
with open("net_io_names.json", "w") as f:
json.dump(names, f)
elif (
mode.lower() == "inference"
or mode.lower() == "prediction"
or mode.lower() == "pred"
):
pass
else:
raise ValueError("unknown mode for network construction {0:}".format(mode))
net_name = "unet_" + mode
tf.train.export_meta_graph(filename=net_name + ".meta")
return net_name, input_size_actual, output_shape
def make_net(
net_name,
unet,
n_out,
added_context,
sigma=1.0,
lamb=1.0,
input_name="raw",
output_names=None,
loss_name="loss_total",
mode="train",
):
names = dict()
input_size = unet.min_input_shape
if unet.padding == "valid":
assert np.all(np.array(added_context) % np.array(unet.step_valid_shape) == 0), "input shape not suitable for " \
"valid padding"
else:
if not np.all(np.array(added_context) > 0):
logging.warning(
"Small input shape does not generate any output elements free of influence from padding"
)
input_size_actual = (np.array(input_size) +
np.array(added_context)).astype(np.int)
input = tf.placeholder(tf.float32, shape=tuple(input_size_actual))
names[input_name] = input.name
input_bc = tf.reshape(input, (1, 1) + tuple(input_size_actual))
last_fmap, fov, anisotropy = unet.build(input_bc)
output_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=n_out,
activation=None,
padding=unet.padding,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = output_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:]
output_c = tf.reshape(output_bc, output_shape_c)
names["output"] = output_c.name
network_outputs = tf.unstack(output_c, n_out, axis=0)
blurred_full = ops3d.gaussian_blur(input_bc, sigma)
blurred_bc = ops3d.crop_zyx(blurred_full, output_shape_bc)
blurred_c = tf.reshape(blurred_bc, output_shape_c)
blurred = tf.reshape(blurred_c, output_shape)
names["blurred"] = blurred_c.name
if output_names is None:
output_names = ["output_{0:}".format(n) for n in range(n_out)]
assert len(output_names) == n_out
if mode.lower() == "training" or mode.lower() == "forward":
target = []
for tgt in range(n_out):
target.append(tf.placeholder(tf.float32, shape=output_shape))
loss_l2 = []
loss_l1 = []
loss_l2_gauss = []
loss_l1_gauss = []
for output_it, tgt_it, out_name in zip(network_outputs, target,
output_names):
names[out_name + "_predicted"] = output_it.name
names[out_name + "_target"] = tgt_it.name
l2 = tf.losses.mean_squared_error(tgt_it, output_it)
loss_l2.append(l2)
tf.summary.scalar("l2_" + out_name, l2)
names[out_name + "_l2"] = l2.name
l1 = tf.losses.absolute_difference(tgt_it, output_it)
loss_l1.append(l1)
tf.summary.scalar("l1_" + out_name, l1)
names[out_name + "_l1"] = l1.name
l2_gauss = tf.losses.mean_squared_error(blurred, output_it)
loss_l2_gauss.append(l2_gauss)
tf.summary.scalar("l2_gauss_" + out_name, l2_gauss)
names[out_name + "_l2_gauss"] = l2_gauss.name
l1_gauss = tf.losses.absolute_difference(blurred, output_it)
loss_l1_gauss.append(l1_gauss)
tf.summary.scalar("l1_gauss_" + out_name, l1_gauss)
names[out_name + "_l1_gauss"] = l1_gauss.name
l2_total = tf.add_n(loss_l2)
tf.summary.scalar("l2_total", l2_total)
l2_gp_readout = tf.reshape(l2_total, (1, ) * 3)
names["L2"] = l2_gp_readout.name
l1_total = tf.add_n(loss_l1)
tf.summary.scalar("l1_total", l1_total)
l1_gp_readout = tf.reshape(l1_total, (1, ) * 3)
names["L1"] = l1_gp_readout.name
l2_gauss_total = tf.add_n(loss_l2_gauss)
tf.summary.scalar("l2_gauss_total", l2_gauss_total)
l2_gauss_gp_readout = tf.reshape(l2_gauss_total, (1, ) * 3)
names["L2gauss"] = l2_gauss_gp_readout.name
l1_gauss_total = tf.add_n(loss_l1_gauss)
tf.summary.scalar("l1_gauss_total", l1_gauss_total)
