An issue on implementation of SemanticConnectivityloss

[rose-jinyang]

Hi @LutaoChu and @dreamer121121
In the forward function of the SemanticConnectivityLoss implementation(https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.4/paddleseg/models/losses/semantic_connectivity_loss.py), are the parameter logits the probabilities map or logit map without applying softmax?
BTW:

There is an issue with the above function implementation.
pred_conn has labels greater than pred_num_conn value.
That's because pred_num_conn is limited as smaller than 10.

This results in an issue in the below line.
pred_conn = F.one_hot(pred_conn.long(), pred_num_conn)

How can we fix this issue?

[LutaoChu]

Hi, Logits is logit map without applying softmax.

Regarding the above issue, it looks like there is indeed a problem. We need to slice pred_conn.
But it is very strange that I did not find any errors during the actual training process. Can you provide your error message?

[LutaoChu]

I tested it for paddle. If the range of values in pred_conn exceeds the pred_num_conn value, F.one_hot can still be calculated

[rose-jinyang]

In fact, I am using this loss in PyTorch.
In PyTorch, F.one_hot does not handle such case.

[rose-jinyang]

In case of using Paddle, how will the values in pred_conn exceeds the pred_num_conn value be processed?

[LutaoChu]

In fact, I am using this loss in PyTorch. In PyTorch, F.one_hot does not handle such case.

You can try it as the following

[LutaoChu]

In case of using Paddle, how will the values in pred_conn exceeds the pred_num_conn value be processed?

Exceeded values will be ignored. The corresponding one hot codes returned are all 0

import paddle
import paddle.nn.functional as F

a = paddle.randint(0, 10, (4, 4))
b = F.one_hot(a, 5)
print(a)
print(b)

Output

Tensor(shape=[4, 4], dtype=int64, place=CUDAPlace(0), stop_gradient=True,
       [[7, 2, 3, 3],
        [8, 6, 2, 6],
        [8, 1, 7, 4],
        [4, 1, 8, 6]])
Tensor(shape=[4, 4, 5], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
       [[[0., 0., 0., 0., 0.],
         [0., 0., 1., 0., 0.],
         [0., 0., 0., 1., 0.],
         [0., 0., 0., 1., 0.]],

        [[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 1., 0., 0.],
         [0., 0., 0., 0., 0.]],

        [[0., 0., 0., 0., 0.],
         [0., 1., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 1.]],

        [[0., 0., 0., 0., 1.],
         [0., 1., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]])

[rose-jinyang]

I've got the below issue with the above code.

[rose-jinyang]

What is your paddlepaddle version? Mine is paddlepaddle 2.2.2.

[LutaoChu]

Mine is 2.2.1
Please give your running command and I'll repeat the problem

[rose-jinyang]

I've installed paddle 2.2.1.
But the issue appears still.

[LutaoChu]

I have no problem here
What is your running command?

[rose-jinyang]

import paddle
import paddle.nn.functional as F

a = paddle.randint(0, 10, (4, 4))
b = F.one_hot(a, 5)
print(a)
print(b)

This is the script.

[LutaoChu]

So strange. I can run it

[rose-jinyang]

The difference is only that I tried it on CPU without GPU.

[LutaoChu]

I have the same problem with CPU. This is an OP bug in paddle. You can submit the issue here https://github.com/PaddlePaddle/Paddle

[LutaoChu]

In fact, I am using this loss in PyTorch. In PyTorch, F.one_hot does not handle such case.

You can try it as the following

If you only have CPU, you can modify it according to this

[rose-jinyang]

One of Paddle developers recognized this issue on GPU.
I think that your loss implementation should be updated under consideration of this point.

[rose-jinyang]

Could u provide the implementation of this loss in PyTorch.
That's because some issues appears when running this loss implemented by myself in PyTorch.

[LutaoChu]

In fact, I think that the input element can support greater than class_num is a feature, not a bug.

I will try the pytorch loss implementation

[rose-jinyang]

Hi @LutaoChu
Did u implement PyTorch version of this loss?

[LutaoChu]

Hi，I fix this bug on CPU with Paddle. You can use it after the PR is merged