Upload ProbUNet
Browse files- PULASki.py +48 -0
- PULASkiConfigs.py +26 -0
- ProbUNet_model.py +731 -0
- ProbUNet_utils.py +224 -0
- config.json +21 -0
- model.safetensors +3 -0
PULASki.py
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
import sys
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
import torch.nn as nn
|
| 6 |
+
|
| 7 |
+
from transformers import PreTrainedModel
|
| 8 |
+
|
| 9 |
+
from .ProbUNet_model import InjectionConvEncoder2D, InjectionUNet2D, InjectionConvEncoder3D, InjectionUNet3D, ProbabilisticSegmentationNet
|
| 10 |
+
from .PULASkiConfigs import ProbUNetConfig
|
| 11 |
+
|
| 12 |
+
class ProbUNet(PreTrainedModel):
|
| 13 |
+
config_class = ProbUNetConfig
|
| 14 |
+
def __init__(self, config):
|
| 15 |
+
super().__init__(config)
|
| 16 |
+
|
| 17 |
+
if config.dim == 2:
|
| 18 |
+
task_op = InjectionUNet2D
|
| 19 |
+
prior_op = InjectionConvEncoder2D
|
| 20 |
+
posterior_op = InjectionConvEncoder2D
|
| 21 |
+
elif config.dim == 3:
|
| 22 |
+
task_op = InjectionUNet3D
|
| 23 |
+
prior_op = InjectionConvEncoder3D
|
| 24 |
+
posterior_op = InjectionConvEncoder3D
|
| 25 |
+
else:
|
| 26 |
+
sys.exit("Invalid dim! Only configured for dim 2 and 3.")
|
| 27 |
+
|
| 28 |
+
if config.latent_distribution == "normal":
|
| 29 |
+
latent_distribution = torch.distributions.Normal
|
| 30 |
+
else:
|
| 31 |
+
sys.exit("Invalid latent_distribution. Only normal has been implemented.")
|
| 32 |
+
|
| 33 |
+
self.model = ProbabilisticSegmentationNet(in_channels=config.in_channels,
|
| 34 |
+
out_channels=config.out_channels,
|
| 35 |
+
num_feature_maps=config.num_feature_maps,
|
| 36 |
+
latent_size=config.latent_size,
|
| 37 |
+
depth=config.depth,
|
| 38 |
+
latent_distribution=latent_distribution,
|
| 39 |
+
task_op=task_op,
|
| 40 |
+
task_kwargs={"output_activation_op": nn.Identity if config.no_outact_op else nn.Sigmoid,
|
| 41 |
+
"activation_kwargs": {"inplace": True}, "injection_at": config.prob_injection_at},
|
| 42 |
+
prior_op=prior_op,
|
| 43 |
+
prior_kwargs={"activation_kwargs": {"inplace": True}, "norm_depth": 2},
|
| 44 |
+
posterior_op=posterior_op,
|
| 45 |
+
posterior_kwargs={"activation_kwargs": {"inplace": True}, "norm_depth": 2},
|
| 46 |
+
)
|
| 47 |
+
def forward(self, x):
|
| 48 |
+
return self.model(x)
|
PULASkiConfigs.py
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import PretrainedConfig
|
| 2 |
+
|
| 3 |
+
class ProbUNetConfig(PretrainedConfig):
|
| 4 |
+
model_type = "ProbUNet"
|
| 5 |
+
def __init__(
|
| 6 |
+
self,
|
| 7 |
+
dim=2,
|
| 8 |
+
in_channels=1,
|
| 9 |
+
out_channels=1,
|
| 10 |
+
num_feature_maps=24,
|
| 11 |
+
latent_size=3,
|
| 12 |
+
depth=5,
|
| 13 |
+
latent_distribution="normal",
|
| 14 |
+
no_outact_op=False,
|
| 15 |
+
prob_injection_at="end",
|
| 16 |
+
**kwargs):
|
| 17 |
+
self.dim = dim
|
| 18 |
+
self.in_channels = in_channels
|
| 19 |
+
self.out_channels = out_channels
|
| 20 |
+
self.num_feature_maps = num_feature_maps
|
| 21 |
+
self.latent_size = latent_size
|
| 22 |
+
self.depth = depth
|
| 23 |
+
self.latent_distribution = latent_distribution
|
| 24 |
+
self.no_outact_op = no_outact_op
|
| 25 |
+
self.prob_injection_at = prob_injection_at
|
| 26 |
+
super().__init__(**kwargs)
|
ProbUNet_model.py
ADDED
|
@@ -0,0 +1,731 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
|
| 4 |
+
from .ProbUNet_utils import make_onehot as make_onehot_segmentation, make_slices, match_to
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def is_conv(op):
|
| 8 |
+
conv_types = (nn.Conv1d,
|
| 9 |
+
nn.Conv2d,
|
| 10 |
+
nn.Conv3d,
|
| 11 |
+
nn.ConvTranspose1d,
|
| 12 |
+
nn.ConvTranspose2d,
|
| 13 |
+
nn.ConvTranspose3d)
|
| 14 |
+
if type(op) == type and issubclass(op, conv_types):
|
| 15 |
+
return True
|
| 16 |
+
elif type(op) in conv_types:
|
| 17 |
+
return True
|
| 18 |
+
else:
|
| 19 |
+
return False
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
class ConvModule(nn.Module):
|
| 24 |
+
|
| 25 |
+
def __init__(self, *args, **kwargs):
|
| 26 |
+
|
| 27 |
+
super(ConvModule, self).__init__()
|
| 28 |
+
|
| 29 |
+
def init_weights(self, init_fn, *args, **kwargs):
|
| 30 |
+
|
| 31 |
+
class init_(object):
|
| 32 |
+
|
| 33 |
+
def __init__(self):
|
| 34 |
+
self.fn = init_fn
|
| 35 |
+
self.args = args
|
| 36 |
+
self.kwargs = kwargs
|
| 37 |
+
|
| 38 |
+
def __call__(self, module):
|
| 39 |
+
if is_conv(type(module)):
|
| 40 |
+
module.weight = self.fn(module.weight, *self.args, **self.kwargs)
|
| 41 |
+
|
| 42 |
+
_init_ = init_()
|
| 43 |
+
self.apply(_init_)
|
| 44 |
+
|
| 45 |
+
def init_bias(self, init_fn, *args, **kwargs):
|
| 46 |
+
|
| 47 |
+
class init_(object):
|
| 48 |
+
|
| 49 |
+
def __init__(self):
|
| 50 |
+
self.fn = init_fn
|
| 51 |
+
self.args = args
|
| 52 |
+
self.kwargs = kwargs
|
| 53 |
+
|
| 54 |
+
def __call__(self, module):
|
| 55 |
+
if is_conv(type(module)) and module.bias is not None:
|
| 56 |
+
module.bias = self.fn(module.bias, *self.args, **self.kwargs)
|
| 57 |
+
|
| 58 |
+
_init_ = init_()
|
| 59 |
+
self.apply(_init_)
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
class ConcatCoords(nn.Module):
|
| 64 |
+
|
| 65 |
+
def forward(self, input_):
|
| 66 |
+
|
