Hi, thanks for sharing this interesting work!

I noticed that the README file seems incomplete, i.e., the torchvision version is missing and sections like "Data Preparation" are empty. Can we expect some update of the documentation?

Best
Jakub

const.py文件中，import os
from pathlib import Path

Use environment variables to auto-detect whether we are running an a Compute Canada cluster:

Thanks to https://github.com/DM-Berger/unet-learn/blob/master/src/train/load.py for this trick.

COMPUTECANADA = False
TMP = os.environ.get("SLURM_TMPDIR")

if TMP:
COMPUTECANADA = True

if COMPUTECANADA:
INPUT_FOLDER = Path(str(TMP)).resolve() / "work" / "inputs"
MASK_FOLDER = Path(str(TMP)).resolve() / "work" / "inputs" / "masks"
PRETRAINED_MODEL_FOLDER = Path(str(TMP)).resolve() / "work" / "trained_models"
PRETRAINED_MODEL_DDPM_PATH = (
Path(str(TMP)).resolve() / "work" / "trained_models" / "ddpm"
)
PRETRAINED_MODEL_VAE_PATH = (
Path(str(TMP)).resolve() / "work" / "trained_models" / "vae"
)
PRETRAINED_MODEL_DECODER_PATH = (
Path(str(TMP)).resolve() / "work" / "trained_models" / "decoder"
)
PRETRAINED_MODEL_VGG_PATH = (
Path(str(TMP)).resolve() / "work" / "trained_models" / "vgg16.pt"
)
OUTPUT_FOLDER = Path(str(TMP)).resolve() / "work" / "outputs"
else:
INPUT_FOLDER = Path(file).resolve().parent.parent.parent / "data" / "IXI"
MASK_FOLDER = Path(file).resolve().parent.parent / "masks"
OASIS_FOLDER = Path(file).resolve().parent.parent.parent / "data" / "OASIS"
PRETRAINED_MODEL_FOLDER = (
Path(file).resolve().parent.parent.parent / "data" / "trained_models"
)这些预训练的模型都在哪里啊，数据在哪啊？

Hi, this is an interesting work. By the way, when do you plan to release the code?

Hi, I'm wondering if you have the vgg16.pt file as I cannot find it in the google drive or is it fine to use any vgg16 weights online? Thank you in advance!

Hi, there is no details about the encoder and decoder in the paper and code. I would like to ask what the specific model structure of the autocoder? And what's the information of training data, do they include the super resolution ground truth? Looking forward to your reply, thanks~

Hi!
I'm facing constant memory issues with running the ddim pipeline on my CPU, so I am wondering, what hardware did you run your experiments on? Most importantly, what was your RAM memory size?

Thanks for all previous answers
Jakub

Hi!
Would you mind sharing latent vectors file (latent_vector_ddpm_samples_100000.pt) or just the precomputed stats, that serve as the initialization of the latent code for the InverseSR(decoder) pipeline?

Thanks in advance
Jakub

@jueqiw What are the pre trained vae and vgg16 models being used here? There are no files in the drive. Is the files somewhere or its being trained by yourself?

Hello, author of InverseSR.
We are trying to utilize your nice work as our baseline in our paper for SR.
Due to the large resource requirement, we just want to use the same test images and then compare the evaluated PSNR, .. and other factors written in your paper.
I wonder the test images are the list below ( I saw it on your github code).
Thank you so much for your nice work.

Dear Jueqi,

Thanks for sharing this work! However, I noticed that some essential files are missing from the code repository. Specifically, the readme file, dataloader file, and IXI_T1_069.pth file are not included. I would greatly appreciate it if you could kindly upload the complete code for InverseSR, including all the necessary files.

Thank you very much for your attention to this matter. Looking forward to your positive response.

Best regards,
Adele

Hi!
Would you mind sharing the mask files used for acquiring 4mm and 8mm image slices?

Thanks in advance
Jakub

In the paper, it introduce a corruption function that generates masks for non-acquired slices, enabling our method to in-paint the missing slices. For instance, on 1 × 1 × 4mm3 undersampled volumes, we create masks for three slices every four slices on the generated HR 1 × 1 × 1mm3 volumes.

but I cannot find it in code. I do found a downsample function though, but it downsamples all the planes.
`class ForwardDownsample(ForwardAbstract):
def init(self, factor):
self.factor = factor

# resolution of input x can be anything, but aspect ratio should be 1:1
def __call__(self, x):
    x_down = F.interpolate(
        x,
        scale_factor=1 / self.factor,
        mode="trilinear",
        recompute_scale_factor=True,
        align_corners=False,
    )  # BCHW
    return x_down`