AttributeError: 'numpy.ndarray' object has no attribute 'log'

1

File “/home/ali/BioNet_project/pytorch_version/utils.py”, line 90, in train
soft_score = soft_loss(y, x, beta=0.95)
File “/home/ali/BioNet_project/pytorch_version/metrics.py”, line 23, in soft_loss
cross_entropy = F.nll_loss(y_pred.log(), y_true, size_average=False)
AttributeError: ‘numpy.ndarray’ object has no attribute ‘log’

2

It seems you are trying to pass a numpy array to F.nll_loss, while a PyTorch tensor is expected.
I’m not sure how y_pred is calculated, but note that using numpy array would detach them from the computation graph, so you should stick to PyTorch tensors and operations, if possible.

3

thanks for your reply. yes, I am trying to calculate the soft_loss but stuck in this problem and don’t know what to do.

metrics

        x, y = output.detach().cpu().numpy(), y.detach().cpu().numpy()
        soft_score = soft_loss(y, x, beta=0.95)

def soft_loss(y_true, y_pred, beta=0.95):
cross_entropy = F.nll_loss(y_pred.log(), y_true, size_average=False)
soft_reed = -y_pred * torch.log(y_pred + 1e-8)
return beta * cross_entropy + (1 - beta) * torch.sum(soft_reed)

4

Remove the .numpy() operations and pass the tensors to the soft_loss function.
As described before, you won’t be able to call backward() on the loss and calculate gradients with it, since you are also explicitly detaching the tensors, but I assume that’s wanted.

5

thanks for your suggestion and i have followed your words but facing now this issue.
####modification######
# # loss
# l = criterion(output, y)
# # tot_loss += l.item()
# l.backward()
# optimizer.step()

        # metrics
        x, y = output.detach().cpu(), y.detach().cpu()

        # iou_score = iou(y, x)
        # dice_score = dice_coef(y, x)
        soft_score = soft_loss(y, x)
        hard_score = hard_loss(y, x)

######metrics###########
import numpy as np
import torch.nn.functional as F

def soft_loss(y_true, y_pred, beta=0.95):
cross_entropy = F.nll_loss(y_pred.log(), y_true, size_average=False)
soft_reed = -y_pred * torch.log(y_pred + 1e-8)
return beta * cross_entropy + (1 - beta) * torch.sum(soft_reed)

#####errorrr####
File “/home/ali/BioNet_project/pytorch_version/utils.py”, line 90, in train
soft_score = soft_loss(y, x)
File “/home/ali/BioNet_project/pytorch_version/metrics.py”, line 23, in soft_loss
cross_entropy = F.nll_loss(y_pred.log(), y_true, size_average=False)
File “/home/ali/miniconda3/envs/bio/lib/python3.6/site-packages/torch/nn/functional.py”, line 2117, in nll_loss
ret = torch._C._nn.nll_loss2d(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
RuntimeError: only batches of spatial targets supported (3D tensors) but got targets of dimension: 4

6

This is the shape of my data ‘‘data shape: (240, 512, 512, 3) (240, 512, 512, 1)’’.