| 67 |
+
dim = input_.dim() - 2
|
| 68 |
+
coord_channels = []
|
| 69 |
+
for i in range(dim):
|
| 70 |
+
view = [1, ] * dim
|
| 71 |
+
view[i] = -1
|
| 72 |
+
repeat = list(input_.shape[2:])
|
| 73 |
+
repeat[i] = 1
|
| 74 |
+
coord_channels.append(
|
| 75 |
+
torch.linspace(-0.5, 0.5, input_.shape[i+2])
|
| 76 |
+
.view(*view)
|
| 77 |
+
.repeat(*repeat)
|
| 78 |
+
.to(device=input_.device, dtype=input_.dtype))
|
| 79 |
+
coord_channels = torch.stack(coord_channels).unsqueeze(0)
|
| 80 |
+
repeat = [1, ] * input_.dim()
|
| 81 |
+
repeat[0] = input_.shape[0]
|
| 82 |
+
coord_channels = coord_channels.repeat(*repeat).contiguous()
|
| 83 |
+
|
| 84 |
+
return torch.cat([input_, coord_channels], 1)
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
class InjectionConvEncoder(ConvModule):
|
| 89 |
+
|
| 90 |
+
_default_activation_kwargs = dict(inplace=True)
|
| 91 |
+
_default_norm_kwargs = dict()
|
| 92 |
+
_default_conv_kwargs = dict(kernel_size=3, padding=1)
|
| 93 |
+
_default_pool_kwargs = dict(kernel_size=2)
|
| 94 |
+
_default_dropout_kwargs = dict()
|
| 95 |
+
_default_global_pool_kwargs = dict()
|
| 96 |
+
|
| 97 |
+
def __init__(self,
|
| 98 |
+
in_channels=1,
|
| 99 |
+
out_channels=6,
|
| 100 |
+
depth=4,
|
| 101 |
+
injection_depth="last",
|
| 102 |
+
injection_channels=0,
|
| 103 |
+
block_depth=2,
|
| 104 |
+
num_feature_maps=24,
|
| 105 |
+
feature_map_multiplier=2,
|
| 106 |
+
activation_op=nn.LeakyReLU,
|
| 107 |
+
activation_kwargs=None,
|
| 108 |
+
norm_op=nn.InstanceNorm2d,
|
| 109 |
+
norm_kwargs=None,
|
| 110 |
+
norm_depth=0,
|
| 111 |
+
conv_op=nn.Conv2d,
|
| 112 |
+
conv_kwargs=None,
|
| 113 |
+
pool_op=nn.AvgPool2d,
|
| 114 |
+
pool_kwargs=None,
|
| 115 |
+
dropout_op=None,
|
| 116 |
+
dropout_kwargs=None,
|
| 117 |
+
global_pool_op=nn.AdaptiveAvgPool2d,
|
| 118 |
+
global_pool_kwargs=None,
|
| 119 |
+
**kwargs):
|
| 120 |
+
|
| 121 |
+
super(InjectionConvEncoder, self).__init__(**kwargs)
|
| 122 |
+
|
| 123 |
+
self.in_channels = in_channels
|
| 124 |
+
self.out_channels = out_channels
|
| 125 |
+
self.depth = depth
|
| 126 |
+
self.injection_depth = depth - 1 if injection_depth == "last" else injection_depth
|
| 127 |
+
self.injection_channels = injection_channels
|
| 128 |
+
self.block_depth = block_depth
|
| 129 |
+
self.num_feature_maps = num_feature_maps
|
| 130 |
+
self.feature_map_multiplier = feature_map_multiplier
|
| 131 |
+
|
| 132 |
+
self.activation_op = activation_op
|
| 133 |
+
self.activation_kwargs = self._default_activation_kwargs
|
| 134 |
+
if activation_kwargs is not None:
|
| 135 |
+
self.activation_kwargs.update(activation_kwargs)
|
| 136 |
+
|
| 137 |
+
self.norm_op = norm_op
|
| 138 |
+
self.norm_kwargs = self._default_norm_kwargs
|
| 139 |
+
if norm_kwargs is not None:
|
| 140 |
+
self.norm_kwargs.update(norm_kwargs)
|
| 141 |
+
self.norm_depth = depth if norm_depth == "full" else norm_depth
|
| 142 |
+
|
| 143 |
+
self.conv_op = conv_op
|
| 144 |
+
self.conv_kwargs = self._default_conv_kwargs
|
| 145 |
+
if conv_kwargs is not None:
|
| 146 |
+
self.conv_kwargs.update(conv_kwargs)
|
| 147 |
+
|
| 148 |
+
self.pool_op = pool_op
|
| 149 |
+
self.pool_kwargs = self._default_pool_kwargs
|
| 150 |
+
if pool_kwargs is not None:
|
| 151 |
+
self.pool_kwargs.update(pool_kwargs)
|
| 152 |
+
|
| 153 |
+
self.dropout_op = dropout_op
|
| 154 |
+
self.dropout_kwargs = self._default_dropout_kwargs
|
| 155 |
+
if dropout_kwargs is not None:
|
| 156 |
+
self.dropout_kwargs.update(dropout_kwargs)
|
| 157 |
+
|
| 158 |
+
self.global_pool_op = global_pool_op
|
| 159 |
+
self.global_pool_kwargs = self._default_global_pool_kwargs
|
| 160 |
+
if global_pool_kwargs is not None:
|
| 161 |
+
self.global_pool_kwargs.update(global_pool_kwargs)
|
| 162 |
+
|
| 163 |
+
for d in range(self.depth):
|
| 164 |
+
|
| 165 |
+
in_ = self.in_channels if d == 0 else self.num_feature_maps * (self.feature_map_multiplier**(d-1))
|
| 166 |
+
out_ = self.num_feature_maps * (self.feature_map_multiplier**d)
|
| 167 |
+
|
| 168 |
+
if d == self.injection_depth + 1:
|
| 169 |
+
in_ += self.injection_channels
|
| 170 |
+
|
| 171 |
+
layers = []
|
| 172 |
+
if d > 0:
|
| 173 |
+
layers.append(self.pool_op(**self.pool_kwargs))
|
| 174 |
+
for b in range(self.block_depth):
|
| 175 |
+
current_in = in_ if b == 0 else out_
|
| 176 |
+
layers.append(self.conv_op(current_in, out_, **self.conv_kwargs))
|
| 177 |
+
if self.norm_op is not None and d < self.norm_depth:
|
| 178 |
+
layers.append(self.norm_op(out_, **self.norm_kwargs))
|
| 179 |
+
if self.activation_op is not None:
|
| 180 |
+
layers.append(self.activation_op(**self.activation_kwargs))
|
| 181 |
+
if self.dropout_op is not None:
|
| 182 |
+
layers.append(self.dropout_op(**self.dropout_kwargs))
|
| 183 |
+
if d == self.depth - 1:
|
| 184 |
+
current_conv_kwargs = self.conv_kwargs.copy()
|
| 185 |
+
current_conv_kwargs["kernel_size"] = 1
|
| 186 |
+
current_conv_kwargs["padding"] = 0
|
| 187 |
+
current_conv_kwargs["bias"] = False
|
| 188 |
+
layers.append(self.conv_op(out_, out_channels, **current_conv_kwargs))
|
| 189 |
+
|
| 190 |
+
self.add_module("encode_{}".format(d), nn.Sequential(*layers))
|
| 191 |
+
|
| 192 |
+
