ethanrosenthal / spacecutter Goto Github PK

it's strange for me, why you use deepcopy in there? I have tested the model can work without deepcopy, is there any reason for use it?

line 77: self.predictor = deepcopy(predictor)

RuntimeError: The size of tensor a (5) must match the size of tensor b (6) at non-singleton dimension 1
I'm working on a six classification task[0, 1, 2, 3, 4, 5].I found that cutpoints = num_ class - 1 in the code。 But in this way, the dimensions of cutpoints and X are not equal. I've seen your example, but I still don't know how to solve it.
thank you!

Dear Torch friends

Perhaps I am missing something, but I think the docs of "reduction_" function is not correct. Shall it be?

def _reduction(loss: torch.Tensor, reduction: str) -> torch.Tensor:
    """
    Reduce loss
    Parameters
    ----------
    loss : torch.Tensor, [batch_size, 1]
        Batch losses.
    reduction : str
        Method for reducing the loss. Options include 'elementwise_mean',
        'none', and 'sum'.
    Returns
    -------
    loss : torch.Tensor
        Reduced loss.
    """
    if reduction == 'elementwise_mean':
        return loss.mean()
    elif reduction == 'none':
        return loss
    elif reduction == 'sum':
        return loss.sum()
    else:
        raise ValueError(f'{reduction} is not a valid reduction')

?

Hello,
Firstly, thanks a lot for making this useful repo public.
My question is regarding the conversion of an OrdinalLogisticModel() wrapped model to its tensorRT equivalent.
I am unable to do it and I think it has something to do with the LogisticCumulativeLink at the end of the wrapper class.
Have you personally tried converting the model to tensorrt for faster inference?
If yes, would you be kind enough to let me know.
Thank you

solution: instead of accepting X, accept *args and **kwargs

happy to submit a PR to fix this

Hi,

I am trying to train an OrdinalLogit model and have stripped down the model to this:

pred_dosages_tensor = torch.tensor(pred_dosages)
true_dosages_tensor = torch.tensor(true_dosages, dtype=torch.long)
predictor = torch.nn.Sequential()
num_classes = len(np.unique(true_dosages))

scaling = NeuralNet(
	  module=OrdinalLogisticModel,
	  module__predictor=predictor,
	  module__num_classes=num_classes,
	  criterion=CumulativeLinkLoss,
	  train_split=None,
	  callbacks=[
	      ('ascension', AscensionCallback()),
	  ],
)
scaling.fit(true_dosages_tensor, pred_dosages_tensor)

However, when I try to run this code, I get the following error:

File "/home/unix/ssadhuka/.conda/envs/shuvomenv/lib/python3.7/site-packages/spacecutter/losses.py", line 68, in cumulative_link_loss
    likelihoods = torch.clamp(torch.gather(y_pred, 1, y_true), eps, 1 - eps)
RuntimeError: gather_out_cpu(): Expected dtype int64 for index

EDIT: Resolved

Fix:
In losses.py, change
likelihoods = torch.clamp(torch.gather(y_pred, 1, y_true), eps, 1 - eps) to likelihoods = torch.clamp(torch.gather(y_pred, 1, y_true.to(torch.int64)), eps, 1 - eps)

Actually I dont want to use skorch (more flexible), so how should i change my training code? i dont know how to add callback to my train code ..... Thanks very much !!!

Hi could u help understand the below peace of code.
Why do we subtract the elements in linkmat ,then concatinating them .
Isnt just cutpoints-X is sufficient ?

sigmoids=cutpoints-X
 link_mat = sigmoids[:, 1:] - sigmoids[:, :-1]
        link_mat = torch.cat((
                sigmoids[:, [0]],
                link_mat,
                (1 - sigmoids[:, [-1]])
            ),
            dim=1

2. when does this AscensionCallback gets called up.. start of every batch,epoch,or end of batch or epoch

Hi! Thanks for your work, I think it's useful and inspiring, however, Skorch may support torch 0.4.1 no more. When I try to run 'from Skorch.callbacks import Callback, ProgressBar', I met 'ImportWarning: Skorch depends on a newer version of PyTorch (at least 1.1.0, not 0.4.1). Visit https://pytorch.org for installation details', may be spacecutter should upgrade torch version? Thanks a lot!