l1_gauss_gp_readout = tf.reshape(l1_gauss_total, (1, ) * 3)
names["L1gauss"] = l1_gauss_gp_readout.name
if loss_name == "L2":
loss_opt = l2_total
elif loss_name == "L1":
loss_opt = l1_total
elif loss_name == "L2+L2gauss":
loss_opt = l2_total + lamb * l2_gauss_total
elif loss_name == "L2+L1gauss":
loss_opt = l2_total + lamb * l1_gauss_total
elif loss_name == "L1+L2gauss":
loss_opt = l1_total + lamb * l2_gauss_total
elif loss_name == "L1+L1gauss":
loss_opt = l1_total + lamb * l1_gauss_total
else:
raise ValueError(loss_name + "not defined")
names["loss"] = loss_opt.name
if mode.lower() == "training":
opt = tf.train.AdamOptimizer(learning_rate=0.5e-4,
beta1=0.95,
beta2=0.999,
epsilon=1e-8)
optimizer = opt.minimize(loss_opt)
names["optimizer"] = optimizer.name
merged = tf.summary.merge_all()
names["summary"] = merged.name
with open("{0:}_io_names.json".format(net_name), "w") as f:
json.dump(names, f)
elif (mode.lower() == "inference" or mode.lower() == "prediction"
or mode.lower() == "pred"):
pass
else:
raise ValueError(
"unknown mode for network construction {0:}".format(mode))
tf.train.export_meta_graph(filename=net_name + "_" + mode + ".meta")
return net_name, input_size_actual, output_shape
def train_net():
input_shape = (43, 430, 430)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = unet.unet(
raw_bc,
12,
6,
[[1, 3, 3], [1, 3, 3], [3, 3, 3]],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(10, 1, 1),
fov=(10, 1, 1),
)
logits_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=2,
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = logits_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:] # strip the channel dimension
flat_logits = tf.transpose(tf.reshape(tensor=logits_bc, shape=(2, -1)))
gt_labels = tf.placeholder(tf.float32, shape=output_shape)
gt_labels_flat = tf.reshape(gt_labels, (-1,))
gt_bg = tf.to_float(tf.not_equal(gt_labels_flat, 1))
flat_ohe = tf.stack(values=[gt_labels_flat, gt_bg], axis=1)
loss_weights = tf.placeholder(tf.float32, shape=output_shape)
loss_weights_flat = tf.reshape(loss_weights, (-1,))
mask = tf.placeholder(tf.float32, shape=output_shape)
mask_flat = tf.reshape(mask, (-1,))
probabilities = tf.reshape(tf.nn.softmax(logits_bc, dim=1)[0], output_shape_c)
predictions = tf.argmax(probabilities, axis=0)
ce_loss_balanced = tf.losses.softmax_cross_entropy(
flat_ohe, flat_logits, weights=loss_weights_flat
)
ce_loss_unbalanced = tf.losses.softmax_cross_entropy(
flat_ohe, flat_logits, weights=mask_flat
)
tf.summary.scalar("loss_balanced_syn", ce_loss_balanced)
tf.summary.scalar("loss_unbalanced_syn", ce_loss_unbalanced)
opt = tf.train.AdamOptimizer(
learning_rate=0.5e-4, beta1=0.95, beta2=0.999, epsilon=1e-8
)
optimizer = opt.minimize(ce_loss_balanced)
merged = tf.summary.merge_all()
tf.train.export_meta_graph(filename="unet.meta")
names = {
"raw": raw.name,
"probabilities": probabilities.name,
"predictions": predictions.name,
"gt_labels": gt_labels.name,
"loss_balanced_syn": ce_loss_balanced.name,
"loss_unbalanced_syn": ce_loss_unbalanced.name,
"loss_weights": loss_weights.name,
"mask": mask.name,
"optimizer": optimizer.name,
"summary": merged.name,
}
with open("net_io_names.json", "w") as f:
json.dump(names, f)
def train_net(labels):
input_shape = (196, 196, 196)
raw = tf.placeholder(tf.float32, shape=input_shape)
raw_bc = tf.reshape(raw, (1, 1) + input_shape)
last_fmap, fov, anisotropy = unet.unet(
raw_bc,
[12, 12 * 6, 12 * 6 * 6, 12 * 6 * 6 * 6],
[48, 12 * 6, 12 * 6 * 6, 12 * 6 * 6 * 6],
[[2, 2, 2], [2, 2, 2], [3, 3, 3]],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(1, 1, 1),
fov=(1, 1, 1),
