Why I have negative loss in training?

Question

I have the following model.

import torch
from torchvision import datasets, transforms
import torch.nn.functional as f
class CNN(nn.Module):
    def __init__(s