shouldn't it be something like

k[i, j] = np.cos(np.deg2rad(deg)) * kd[i, j] + np.sin(np.deg2rad(deg)) * kd[i, j]

???

Could you share with us the training code and the loss function?

Hello @jannctu!

What are the initial 16 directional gradient kernels?

I want to submit a PR with a TensorFlow implementation of TIN2, but the initial gradient kernels aren't defined in the paper or the code.

Hello,

At the Dataloader I am getting following errors:
ValueError: Caught ValueError in DataLoader worker process 0.
ValueError: not enough values to unpack dataload

Btw I am using MacOS, and I don't have NVidia Graphics card so I am not using CUDA. Thus, at inference.py I closed the following line:
#model.cuda()

and I modified the following line as:
checkpoint = torch.load(weight_file, map_location=torch.device('cpu'))

Do you have any idea on how i can solve this issue?

Thank you!
Sinem

Hello Jan!

Could you please re-explain and detail a bit more the function of the Summarizer module?

It is writing in the paper that:

The last module tries to summarize the features generated by the
Enrichment modules and to produce the final edges. We use eight
1x1 convolutional layers together with a sigmoid activation function.
Fig. 5(a) shows the 8 channel outputs of Summarizer 1 (Point C on
Fig. 6), and (b) shows the final output of two summarizers (Point D
in Fig. 6).

How summarization is performed? Can we consider that convolution operation something like a sum operation and sigmoid function something like thresholding?

Thank you,
Sinem

Can you provide training code?