if self.global_pool_op is not None:
|
| 193 |
+
self.add_module("global_pool", self.global_pool_op(1, **self.global_pool_kwargs))
|
| 194 |
+
|
| 195 |
+
def forward(self, x, injection=None):
|
| 196 |
+
|
| 197 |
+
for d in range(self.depth):
|
| 198 |
+
x = self._modules["encode_{}".format(d)](x)
|
| 199 |
+
if d == self.injection_depth and self.injection_channels > 0:
|
| 200 |
+
injection = match_to(injection, x, self.injection_channels)
|
| 201 |
+
x = torch.cat([x, injection], 1)
|
| 202 |
+
if hasattr(self, "global_pool"):
|
| 203 |
+
x = self.global_pool(x)
|
| 204 |
+
|
| 205 |
+
return x
|
| 206 |
+
|
| 207 |
+
|
| 208 |
+
class InjectionConvEncoder3D(InjectionConvEncoder):
|
| 209 |
+
|
| 210 |
+
def __init__(self, *args, **kwargs):
|
| 211 |
+
|
| 212 |
+
update_kwargs = dict(
|
| 213 |
+
norm_op=nn.InstanceNorm3d,
|
| 214 |
+
conv_op=nn.Conv3d,
|
| 215 |
+
pool_op=nn.AvgPool3d,
|
| 216 |
+
global_pool_op=nn.AdaptiveAvgPool3d
|
| 217 |
+
)
|
| 218 |
+
|
| 219 |
+
for (arg, val) in update_kwargs.items():
|
| 220 |
+
if arg not in kwargs: kwargs[arg] = val
|
| 221 |
+
|
| 222 |
+
super(InjectionConvEncoder3D, self).__init__(*args, **kwargs)
|
| 223 |
+
|
| 224 |
+
class InjectionConvEncoder2D(InjectionConvEncoder): #Created by Soumick
|
| 225 |
+
|
| 226 |
+
def __init__(self, *args, **kwargs):
|
| 227 |
+
|
| 228 |
+
update_kwargs = dict(
|
| 229 |
+
norm_op=nn.InstanceNorm2d,
|
| 230 |
+
conv_op=nn.Conv2d,
|
| 231 |
+
pool_op=nn.AvgPool2d,
|
| 232 |
+
global_pool_op=nn.AdaptiveAvgPool2d
|
| 233 |
+
)
|
| 234 |
+
|
| 235 |
+
for (arg, val) in update_kwargs.items():
|
| 236 |
+
if arg not in kwargs: kwargs[arg] = val
|
| 237 |
+
|
| 238 |
+
super(InjectionConvEncoder2D, self).__init__(*args, **kwargs)
|
| 239 |
+
|
| 240 |
+
class InjectionUNet(ConvModule):
|
| 241 |
+
|
| 242 |
+
def __init__(
|
| 243 |
+
self,
|
| 244 |
+
depth=5,
|
| 245 |
+
in_channels=4,
|
| 246 |
+
out_channels=4,
|
| 247 |
+
kernel_size=3,
|
| 248 |
+
dilation=1,
|
| 249 |
+
num_feature_maps=24,
|
| 250 |
+
block_depth=2,
|
| 251 |
+
num_1x1_at_end=3,
|
| 252 |
+
injection_channels=3,
|
| 253 |
+
injection_at="end",
|
| 254 |
+
activation_op=nn.LeakyReLU,
|
| 255 |
+
activation_kwargs=None,
|
| 256 |
+
pool_op=nn.AvgPool2d,
|
| 257 |
+
pool_kwargs=dict(kernel_size=2),
|
| 258 |
+
dropout_op=None,
|
| 259 |
+
dropout_kwargs=None,
|
| 260 |
+
norm_op=nn.InstanceNorm2d,
|
| 261 |
+
norm_kwargs=None,
|
| 262 |
+
conv_op=nn.Conv2d,
|
| 263 |
+
conv_kwargs=None,
|
| 264 |
+
upconv_op=nn.ConvTranspose2d,
|
| 265 |
+
upconv_kwargs=None,
|
| 266 |
+
output_activation_op=None,
|
| 267 |
+
output_activation_kwargs=None,
|
| 268 |
+
return_bottom=False,
|
| 269 |
+
coords=False,
|
| 270 |
+
coords_dim=2,
|
| 271 |
+
**kwargs
|
| 272 |
+
):
|
| 273 |
+
|
| 274 |
+
super(InjectionUNet, self).__init__(**kwargs)
|
| 275 |
+
|
| 276 |
+
self.depth = depth
|
| 277 |
+
self.in_channels = in_channels
|
| 278 |
+
self.out_channels = out_channels
|
| 279 |
+
self.kernel_size = kernel_size
|
| 280 |
+
self.dilation = dilation
|
| 281 |
+
self.padding = (self.kernel_size + (self.kernel_size-1) * (self.dilation-1)) // 2
|
| 282 |
+
self.num_feature_maps = num_feature_maps
|
| 283 |
+
self.block_depth = block_depth
|
| 284 |
+
self.num_1x1_at_end = num_1x1_at_end
|
| 285 |
+
self.injection_channels = injection_channels
|
| 286 |
+
self.injection_at = injection_at
|
| 287 |
+
self.activation_op = activation_op
|
| 288 |
+
self.activation_kwargs = {} if activation_kwargs is None else activation_kwargs
|
| 289 |
+
self.pool_op = pool_op
|
| 290 |
+
self.pool_kwargs = {} if pool_kwargs is None else pool_kwargs
|
| 291 |
+
self.dropout_op = dropout_op
|
| 292 |
+
self.dropout_kwargs = {} if dropout_kwargs is None else dropout_kwargs
|
| 293 |
+
self.norm_op = norm_op
|
| 294 |
+
self.norm_kwargs = {} if norm_kwargs is None else norm_kwargs
|
| 295 |
+
self.conv_op = conv_op
|
| 296 |
+
self.conv_kwargs = {} if conv_kwargs is None else conv_kwargs
|
| 297 |
+
self.upconv_op = upconv_op
|
| 298 |
+
self.upconv_kwargs = {} if upconv_kwargs is None else upconv_kwargs
|
| 299 |
+
self.output_activation_op = output_activation_op
|
| 300 |
+
self.output_activation_kwargs = {} if output_activation_kwargs is None else output_activation_kwargs
|
| 301 |
+
self.return_bottom = return_bottom
|
| 302 |
+
if not coords:
|
| 303 |
+
self.coords = [[], []]
|
| 304 |
+
elif coords is True:
|
| 305 |
+
self.coords = [list(range(depth)), []]
|
| 306 |
+
else:
|
| 307 |
+
self.coords = coords
|
| 308 |
+
self.coords_dim = coords_dim
|
| 309 |
+
|
| 310 |
+
self.last_activations = None
|
| 311 |
+
|
| 312 |
+
# BUILD ENCODER
|
| 313 |
+
for d in range(self.depth):
|
| 314 |
+
|
| 315 |
+
block = []
|
| 316 |
+
if d > 0:
|
| 317 |
+
block.append(self.pool_op(**self.pool_kwargs))
|
| 318 |
+
|
| 319 |
+
for i in range(self.block_depth):
|
| 320 |
+
|
| 321 |
+
# bottom block fixed to have depth 1
|
| 322 |
+
if d == self.depth - 1 and i > 0:
|
| 323 |
+
continue
|
| 324 |
+
|
| 325 |
+
out_size = self.num_feature_maps * 2**d
|
| 326 |
+
if d == 0 and i == 0:
|
| 327 |
+
in_size = self.in_channels
|
| 328 |
+
elif i == 0:
|
| 329 |
+
