HKervadec
commented
5 years ago (1) I am not used to Tensorflow or Keras, so is it difficult for me to help you. Especially since the paradigm is slightly different (where stuff runs in a session). Are you sure the two are on the same device ? The _device variable seems different in your error message.
As you can see, the results of one_hot2dist will be the same shape as its input. But you could add an assertion at the end just to be sure:
res = np.zeros_like(seg)
...
assert res.shape == seg.shape
return resIf this pass, you should also make sure that y_true and y_pred have the same shape (guess you will have to use tf.assert for that, don't know the details). Other than that, I think it would be a problem different than the shape. Perhaps because the y_pred are floats, while the distance maps are doubles (I think NumPy defaults to float64 and not float32 as many deep learning frameworks).
(2): Yes it is not obvious at first. I actually give a list of losses to the main script, which are then summed:
def setup(...):
...
losses = eval(args.losses)
loss_fns: List[Callable] = []
for loss_name, loss_params, _, _, fn, _ in losses:
loss_class = getattr(__import__('losses'), loss_name)
loss_fns.append(loss_class(**loss_params, fn=fn))
loss_weights: List[float] = map_(itemgetter(5), losses)
...
def do_epoch(...):
...
for j, data in enumerate(loader):
...
# Forward
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
assert len(bounds) == len(loss_fns) == len(loss_weights)
ziped = zip(loss_fns, labels, loss_weights, bounds)
# Apply each loss, creating a list of loss values
losses = [w * loss_fn(pred_probs, label, bound) for loss_fn, label, w, bound in ziped]
# Then sum them
loss = reduce(add, losses)
assert loss.shape == (), loss.shape
# Backward
if optimizer:
loss.backward()
optimizer.step()
...
...While this is more complicated at first, it gives much more flexibility then to experiment and avoid code duplication.
(3): I update the loss weights in the schedulers:
optimizer, loss_fns, loss_weights = scheduler(i, optimizer, loss_fns, loss_weights)where scheduler is one of the strategies defined in the scheduler.py file. There is --schedule, --scheduler, and --scheduler_params to define what you wan to use. Many examples of options can be found in the makefile
(4) Sigmoid or softmax, that is actually a hyper-parameter (just as the network architecture and the activation functions inside the network). So for some problem one will be better, and on other tasks it will be the opposite. The main difference will be wrt the gradient of the loss.
In addition, I define the SurfaceLoss by 'def ', rather than by 'class' . Is this different?
No, that is the same. I use a class (where I define the __call__ function) so I can easily add parameters to the loss (which is the case for some of my other projects). But in this case, and function and a callable object are effectively the same.
WMeixiang
AmericaBG
Thanks for sharing a very good idea. I am interested in the surface loss function.
(1): In the SurfaceLoss(), the dist_maps is Tensor, while dist_maps is from
def one_hot2dist(seg: np.ndarray) -> np.ndarray:， the format of input and output is ndarray, how to convert the format of data? This question confused me several days, I want to rewrite SurfaceLoss in keras, The following is the code I am rewriting：if I use generalized_dice_loss as cost, my code is ok, but if I use SurfaceLoss, there are some error:
---> I guess the problem is the dimension problem about output, I try to modify, but it always have this problem, could you help me?
(2): your final_loss is formula(6) in your paper midl, combine the generalized_dice_loss and surfaceloss, but I did not find final_loss in your main.py, how to operate?
(3): In formula(6), alpha is dynamic, is related to epoch, how to generate alpha? In your project, I didn't find the definition of function about alpha
(4): For two classifications， the activation function of final layer: softmax, sigmoid. what the difference between softmax and sigmoid? when I used generalized_dice_loss as cost in my code, the sigmoid function perform better than softmax, I have no idea why? Looking forward to your reply.