)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=len(labels),
activation=None,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:] # strip the batch dimension
output_shape = output_shape_c[1:]
dist_c = tf.reshape(dist_bc, output_shape_c)
network_outputs = tf.unstack(dist_c, len(labels), axis=0)
mask = tf.placeholder(tf.float32, shape=output_shape)
gt = []
w = []
for l in range(len(labels)):
gt.append(tf.placeholder(tf.float32, shape=output_shape))
w.append(tf.placeholder(tf.float32, shape=output_shape))
lb = []
lub = []
for output_it, gt_it, w_it in zip(network_outputs, gt, w):
lb.append(tf.losses.mean_squared_error(gt_it, output_it, w_it))
lub.append(tf.losses.mean_squared_error(gt_it, output_it, mask))
for label, lb_it, lub_it in zip(labels, lb, lub):
tf.summary.scalar("lb_" + label, lb_it)
tf.summary.scalar("lub_" + label, lub_it)
loss_total = tf.add_n(lb)
loss_total_unbalanced = tf.add_n(lub)
tf.summary.scalar("loss_total", loss_total)
tf.summary.scalar("loss_total_unbalanced", loss_total_unbalanced)
opt = tf.train.AdamOptimizer(learning_rate=0.5e-4,
beta1=0.95,
beta2=0.999,
epsilon=1e-8)
optimizer = opt.minimize(loss_total)
merged = tf.summary.merge_all()
tf.train.export_meta_graph(filename="unet.meta")
names = {
"raw": raw.name,
"dist": dist_c.name,
"loss_total": loss_total.name,
"loss_total_unbalanced": loss_total_unbalanced.name,
"mask": mask.name,
"optimizer": optimizer.name,
"summary": merged.name,
}
for label, output_it, gt_it, w_it, lb_it, lub_it in zip(
labels, network_outputs, gt, w, lb, lub):
names[label] = output_it.name
names["gt_" + label] = gt_it.name
names["w_" + label] = w_it.name
names["lb_" + label] = lb_it.name
names["lub_" + label] = lub_it.name
with open("net_io_names.json", "w") as f:
json.dump(names, f)
def make_net(unet, added_steps, loss_name="loss_total", padding="valid", mode="train"):
# input_shape = (43, 430, 430)
names = dict()
if padding == "valid":
input_size = unet.min_input_shape
else:
input_size = np.array((0, 0, 0))
input_size_actual = (input_size + added_steps * unet.step_valid_shape).astype(
np.int
)
raw = tf.placeholder(tf.float32, shape=tuple(input_size_actual))
names["raw"] = raw.name
raw_bc = tf.reshape(raw, (1, 1) + tuple(input_size_actual))
last_fmap, fov, anisotropy = unet.build(raw_bc)
dist_bc, fov = ops3d.conv_pass(
last_fmap,
kernel_size=[[1, 1, 1]],
num_fmaps=3,
activation=None,
padding=padding,
fov=fov,
voxel_size=anisotropy,
)
output_shape_bc = dist_bc.get_shape().as_list()
output_shape_c = output_shape_bc[1:]
output_shape = output_shape_c[1:]
dist_c = tf.reshape(dist_bc, shape=output_shape_c)
names["dist"] = dist_c.name
cleft_dist, pre_dist, post_dist = tf.unstack(dist_c, 3, axis=0)
names["cleft_dist"] = cleft_dist.name
names["pre_dist"] = pre_dist.name
names["post_dist"] = post_dist.name
if mode.lower() == "train" or mode.lower() == "training":
gt_cleft_dist = tf.placeholder(tf.float32, shape=output_shape)
gt_pre_dist = tf.placeholder(tf.float32, shape=output_shape)
gt_post_dist = tf.placeholder(tf.float32, shape=output_shape)
names["gt_cleft_dist"] = gt_cleft_dist.name
names["gt_pre_dist"] = gt_pre_dist.name
names["gt_post_dist"] = gt_post_dist.name
loss_weights_cleft = tf.placeholder(tf.float32, shape=output_shape)
loss_weights_pre = tf.placeholder(tf.float32, shape=output_shape)
loss_weights_post = tf.placeholder(tf.float32, shape=output_shape)
names["loss_weights_cleft"] = loss_weights_cleft.name
names["loss_weights_pre"] = loss_weights_pre.name
names["loss_weights_post"] = loss_weights_post.name
cleft_mask = tf.placeholder(tf.float32, shape=output_shape)