in_size = self.num_feature_maps * 2**(d - 1)
|
| 330 |
+
else:
|
| 331 |
+
in_size = out_size
|
| 332 |
+
|
| 333 |
+
# check for coord appending at this depth
|
| 334 |
+
if d in self.coords[0] and i == 0:
|
| 335 |
+
block.append(ConcatCoords())
|
| 336 |
+
in_size += self.coords_dim
|
| 337 |
+
|
| 338 |
+
block.append(self.conv_op(in_size,
|
| 339 |
+
out_size,
|
| 340 |
+
self.kernel_size,
|
| 341 |
+
padding=self.padding,
|
| 342 |
+
dilation=self.dilation,
|
| 343 |
+
**self.conv_kwargs))
|
| 344 |
+
if self.dropout_op is not None:
|
| 345 |
+
block.append(self.dropout_op(**self.dropout_kwargs))
|
| 346 |
+
if self.norm_op is not None:
|
| 347 |
+
block.append(self.norm_op(out_size, **self.norm_kwargs))
|
| 348 |
+
block.append(self.activation_op(**self.activation_kwargs))
|
| 349 |
+
|
| 350 |
+
self.add_module("encode-{}".format(d), nn.Sequential(*block))
|
| 351 |
+
|
| 352 |
+
# BUILD DECODER
|
| 353 |
+
for d in reversed(range(self.depth)):
|
| 354 |
+
|
| 355 |
+
block = []
|
| 356 |
+
|
| 357 |
+
for i in range(self.block_depth):
|
| 358 |
+
|
| 359 |
+
# bottom block fixed to have depth 1
|
| 360 |
+
if d == self.depth - 1 and i > 0:
|
| 361 |
+
continue
|
| 362 |
+
|
| 363 |
+
out_size = self.num_feature_maps * 2**(d)
|
| 364 |
+
if i == 0 and d < self.depth - 1:
|
| 365 |
+
in_size = self.num_feature_maps * 2**(d+1)
|
| 366 |
+
elif i == 0 and self.injection_at == "bottom":
|
| 367 |
+
in_size = out_size + self.injection_channels
|
| 368 |
+
else:
|
| 369 |
+
in_size = out_size
|
| 370 |
+
|
| 371 |
+
# check for coord appending at this depth
|
| 372 |
+
if d in self.coords[0] and i == 0 and d < self.depth - 1:
|
| 373 |
+
block.append(ConcatCoords())
|
| 374 |
+
in_size += self.coords_dim
|
| 375 |
+
|
| 376 |
+
block.append(self.conv_op(in_size,
|
| 377 |
+
out_size,
|
| 378 |
+
self.kernel_size,
|
| 379 |
+
padding=self.padding,
|
| 380 |
+
dilation=self.dilation,
|
| 381 |
+
**self.conv_kwargs))
|
| 382 |
+
if self.dropout_op is not None:
|
| 383 |
+
block.append(self.dropout_op(**self.dropout_kwargs))
|
| 384 |
+
if self.norm_op is not None:
|
| 385 |
+
block.append(self.norm_op(out_size, **self.norm_kwargs))
|
| 386 |
+
block.append(self.activation_op(**self.activation_kwargs))
|
| 387 |
+
|
| 388 |
+
if d > 0:
|
| 389 |
+
block.append(self.upconv_op(out_size,
|
| 390 |
+
out_size // 2,
|
| 391 |
+
self.kernel_size,
|
| 392 |
+
2,
|
| 393 |
+
padding=self.padding,
|
| 394 |
+
dilation=self.dilation,
|
| 395 |
+
output_padding=1,
|
| 396 |
+
**self.upconv_kwargs))
|
| 397 |
+
|
| 398 |
+
self.add_module("decode-{}".format(d), nn.Sequential(*block))
|
| 399 |
+
|
| 400 |
+
if self.injection_at == "end":
|
| 401 |
+
out_size += self.injection_channels
|
| 402 |
+
in_size = out_size
|
| 403 |
+
for i in range(self.num_1x1_at_end):
|
| 404 |
+
if i == self.num_1x1_at_end - 1:
|
| 405 |
+
out_size = self.out_channels
|
| 406 |
+
current_conv_kwargs = self.conv_kwargs.copy()
|
| 407 |
+
current_conv_kwargs["bias"] = True
|
| 408 |
+
self.add_module("reduce-{}".format(i), self.conv_op(in_size, out_size, 1, **current_conv_kwargs))
|
| 409 |
+
if i != self.num_1x1_at_end - 1:
|
| 410 |
+
self.add_module("reduce-{}-nonlin".format(i), self.activation_op(**self.activation_kwargs))
|
| 411 |
+
if self.output_activation_op is not None:
|
| 412 |
+
self.add_module("output-activation", self.output_activation_op(**self.output_activation_kwargs))
|
| 413 |
+
|
| 414 |
+
def reset(self):
|
| 415 |
+
|
| 416 |
+
self.last_activations = None
|
| 417 |
+
|
| 418 |
+
def forward(self, x, injection=None, reuse_last_activations=False, store_activations=False):
|
| 419 |
+
|
| 420 |
+
if self.injection_at == "bottom": # not worth it for now
|
| 421 |
+
reuse_last_activations = False
|
| 422 |
+
store_activations = False
|
| 423 |
+
|
| 424 |
+
if self.last_activations is None or reuse_last_activations is False:
|
| 425 |
+
|
| 426 |
+
enc = [x]
|
| 427 |
+
|
| 428 |
+
for i in range(self.depth - 1):
|
| 429 |
+
enc.append(self._modules["encode-{}".format(i)](enc[-1]))
|
| 430 |
+
|
| 431 |
+
bottom_rep = self._modules["encode-{}".format(self.depth - 1)](enc[-1])
|
| 432 |
+
|
| 433 |
+
if self.injection_at == "bottom" and self.injection_channels > 0:
|
| 434 |
+
injection = match_to(injection, bottom_rep, (0, 1))
|
| 435 |
+
bottom_rep = torch.cat((bottom_rep, injection), 1)
|
| 436 |
+
|
| 437 |
+
x = self._modules["decode-{}".format(self.depth - 1)](bottom_rep)
|
| 438 |
+
|
| 439 |
+
for i in reversed(range(self.depth - 1)):
|
| 440 |
+
x = self._modules["decode-{}".format(i)](torch.cat((enc[-(self.depth - 1 - i)], x), 1))
|
| 441 |
+
|
| 442 |
+
if store_activations:
|
| 443 |
+
self.last_activations = x.detach()
|
| 444 |
+
|
| 445 |
+
else:
|
| 446 |
+
|
| 447 |
+
x = self.last_activations
|
| 448 |
+
|
| 449 |
+
if self.injection_at == "end" and self.injection_channels > 0:
|
| 450 |
+
injection = match_to(injection, x, (0, 1))
|
| 451 |
+
x = torch.cat((x, injection), 1)
|
| 452 |
+
|
| 453 |
+
for i in range(self.num_1x1_at_end):
|
| 454 |
+
x = self._modules["reduce-{}".format(i)](x)
|
| 455 |
+
if self.output_activation_op is not None:
|
| 456 |
+
x = self._modules["output-activation"](x)
|
