jcreinhold

#!/usr/bin/env python
from torch import nn
from torch.nn import functional as F
import torch

test_weight_zero = True
test_bias_zero = False
n_steps = 3

############################### 1-layer conv network ###########################

jcreinhold

from typing import Tuple

import torch
from torch import nn
import torch.nn.functional as F

activation = nn.ReLU


class UncertainLinear(nn.Module):

jcreinhold

jcreinhold

jcreinhold

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
3D Tiramisu network for FLAIR-based T2-lesion segmentation

This code is unfortunately a huge mess. But, given the CSV files with
the appropriate setup, you can run the below command (starting with 
"python -u ...") to generate the network used to generate the 
segmentation results in the paper:
"A Structural Causal Model for MR Images of Multiple Sclerosis"

jcreinhold

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
below we calculate and print the stationary dist.
for the transition matrix P with variable p

Note that A is I - P', where P' is P with an
additional row of 1's to represent the condition
that the stationary distribution must add to 1.

jcreinhold

jcreinhold

jcreinhold

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
grid attention blocks for gated attention networks
Based on: https://github.com/ozan-oktay/Attention-Gated-Networks
Author: Jacob Reinhold ([email protected])
"""

__all__ = ['GridAttentionBlock2d',
           'GridAttentionBlock3d']

jcreinhold

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
PyTorch implementation of the Tiramisu network architecture [1]
(2D) Implementation based on [2].

Changes from [2] include:
  1) removal of bias from conv layers,
  2) change zero padding to replication padding,
  3) use of GELU for default activation,