pre_mask = tf.placeholder(tf.float32, shape=output_shape)
post_mask = tf.placeholder(tf.float32, shape=output_shape)
names["cleft_mask"] = cleft_mask.name
names["pre_mask"] = pre_mask.name
names["post_mask"] = post_mask.name
loss_balanced_cleft = tf.losses.mean_squared_error(
gt_cleft_dist, cleft_dist, loss_weights_cleft * cleft_mask
)
loss_balanced_pre = tf.losses.mean_squared_error(
gt_pre_dist, pre_dist, loss_weights_pre * pre_mask
)
loss_balanced_post = tf.losses.mean_squared_error(
gt_post_dist, post_dist, loss_weights_post * post_mask
)
names["loss_balanced_cleft"] = loss_balanced_cleft.name
names["loss_balanced_pre"] = loss_balanced_pre.name
names["loss_balanced_post"] = loss_balanced_post.name
loss_unbalanced_cleft = tf.losses.mean_squared_error(
gt_cleft_dist, cleft_dist, cleft_mask
)
loss_unbalanced_pre = tf.losses.mean_squared_error(
gt_pre_dist, pre_dist, pre_mask
)
loss_unbalanced_post = tf.losses.mean_squared_error(
gt_post_dist, post_dist, post_mask
)
names["loss_unbalanced_cleft"] = loss_unbalanced_cleft.name
names["loss_unbalanced_pre"] = loss_unbalanced_pre.name
names["loss_unbalanced_post"] = loss_unbalanced_post.name
loss_total = loss_balanced_cleft + loss_balanced_pre + loss_balanced_post
loss_total_unbalanced = (
loss_unbalanced_cleft + loss_unbalanced_pre + loss_unbalanced_post
)
names["loss_total"] = loss_total.name
names["loss_total_unbalanced"] = loss_total_unbalanced.name
tf.summary.scalar("loss_balanced_cleft", loss_balanced_cleft)
tf.summary.scalar("loss_balanced_pre", loss_balanced_pre)
tf.summary.scalar("loss_balanced_post", loss_balanced_post)
tf.summary.scalar("loss_unbalanced_cleft", loss_unbalanced_cleft)
tf.summary.scalar("loss_unbalanced_pre", loss_unbalanced_pre)
tf.summary.scalar("loss_unbalanced_post", loss_unbalanced_post)
tf.summary.scalar("loss_total", loss_total)
tf.summary.scalar("loss_total_unbalanced", loss_total_unbalanced)
opt = tf.train.AdamOptimizer(
learning_rate=0.5e-4, beta1=0.95, beta2=0.999, epsilon=1e-8
)
if loss_name == "loss_total":
optimizer = opt.minimize(loss_total)
elif loss_name == "loss_total_unbalanced":
optimizer = opt.minimize(loss_total_unbalanced)
else:
raise ValueError(loss_name + " not defined")
names["optimizer"] = optimizer.name
merged = tf.summary.merge_all()
names["summary"] = merged.name
with open("net_io_names.json", "w") as f:
json.dump(names, f)
elif (
mode.lower() == "inference"
or mode.lower() == "prediction"
or mode.lower() == "pred"
):
pass
else:
raise ValueError("unknown mode for netowrk construction: {0:}".format(mode))
net_name = "unet_" + mode
tf.train.export_meta_graph(filename=net_name + ".meta")
return net_name, input_size_actual, output_shape
def unet(
fmaps_in,
num_fmaps_down,
num_fmaps_up,
downsample_factors,
kernel_size_down,
kernel_size_up,
activation="relu",
layer=0,
fov=(1, 1, 1),
voxel_size=(1, 1, 1),
constant_upsample=False,
):
"""Create a U-Net::
f_in --> f_left --------------------------->> f_right--> f_out
| ^
v |
g_in --> g_left ------->> g_right --> g_out
| ^
v |
...
where each ``-->`` is a convolution pass (see ``conv_pass``), each `-->>` a
crop, and down and up arrows are max-pooling and transposed convolutions,
respectively.
The U-Net expects tensors to have shape ``(batch=1, channels, depth, height,
width)``.
This U-Net performs only "valid" convolutions, i.e., sizes of the feature
maps decrease after each convolution.
Args:
fmaps_in:
The input tensor.
num_fmaps:
The number of feature maps in the first layer. This is also the
number of output feature maps.
fmap_inc_factors:
By how much to multiply the number of feature maps between layers.