| 457 |
+
|
| 458 |
+
if self.return_bottom and not reuse_last_activations:
|
| 459 |
+
return x, bottom_rep
|
| 460 |
+
else:
|
| 461 |
+
return x
|
| 462 |
+
|
| 463 |
+
|
| 464 |
+
|
| 465 |
+
class InjectionUNet3D(InjectionUNet):
|
| 466 |
+
|
| 467 |
+
def __init__(self, *args, **kwargs):
|
| 468 |
+
|
| 469 |
+
update_kwargs = dict(
|
| 470 |
+
pool_op=nn.AvgPool3d,
|
| 471 |
+
norm_op=nn.InstanceNorm3d,
|
| 472 |
+
conv_op=nn.Conv3d,
|
| 473 |
+
upconv_op=nn.ConvTranspose3d,
|
| 474 |
+
coords_dim=3
|
| 475 |
+
)
|
| 476 |
+
|
| 477 |
+
for (arg, val) in update_kwargs.items():
|
| 478 |
+
if arg not in kwargs: kwargs[arg] = val
|
| 479 |
+
|
| 480 |
+
super(InjectionUNet3D, self).__init__(*args, **kwargs)
|
| 481 |
+
|
| 482 |
+
class InjectionUNet2D(InjectionUNet): #Created by Soumick
|
| 483 |
+
|
| 484 |
+
def __init__(self, *args, **kwargs):
|
| 485 |
+
|
| 486 |
+
update_kwargs = dict(
|
| 487 |
+
pool_op=nn.AvgPool2d,
|
| 488 |
+
norm_op=nn.InstanceNorm2d,
|
| 489 |
+
conv_op=nn.Conv2d,
|
| 490 |
+
upconv_op=nn.ConvTranspose2d,
|
| 491 |
+
coords_dim=2
|
| 492 |
+
)
|
| 493 |
+
|
| 494 |
+
for (arg, val) in update_kwargs.items():
|
| 495 |
+
if arg not in kwargs: kwargs[arg] = val
|
| 496 |
+
|
| 497 |
+
super(InjectionUNet2D, self).__init__(*args, **kwargs)
|
| 498 |
+
|
| 499 |
+
class ProbabilisticSegmentationNet(ConvModule):
|
| 500 |
+
|
| 501 |
+
def __init__(self,
|
| 502 |
+
in_channels=4,
|
| 503 |
+
out_channels=4,
|
| 504 |
+
num_feature_maps=24,
|
| 505 |
+
latent_size=3,
|
| 506 |
+
depth=5,
|
| 507 |
+
latent_distribution=torch.distributions.Normal,
|
| 508 |
+
task_op=InjectionUNet3D,
|
| 509 |
+
task_kwargs=None,
|
| 510 |
+
prior_op=InjectionConvEncoder3D,
|
| 511 |
+
prior_kwargs=None,
|
| 512 |
+
posterior_op=InjectionConvEncoder3D,
|
| 513 |
+
posterior_kwargs=None,
|
| 514 |
+
**kwargs):
|
| 515 |
+
|
| 516 |
+
super(ProbabilisticSegmentationNet, self).__init__(**kwargs)
|
| 517 |
+
|
| 518 |
+
self.task_op = task_op
|
| 519 |
+
self.task_kwargs = {} if task_kwargs is None else task_kwargs
|
| 520 |
+
self.prior_op = prior_op
|
| 521 |
+
self.prior_kwargs = {} if prior_kwargs is None else prior_kwargs
|
| 522 |
+
self.posterior_op = posterior_op
|
| 523 |
+
self.posterior_kwargs = {} if posterior_kwargs is None else posterior_kwargs
|
| 524 |
+
|
| 525 |
+
default_task_kwargs = dict(
|
| 526 |
+
in_channels=in_channels,
|
| 527 |
+
out_channels=out_channels,
|
| 528 |
+
num_feature_maps=num_feature_maps,
|
| 529 |
+
injection_size=latent_size,
|
| 530 |
+
depth=depth
|
| 531 |
+
)
|
| 532 |
+
|
| 533 |
+
default_prior_kwargs = dict(
|
| 534 |
+
in_channels=in_channels,
|
| 535 |
+
out_channels=latent_size*2, #Soumick
|
| 536 |
+
num_feature_maps=num_feature_maps,
|
| 537 |
+
z_dim=latent_size,
|
| 538 |
+
depth=depth
|
| 539 |
+
)
|
| 540 |
+
|
| 541 |
+
default_posterior_kwargs = dict(
|
| 542 |
+
in_channels=in_channels+out_channels,
|
| 543 |
+
out_channels=latent_size*2, #Soumick
|
| 544 |
+
num_feature_maps=num_feature_maps,
|
| 545 |
+
z_dim=latent_size,
|
| 546 |
+
depth=depth
|
| 547 |
+
)
|
| 548 |
+
|
| 549 |
+
default_task_kwargs.update(self.task_kwargs)
|
| 550 |
+
self.task_kwargs = default_task_kwargs
|
| 551 |
+
default_prior_kwargs.update(self.prior_kwargs)
|
| 552 |
+
self.prior_kwargs = default_prior_kwargs
|
| 553 |
+
default_posterior_kwargs.update(self.posterior_kwargs)
|
| 554 |
+
self.posterior_kwargs = default_posterior_kwargs
|
| 555 |
+
|
| 556 |
+
self.latent_distribution = latent_distribution
|
| 557 |
+
self._prior = None
|
| 558 |
+
self._posterior = None
|
| 559 |
+
|
| 560 |
+
self.make_modules()
|
| 561 |
+
|
| 562 |
+
def make_modules(self):
|
| 563 |
+
|
| 564 |
+
if type(self.task_op) == type:
|
| 565 |
+
self.add_module("task_net", self.task_op(**self.task_kwargs))
|
| 566 |
+
else:
|
| 567 |
+
self.add_module("task_net", self.task_op)
|
| 568 |
+
if type(self.prior_op) == type:
|
| 569 |
+
self.add_module("prior_net", self.prior_op(**self.prior_kwargs))
|
| 570 |
+
else:
|
| 571 |
+
self.add_module("prior_net", self.prior_op)
|
| 572 |
+
if type(self.posterior_op) == type:
|
| 573 |
+
self.add_module("posterior_net", self.posterior_op(**self.posterior_kwargs))
|
| 574 |
+
else:
|
| 575 |
+
self.add_module("posterior_net", self.posterior_op)
|
| 576 |
+
|
| 577 |
+
@property
|
| 578 |
+
def prior(self):
|
| 579 |
+
return self._prior
|
| 580 |
+
|
| 581 |
+
@property
|
| 582 |
+
def posterior(self):
|
| 583 |
+
return self._posterior
|
| 584 |
+
|
| 585 |
+
@property
|
| 586 |
+
def last_activations(self):
|
| 587 |
+
return self.task_net.last_activations
|
| 588 |
+
|
| 589 |
+
def train(self, mode=True):
|
| 590 |
+
|
| 591 |
+
super(ProbabilisticSegmentationNet, self).train(mode)
|
| 592 |
+
self.reset()
|
| 593 |
+
|
| 594 |
+
def reset(self):
|
| 595 |
+
|
| 596 |
+
self.task_net.reset()
|
| 597 |
+
self._prior = None
|
| 598 |
+
self._posterior = None
|
| 599 |
+
|
| 600 |
+
def forward(self, input_, seg=None, make_onehot=True, make_onehot_classes=None, newaxis=False, distlossN=0):
|
| 601 |
+
"""Forward pass includes reparametrization sampling during training, otherwise it'll just take the prior mean."""
