explainableml / bayescap Goto Github PK

Hi,
You mentioned that the err term in the UCE is computed between the reference (y) and the output of the frozen network(\hat{y}). I am wondering if you compute the err term in the UCE for the scratch model in the above way, or you compute the scratch model's err term by using the reference map and the output from the scratch model.

Thank you so much.

Hi,
I'm recently trying to run the evaluation codes using the provided checkpoints. I found that I cannot reproduce some of the results in Table 1. I wonder if I made some mistakes on my part.

On my part, I use the data and model loading parts in .ipynb, I load successfully the provided checkpoints, and then I put eval_BayesCap(NetC, NetG, test_loader) after them.
In eval_BayesCap, I activate line 765 and line 766.

After I ran the evaluation on Set5, I got:
Avg. SSIM: 0.7993451356887817 | Avg. PSNR: 28.397958755493164 | Avg. MSE: 0.0016931496793404222 | Avg. MAE: 0.02605036273598671
UCE: 0.014068946489913017
C.Coeff: [[1. 0.34907256]
[0.34907256 1. ]]
The result is consistent in UCE, but for Coeff, SSIM, and PSNR I cannot get the same results. I don't know if I I missed something.
Thank you for your help in advance.

Hi,
Thanks for your code.
I find that there are two implements of resi, and the one commented out is consistent with function (10) in paper.
So which is actually used in experiments?

# resi = torch.pow(resi*one_over_alpha1, beta1).clamp(min=self.resi_min, max=self.resi_max)
resi = (resi * one_over_alpha1 * beta1).clamp(min=self.resi_min, max=self.resi_max)

Hi,
When reading the code of the TempCombLoss(), I found that you simply give the same target to L1 and L2.
However, it seems the identity mapping term and the Negative log-likelihood term use different target term in Equation 10. If I understand correctly, L1 should use the tensor generated by the frozen model as the target while L2 should use the reference from the dataset as the target. In the paper, there appears two different y term for target in equation 10(y\hat and y).
Could you help explain why the implementation use the same target term ? Correct me if I make a wrong interpretation.

l1 = self.L_l1(mean, target)
l2 = self.L_GenGauss(mean, one_over_alpha, beta, target)
l = T1l1 + T2l2

Thank you very much.

Hi,
I have some questions when I try to re-produce the training for the super-resolution task:
In 4.1 the Super-resolution subsection writes,

the BayesCap is trained on ImageNet patches sized 84 ×84 to perform 4× super-resolution

In the Implementation Details subsection, it writes

a batch size of 2 with images that are resized to 256 × 256

I wonder what I should put for the ? in this line: train_dset = ImgDset(dataroot='./xxx', image_size=(?,?), upscale_factor=4, mode='train') And it is 4 or 2 during training for the batch size (since there is a commented line writes 4 for the batch size in .ipynb).

Thank you.

Hi the authors,

Thanks for making such a great work and i would like to apply this method into my current work. Correct me if i'm wrong but when I try to run BayesCap_SRGAN_train_and_eval.ipynb, it appears that you might lack the function of eval_BayesCap to fully reproduce the complete functionalities of the notebook itself.

Could you kindly help me with this problem. Thanks a lot

I would like to reproduce the results for the described task of MRI Translation, however, I find it difficult to find a complete pipeline to do this. I have downloaded the T1 and T2 data from the IXI Dataset website. You reference your other repo as the guideline for reproducing the results, however, I am still unsure about the data preparation steps taken. The T1 images are [256, 256, 150] dimension and the T2 images are [256,256,130] dimension. From the network code, it does not seem like you are using any 3D Convolutions, but I also cannot find how you convert the images to a suitable input for a 2D Conv model, because I don't think the base model you are using would have >100 input channels.

Therefore, I am hoping that you can point out what the other preprocessing steps are and what the actual input and output dimensions of the expected model are?

Thanks in advance!