Train Accuracy decreses after some epochs

1

After running my model multiple time,

the train accuracy and test accuracy both decreasing after 4th epoch.
how it is possible, what could be reasons?

Train Accuracy
0,==Acc==,45.29
1,==Acc==,67.1
2,==Acc==,74.84
3,==Acc==,80.25
4,==Acc==,83.77
5,==Acc==,56.99
6,==Acc==,46.62
7,==Acc==,46.06
8,==Acc==,46.92
9,==Acc==,46.66
10,==Acc==,47.72
11,==Acc==,48.04
12,==Acc==,46.24
13,==Acc==,48.41
0,==Acc==,6.37
14,==Acc==,29.57
15,==Acc==,47.54
16,==Acc==,47.54
17,==Acc==,46.96
0,==Acc==,10.82
18,==Acc==,47.54

Test Accuracy
0,==Acc==,45.28
1,==Acc==,71.76
2,==Acc==,79.92
3,==Acc==,80.61
4,==Acc==,94.31
5,==Acc==,46.17
6,==Acc==,47.49
7,==Acc==,47.49
8,==Acc==,46.17
9,==Acc==,48.86
10,==Acc==,47.49
11,==Acc==,47.47
12,==Acc==,47.48
13,==Acc==,43.71
14,==Acc==,47.47
15,==Acc==,47.49
16,==Acc==,47.49
17,==Acc==,47.49
18,==Acc==,47.49.

I am using a simple custom crossentropyloss
def custom_categorical_cross_entropy(y_pred, y_true):
y_pred = torch.clamp(y_pred, 1e-9, 1 - 1e-9)
return -(y_true * torch.log(y_pred)).sum(dim=1).mean()

again if I use
def custom_categorical_cross_entropy(y_pred, y_true):
return -(y_true * torch.log(y_pred)).sum(dim=1).mean()

loss is coming NaN.

One more tried LogSoftmax and NLLLoss(). there alse after some epoch loss is NaN.

What could be possible issue.

2

There might be different reasons for the divergence, such as an exploding loss.
I would also not recommend to reimplement already provided loss functions, if the custom ones don’t add any new functionality (e.g. soft labels etc.), as e.g. nn.CrossEntropyLoss uses internally the log-sum-exp trick for numerical stability, which is often missing in custom implementations.

3

I have also tried LogSoftmax and NLLLoss(). In that case alse loss is coming as NaN.
My question is , is it problem with data or the model. or anything else?

4

I am using logsoftmax and NLLLoss.
My training accuracy decreased after 88 epoch.
In 88 epoch : Accuracy is 51 %
Suddenly in 89 epoch accuracy is 1.53 %