|
| 602 |
+
|
| 603 |
+
self.encode_prior(input_)
|
| 604 |
+
|
| 605 |
+
if distlossN == 0:
|
| 606 |
+
if self.training:
|
| 607 |
+
self.encode_posterior(input_, seg, make_onehot, make_onehot_classes, newaxis)
|
| 608 |
+
sample = self.posterior.rsample()
|
| 609 |
+
else:
|
| 610 |
+
sample = self.prior.loc
|
| 611 |
+
return self.task_net(input_, sample, store_activations=not self.training)
|
| 612 |
+
else:
|
| 613 |
+
if self.training:
|
| 614 |
+
self.encode_posterior(input_, seg, make_onehot, make_onehot_classes, newaxis)
|
| 615 |
+
segs = []
|
| 616 |
+
for i in range(distlossN):
|
| 617 |
+
sample = self.posterior.rsample()
|
| 618 |
+
segs.append(self.task_net(input_, sample, store_activations=not self.training))
|
| 619 |
+
return segs #torch.concat(segs, dim=0)
|
| 620 |
+
else: #I'm not totally sure about this!!
|
| 621 |
+
sample = self.prior.loc
|
| 622 |
+
return self.task_net(input_, sample, store_activations=not self.training)
|
| 623 |
+
|
| 624 |
+
|
| 625 |
+
def encode_prior(self, input_):
|
| 626 |
+
|
| 627 |
+
rep = self.prior_net(input_)
|
| 628 |
+
if isinstance(rep, tuple):
|
| 629 |
+
mean, logvar = rep
|
| 630 |
+
elif torch.is_tensor(rep):
|
| 631 |
+
mean, logvar = torch.split(rep, rep.shape[1] // 2, dim=1)
|
| 632 |
+
self._prior = self.latent_distribution(mean, logvar.mul(0.5).exp())
|
| 633 |
+
return self._prior
|
| 634 |
+
|
| 635 |
+
def encode_posterior(self, input_, seg, make_onehot=True, make_onehot_classes=None, newaxis=False):
|
| 636 |
+
|
| 637 |
+
if make_onehot:
|
| 638 |
+
if make_onehot_classes is None:
|
| 639 |
+
make_onehot_classes = tuple(range(self.posterior_net.in_channels - input_.shape[1]))
|
| 640 |
+
seg = make_onehot_segmentation(seg, make_onehot_classes, newaxis=newaxis)
|
| 641 |
+
rep = self.posterior_net(torch.cat((input_, seg.float()), 1))
|
| 642 |
+
if isinstance(rep, tuple):
|
| 643 |
+
mean, logvar = rep
|
| 644 |
+
elif torch.is_tensor(rep):
|
| 645 |
+
mean, logvar = torch.split(rep, rep.shape[1] // 2, dim=1)
|
| 646 |
+
self._posterior = self.latent_distribution(mean, logvar.mul(0.5).exp())
|
| 647 |
+
return self._posterior
|
| 648 |
+
|
| 649 |
+
def sample_prior(self, N=1, out_device=None, input_=None, pred_with_mean=False):
|
| 650 |
+
"""Draw multiple samples from the current prior.
|
| 651 |
+
|
| 652 |
+
* input_ is required if no activations are stored in task_net.
|
| 653 |
+
* If input_ is given, prior will automatically be encoded again.
|
| 654 |
+
* Returns either a single sample or a list of samples.
|
| 655 |
+
|
| 656 |
+
"""
|
| 657 |
+
|
| 658 |
+
if out_device is None:
|
| 659 |
+
if self.last_activations is not None:
|
| 660 |
+
out_device = self.last_activations.device
|
| 661 |
+
elif input_ is not None:
|
| 662 |
+
out_device = input_.device
|
| 663 |
+
else:
|
| 664 |
+
out_device = next(self.task_net.parameters()).device
|
| 665 |
+
with torch.no_grad():
|
| 666 |
+
if self.prior is None or input_ is not None:
|
| 667 |
+
self.encode_prior(input_)
|
| 668 |
+
result = []
|
| 669 |
+
|
| 670 |
+
if input_ is not None:
|
| 671 |
+
result.append(self.task_net(input_, self.prior.sample(), reuse_last_activations=False, store_activations=True).to(device=out_device))
|
| 672 |
+
while len(result) < N:
|
| 673 |
+
result.append(self.task_net(input_,
|
| 674 |
+
self.prior.sample(),
|
| 675 |
+
reuse_last_activations=self.last_activations is not None,
|
| 676 |
+
store_activations=False).to(device=out_device))
|
| 677 |
+
if pred_with_mean:
|
| 678 |
+
result.append(self.task_net(input_, self.prior.mean, reuse_last_activations=False, store_activations=True).to(device=out_device))
|
| 679 |
+
|
| 680 |
+
if len(result) == 1:
|
| 681 |
+
return result[0]
|
| 682 |
+
else:
|
| 683 |
+
return result
|
| 684 |
+
|
| 685 |
+
def reconstruct(self, sample=None, use_posterior_mean=True, out_device=None, input_=None):
|
| 686 |
+
"""Reconstruct a sample or the current posterior mean. Will not compute gradients!"""
|
| 687 |
+
|
| 688 |
+
if self.posterior is None and sample is None:
|
| 689 |
+
raise ValueError("'posterior' is currently None. Please pass an input and a segmentation first.")
|
| 690 |
+
if out_device is None:
|
| 691 |
+
out_device = next(self.task_net.parameters()).device
|
| 692 |
+
if sample is None:
|
| 693 |
+
if use_posterior_mean:
|
| 694 |
+
sample = self.posterior.loc
|
| 695 |
+
else:
|
| 696 |
+
sample = self.posterior.sample()
|
| 697 |
+
else:
|
| 698 |
+
sample = sample.to(next(self.task_net.parameters()).device)
|
| 699 |
+
with torch.no_grad():
|
| 700 |
+
return self.task_net(input_, sample, reuse_last_activations=True).to(device=out_device)
|
| 701 |
+
|
| 702 |
+
def kl_divergence(self):
|
| 703 |
+
"""Compute current KL, requires existing prior and posterior."""
|
| 704 |
+
|
| 705 |
+
if self.posterior is None or self.prior is None:
|
| 706 |
+
raise ValueError("'prior' and 'posterior' must not be None, but prior={} and posterior={}".format(self.prior, self.posterior))
|
| 707 |
+
return torch.distributions.kl_divergence(self.posterior, self.prior).sum()
|
| 708 |
+
|
| 709 |
+
def elbo(self, seg, input_=None, nll_reduction="sum", beta=1.0, make_onehot=True, make_onehot_classes=None, newaxis=False):
|
| 710 |
+
"""Compute the ELBO with seg as ground truth.
|
| 711 |
+
|
| 712 |
+
* Prior is expected and will not be encoded.
|
| 713 |
+
* If input_ is given, posterior will automatically be encoded.
|
| 714 |
+
* Either input_ or stored activations must be available.
|
| 715 |
+
|
| 716 |
+
"""
|
| 717 |
+
|
| 718 |
+
if self.last_activations is None:
|
| 719 |
+
raise ValueError("'last_activations' is currently None. Please pass an input first.")
