Readme

losses.py

@@ -9,6 +9,11 @@
 class Partial_CE(torch.autograd.Function):
     def forward(self, input, target, weakLabels):
         self.save_for_backward(input, target, weakLabels)
+        # b, c, w, h = input.shape
+        # assert target.shape == input.shape
+        # assert weakLabels.shape == (b, 1, w, h)
+
+        # assert np.allclose(input[:, 0, ...].cpu().numpy(), 1 - input[:, 1, ...].cpu().numpy(), atol=1e-2)
 
         eps = 1e-20
 
@@ -25,8 +30,6 @@ def forward(self, input, target, weakLabels):
         lossT = torch.FloatTensor(1)
         lossT.fill_(np.float32(loss).item())
 
-        lossT = lossT.cuda()
-
         return lossT.cuda()   # a single number (averaged loss over batch samples)
 
     def backward(self, grad_output):

main_MIDL.py

@@ -4,13 +4,12 @@
 
 import torch
 import numpy as np
-import medicalDataLoader
 from torch import nn
 from torch.autograd import Variable
 from torch.utils.data import DataLoader
 from torchvision import transforms
 
-
+import medicalDataLoader
 from ENet import ENet
 from utils import to_var
 from utils import computeDiceOneHotBinary, predToSegmentation, inference, DicesToDice, printProgressBar

medicalDataLoader.py

@@ -28,7 +28,7 @@ def make_dataset(root, mode):
         labels.sort()
         labels_weak.sort()
 
-        for it_im, it_gt, it_w in zip(images, labels,labels_weak):
+        for it_im, it_gt, it_w in zip(images, labels, labels_weak):
             item = (os.path.join(train_img_path, it_im), os.path.join(train_mask_path, it_gt), os.path.join(train_mask_weak_path, it_w))
             items.append(item)
 
@@ -45,7 +45,7 @@ def make_dataset(root, mode):
         labels.sort()
         labels_weak.sort()
 
-        for it_im, it_gt, it_w in zip(images, labels,labels_weak):
+        for it_im, it_gt, it_w in zip(images, labels, labels_weak):
             item = (os.path.join(train_img_path, it_im), os.path.join(train_mask_path, it_gt), os.path.join(train_mask_weak_path, it_w))
             items.append(item)
     else:

readme.md

@@ -0,0 +1,23 @@
+# Constrained-CNN losses for weakly supervised segmentation
+Code of our submission https://openreview.net/forum?id=BkIBHb2sG at MIDL 2018
+
+To run it, simply run `main_MIDL.py` (python3.6+, the requirements are specified in the `requirements.txt` file).
+
+The partial ground truth that we used are provided, but not the original dataset: https://www.creatis.insa-lyon.fr/Challenge/acdc/databases.html
+You will need to download and pre-process it yourselves first.
+
+The code was developed for PyTorch 0.3.1, and has been modified slightly to work with PyTorch 0.4. However, a lot of cleanup (removing the variables for instance) still need to be done.
+
+## Loss functions
+The loss functions are located in `losses.py`, and are defined as autograd functions. We implemented manually both the forward and the backward passes with numpy. We use a batch size of 1, and the code might need to be modified before working for more.
+
+The inputs (predictions, labels and weak labels) are all represented as 4-D tensors:
+```python
+b, c, w, h = input.shape
+assert target.shape == input.shape
+assert weakLabels.shape == (b, 1, w, h)
+```
+`b` is the batch size, `c` the number of classes (2), and `w, h` the image size. Since this is a binary problem, the two classes are complementary (minus the rounding errors), both for the predictions and labels:
+```python
+assert np.allclose(input[:, 0, ...].cpu().numpy(), 1 - input[:, 1, ...].cpu().numpy(), atol=1e-2)
+```

requirements.txt

@@ -0,0 +1,3 @@
+numpy
+pytorch>=0.4
+torchvision