If layer 0 has ``k`` feature maps, layer ``l`` will have
``k*fmap_inc_factor**l``.
downsample_factors:
List of lists ``[z, y, x]`` to use to down- and up-sample the
feature maps between layers.
kernel_size_down:
List of lists of tuples ``(z, y, x)`` of kernel sizes. The number of
tuples in a list determines the number of convolutional layers in the
corresponding level of the build on the left side.
kernel_size_up:
List of lists of tuples ``(z, y, x)`` of kernel sizes. The number of
tuples in a list determines the number of convolutional layers in the
corresponding level of the build on the right side. Within one of the
lists going from left to right.
activation:
Which activation to use after a convolution. Accepts the name of any
tensorflow activation function (e.g., ``relu`` for ``tf.nn.relu``).
layer:
Used internally to build the U-Net recursively.
fov:
Initial field of view in physical units
voxel_size:
Size of a voxel in the input data, in physical units
"""
prefix = " " * layer
print(prefix + "Creating U-Net layer %i" % layer)
print(prefix + "f_in: " + str(fmaps_in.shape))
# if isinstance(fmap_inc_factors, int):
# fmap_inc_factors = [fmap_inc_factors]*len(downsample_factors)
assert (
len(num_fmaps_down) - 1
== len(num_fmaps_up) - 1
== len(downsample_factors)
== len(kernel_size_down) - 1
== len(kernel_size_up) - 1
)
# convolve
with tf.name_scope("lev%i" % layer):
f_left, fov = ops3d.conv_pass(
fmaps_in,
kernel_size=kernel_size_down[layer],
num_fmaps=num_fmaps_down[layer],
activation=activation,
name="unet_layer_%i_left" % layer,
fov=fov,
voxel_size=voxel_size,
prefix=prefix,
)
# last layer does not recurse
bottom_layer = layer == len(downsample_factors)
if bottom_layer:
print(prefix + "bottom layer")
print(prefix + "f_out: " + str(f_left.shape))
return f_left, fov, voxel_size
# downsample
g_in, fov, voxel_size = ops3d.downsample(
f_left,
downsample_factors[layer],
"unet_down_%i_to_%i" % (layer, layer + 1),
fov=fov,
voxel_size=voxel_size,
prefix=prefix,
)
# recursive U-net
g_out, fov, voxel_size = unet(
g_in,
num_fmaps_down=num_fmaps_down,
num_fmaps_up=num_fmaps_up,
downsample_factors=downsample_factors,
kernel_size_down=kernel_size_down,
kernel_size_up=kernel_size_up,
activation=activation,
layer=layer + 1,
fov=fov,
voxel_size=voxel_size,
constant_upsample=constant_upsample,
)
print(prefix + "g_out: " + str(g_out.shape))
# upsample
g_out_upsampled, voxel_size = ops3d.upsample(
g_out,
downsample_factors[layer],
num_fmaps_up[layer],
activation=activation,
name="unet_up_%i_to_%i" % (layer + 1, layer),
fov=fov,
voxel_size=voxel_size,
prefix=prefix,
constant_upsample=constant_upsample,
)
print(prefix + "g_out_upsampled: " + str(g_out_upsampled.shape))
# copy-crop
f_left_cropped = ops3d.crop_zyx(f_left, g_out_upsampled.get_shape().as_list())
print(prefix + "f_left_cropped: " + str(f_left_cropped.shape))
# concatenate along channel dimension
f_right = tf.concat([f_left_cropped, g_out_upsampled], 1)
print(prefix + "f_right: " + str(f_right.shape))
# convolve
f_out, fov = ops3d.conv_pass(
f_right,
kernel_size=kernel_size_up[layer],
num_fmaps=num_fmaps_up[layer],
name="unet_layer_%i_right" % layer,
fov=fov,
voxel_size=voxel_size,
prefix=prefix,
)
print(prefix + "f_out: " + str(f_out.shape))
return f_out, fov, voxel_size
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
[
[(1, 3, 3), (1, 3, 3)],
[(1, 3, 3), (1, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
[(3, 3, 3), (3, 3, 3)],
],
voxel_size=(10, 1, 1),
fov=(10, 1, 1),
)
output, full_fov = ops3d.conv_pass(
model,
kernel_size=[(1, 1, 1)],
num_fmaps=1,
activation=None,
fov=ll_fov,
voxel_size=vx,
)
tf.train.export_meta_graph(filename="build.meta")
with tf.Session() as session:
session.run(tf.initialize_all_variables())
tf.summary.FileWriter(".", graph=tf.get_default_graph())
print(model.shape)