|
| 720 |
+
if input_ is not None:
|
| 721 |
+
with torch.no_grad():
|
| 722 |
+
self.encode_posterior(input_, seg, make_onehot=make_onehot, make_onehot_classes=make_onehot_classes, newaxis=newaxis)
|
| 723 |
+
if make_onehot and newaxis:
|
| 724 |
+
pass # seg will already be (B x SPACE)
|
| 725 |
+
elif make_onehot and not newaxis:
|
| 726 |
+
seg = seg[:, 0] # in this case seg will hopefully be (B x 1 x SPACE)
|
| 727 |
+
else:
|
| 728 |
+
seg = torch.argmax(seg, 1, keepdim=False) # seg is already onehot
|
| 729 |
+
kl = self.kl_divergence()
|
| 730 |
+
nll = nn.NLLLoss(reduction=nll_reduction)(self.reconstruct(sample=None, use_posterior_mean=True, out_device=None), seg.long())
|
| 731 |
+
return - (beta * nll + kl)
|
ProbUNet_utils.py
ADDED
|
@@ -0,0 +1,224 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import os
|
| 3 |
+
import shutil
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
# from trixi.util import Config, GridSearch
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
def check_attributes(object_, attributes):
|
| 10 |
+
|
| 11 |
+
missing = []
|
| 12 |
+
for attr in attributes:
|
| 13 |
+
if not hasattr(object_, attr):
|
| 14 |
+
missing.append(attr)
|
| 15 |
+
if len(missing) > 0:
|
| 16 |
+
return False
|
| 17 |
+
else:
|
| 18 |
+
return True
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def set_seeds(seed, cuda=True):
|
| 22 |
+
if not hasattr(seed, "__iter__"):
|
| 23 |
+
seed = (seed, seed, seed)
|
| 24 |
+
np.random.seed(seed[0])
|
| 25 |
+
torch.manual_seed(seed[1])
|
| 26 |
+
if cuda: torch.cuda.manual_seed_all(seed[2])
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def make_onehot(array, labels=None, axis=1, newaxis=False):
|
| 30 |
+
|
| 31 |
+
# get labels if necessary
|
| 32 |
+
if labels is None:
|
| 33 |
+
labels = np.unique(array)
|
| 34 |
+
labels = list(map(lambda x: x.item(), labels))
|
| 35 |
+
|
| 36 |
+
# get target shape
|
| 37 |
+
new_shape = list(array.shape)
|
| 38 |
+
if newaxis:
|
| 39 |
+
new_shape.insert(axis, len(labels))
|
| 40 |
+
else:
|
| 41 |
+
new_shape[axis] = new_shape[axis] * len(labels)
|
| 42 |
+
|
| 43 |
+
# make zero array
|
| 44 |
+
if type(array) == np.ndarray:
|
| 45 |
+
new_array = np.zeros(new_shape, dtype=array.dtype)
|
| 46 |
+
elif torch.is_tensor(array):
|
| 47 |
+
new_array = torch.zeros(new_shape, dtype=array.dtype, device=array.device)
|
| 48 |
+
else:
|
| 49 |
+
raise TypeError("Onehot conversion undefined for object of type {}".format(type(array)))
|
| 50 |
+
|
| 51 |
+
# fill new array
|
| 52 |
+
n_seg_channels = 1 if newaxis else array.shape[axis]
|
| 53 |
+
for seg_channel in range(n_seg_channels):
|
| 54 |
+
for l, label in enumerate(labels):
|
| 55 |
+
new_slc = [slice(None), ] * len(new_shape)
|
| 56 |
+
slc = [slice(None), ] * len(array.shape)
|
| 57 |
+
new_slc[axis] = seg_channel * len(labels) + l
|
| 58 |
+
if not newaxis:
|
| 59 |
+
slc[axis] = seg_channel
|
| 60 |
+
new_array[tuple(new_slc)] = array[tuple(slc)] == label
|
| 61 |
+
|
| 62 |
+
return new_array
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def match_to(x, ref, keep_axes=(1,)):
|
| 66 |
+
|
| 67 |
+
target_shape = list(ref.shape)
|
| 68 |
+
for i in keep_axes:
|
| 69 |
+
target_shape[i] = x.shape[i]
|
| 70 |
+
target_shape = tuple(target_shape)
|
| 71 |
+
if x.shape == target_shape:
|
| 72 |
+
pass
|
| 73 |
+
if x.dim() == 1:
|
| 74 |
+
x = x.unsqueeze(0)
|
| 75 |
+
if x.dim() == 2:
|
| 76 |
+
while x.dim() < len(target_shape):
|
| 77 |
+
x = x.unsqueeze(-1)
|
| 78 |
+
|
| 79 |
+
x = x.expand(*target_shape)
|
| 80 |
+
x = x.to(device=ref.device, dtype=ref.dtype)
|
| 81 |
+
|
| 82 |
+
return x
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
def make_slices(original_shape, patch_shape):
|
| 86 |
+
|
| 87 |
+
working_shape = original_shape[-len(patch_shape):]
|
| 88 |
+
splits = []
|
| 89 |
+
for i in range(len(working_shape)):
|
| 90 |
+
splits.append([])
|
| 91 |
+
for j in range(working_shape[i] // patch_shape[i]):
|
| 92 |
+
splits[i].append(slice(j*patch_shape[i], (j+1)*patch_shape[i]))
|
| 93 |
+
rest = working_shape[i] % patch_shape[i]
|
| 94 |
+
if rest > 0:
|
| 95 |
+
splits[i].append(slice((j+1)*patch_shape[i], (j+1)*patch_shape[i] + rest))
|
| 96 |
+
|
| 97 |
+
# now we have all slices for the individual dimensions
|
| 98 |
+
# we need their combinatorial combinations
|
| 99 |
+
slices = list(itertools.product(*splits))
|
| 100 |
+
for i in range(len(slices)):
|
| 101 |
+
slices[i] = [slice(None), ] * (len(original_shape) - len(patch_shape)) + list(slices[i])
|
| 102 |
+
|
| 103 |
+
return slices
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def coordinate_grid_samples(mean, std, factor_std=5, scale_std=1.):
|
| 107 |
+
|
| 108 |
+
relative = np.linspace(-scale_std*factor_std, scale_std*factor_std, 2*factor_std+1)
|
| 109 |
+
positions = np.array([mean + i * std for i in relative]).T
|
| 110 |
+
axes = np.meshgrid(*positions)
|
| 111 |
+
axes = map(lambda x: list(x.ravel()), axes)
|
| 112 |
+
samples = list(zip(*axes))
|
| 113 |
+
samples = list(map(np.array, samples))
|
| 114 |
+
|
| 115 |
+
return samples
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
def get_default_experiment_parser():
|
| 119 |
+
|
| 120 |
+
parser = argparse.ArgumentParser()
|
| 121 |
+
parser.add_argument("base_dir", type=str, help="Working directory for experiment.")
|
| 122 |
+
parser.add_argument("-c", "--config", type=str, default=None, help="Path to a config file.")
|
| 123 |
+
parser.add_argument("-v", "--visdomlogger", action="store_true", help="Use visdomlogger.")
|
| 124 |
+
parser.add_argument("-tx", "--tensorboardxlogger", type=str, default=None)
|
| 125 |
+
parser.add_argument("-tl", "--telegramlogger", action="store_true")
|
| 126 |
+
parser.add_argument("-dc", "--default_config", type=str, default="DEFAULTS", help="Select a default Config")
|
| 127 |
+
parser.add_argument("-ad", "--automatic_description", action="store_true")
|
| 128 |
+
parser.add_argument("-r", "--resume", type=str, default=None, help="Path to resume from")
|
| 129 |
+
parser.add_argument("-irc", "--ignore_resume_config", action="store_true", help="Ignore Config in experiment we resume from.")
|
| 130 |
+
parser.add_argument("-test", "--test", action="store_true", help="Run test instead of training")
|
| 131 |
+
parser.add_argument("-g", "--grid", type=str, help="Path to a config for grid search")
|
| 132 |
+
parser.add_argument("-s", "--skip_existing", action="store_true", help="Skip configs for which an experiment exists, only for grid search")
|
| 133 |
+
parser.add_argument("-m", "--mods", type=str, nargs="+", default=None, help="Mods are Config stubs to update only relevant parts for a certain setup.")
|
| 134 |
+
parser.add_argument("-ct", "--copy_test", action="store_true", help="Copy test files to original experiment.")
|
| 135 |
+
|
| 136 |
+
return parser
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
def run_experiment(experiment, configs, args, mods=None, **kwargs):
|
| 140 |
+
|
| 141 |
+
# set a few defaults
|
| 142 |
+
if "explogger_kwargs" not in kwargs:
|
| 143 |
+
kwargs["explogger_kwargs"] = dict(folder_format="{experiment_name}_%Y%m%d-%H%M%S")
|
| 144 |
+
if "explogger_freq" not in kwargs:
|
| 145 |
+
kwargs["explogger_freq"] = 1
|
| 146 |
+
if "resume_save_types" not in kwargs:
|
| 147 |
+
kwargs["resume_save_types"] = ("model", "simple", "th_vars", "results")
|
| 148 |
+
|
| 149 |
+
config = Config(file_=args.config) if args.config is not None else Config()
|
| 150 |
+
config.update_missing(configs[args.default_config].deepcopy())
|
| 151 |
+
if args.mods is not None and mods is not None:
|
| 152 |
+
for mod in args.mods:
|
| 153 |
+
config.update(mods[mod])
|
| 154 |
+
config = Config(config=config, update_from_argv=True)
|
| 155 |
+
|
| 156 |
+
# GET EXISTING EXPERIMENTS TO BE ABLE TO SKIP CERTAIN CONFIGS
|
| 157 |
+
if args.skip_existing:
|
| 158 |
+
existing_configs = []
|
| 159 |
+
for exp in os.listdir(args.base_dir):
|
| 160 |
+
try:
|
| 161 |
+
existing_configs.append(Config(file_=os.path.join(args.base_dir, exp, "config", "config.json")))
|
| 162 |
+
except Exception as e:
|
| 163 |
+
pass
|
| 164 |
+
|
| 165 |
+
if args.grid is not None:
|
| 166 |
+
grid = GridSearch().read(args.grid)
|
| 167 |
+
else:
|
| 168 |
+
grid = [{}]
|
| 169 |
+
|
| 170 |
+
for combi in grid:
|
| 171 |
+
|
| 172 |
+
config.update(combi)
|
| 173 |
+
|
| 174 |
+
if args.skip_existing:
|
| 175 |
+
skip_this = False
|
| 176 |
+
for existing_config in existing_configs:
|
| 177 |
+
if existing_config.contains(config):
|
| 178 |
+
skip_this = True
|
| 179 |
+
break
|
| 180 |
+
if skip_this:
|
| 181 |
+
continue
|
| 182 |
+
|
| 183 |
+
if "backup_every" in config:
|
| 184 |
+
kwargs["save_checkpoint_every_epoch"] = config["backup_every"]
|
| 185 |
+
|
| 186 |
+
loggers = {}
|
| 187 |
+
if args.visdomlogger:
|
| 188 |
+
loggers["v"] = ("visdom", {}, 1)
|
| 189 |
+
if args.tensorboardxlogger is not None:
|
| 190 |
+
if args.tensorboardxlogger == "same":
|
| 191 |
+
loggers["tx"] = ("tensorboard", {}, 1)
|
| 192 |
+
else:
|
| 193 |
+
loggers["tx"] = ("tensorboard", {"target_dir": args.tensorboardxlogger}, 1)
|
| 194 |
+
|
| 195 |
+
if args.telegramlogger:
|
| 196 |
+
kwargs["use_telegram"] = True
|
| 197 |
+
|
| 198 |
+
if args.automatic_description:
|
| 199 |
+
difference_to_default = Config.difference_config_static(config, configs["DEFAULTS"]).flat(keep_lists=True, max_split_size=0, flatten_int=True)
|
| 200 |
+
description_str = ""
|
| 201 |
+
for key, val in difference_to_default.items():
|
| 202 |
+
val = val[0]
|
| 203 |
+
description_str = "{} = {}\n{}".format(key, val, description_str)
|
| 204 |
+
config.description = description_str
|
| 205 |
+
|
| 206 |
+
exp = experiment(config=config,
|
| 207 |
+
base_dir=args.base_dir,
|
| 208 |
+
resume=args.resume,
|
| 209 |
+
ignore_resume_config=args.ignore_resume_config,
|
| 210 |
+
loggers=loggers,
|
| 211 |
+
**kwargs)
|
| 212 |
+
|
| 213 |
+
trained = False
|
| 214 |
+
if args.resume is None or args.test is False:
|
| 215 |
+
exp.run()
|
| 216 |
+
trained = True
|
| 217 |
+
if args.test:
|
| 218 |
+
exp.run_test(setup=not trained)
|
| 219 |
+
if isinstance(args.resume, str) and exp.elog is not None and args.copy_test:
|
| 220 |
+
for f in glob.glob(os.path.join(exp.elog.save_dir, "test*")):
|
| 221 |
+
if os.path.isdir(f):
|
| 222 |
+
shutil.copytree(f, os.path.join(args.resume, "save", os.path.basename(f)))
|
| 223 |
+
else:
|
| 224 |
+
shutil.copy(f, os.path.join(args.resume, "save"))
|
config.json
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"ProbUNet"
|
| 4 |
+
],
|
| 5 |
+
"auto_map": {
|
| 6 |
+
"AutoConfig": "PULASkiConfigs.ProbUNetConfig",
|
| 7 |
+
"AutoModel": "PULASki.ProbUNet"
|
| 8 |
+
},
|
| 9 |
+
"depth": 5,
|
| 10 |
+
"dim": 3,
|
| 11 |
+
"in_channels": 1,
|
| 12 |
+
"latent_distribution": "normal",
|
| 13 |
+
"latent_size": 3,
|
| 14 |
+
"model_type": "ProbUNet",
|
| 15 |
+
"no_outact_op": false,
|
| 16 |
+
"num_feature_maps": 24,
|
| 17 |
+
"out_channels": 1,
|
| 18 |
+
"prob_injection_at": "end",
|
| 19 |
+
"torch_dtype": "float32",
|
| 20 |
+
"transformers_version": "4.44.2"
|
| 21 |
+
}
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8b8206054d49de1eb5ac7ca20380daa244abc079237a812db1bf3211e8d20d60
|
| 3 |
+
size 121870000
|