Hi thanks for your work, I was trying to test your code but the datasets apparently were deleted or moved to another directory in google drive (COCO-20 and FS-1000). Can you update the URL where we can download them, please? Thanks in advance.

This is just a minor issue, but I still want to bring it up.

When calling downlaod_from_google() method from ./data/download.py, the requests.Session().get(url, .......) method probably no longer returns any cookies (all I got is just an empty list) for PF-PASCAL and SPair-71k datasets, so no file is downloaded.

I checked the actual absolute url = https://docs.google.com/uc?export=download&id='file-id from download.py'. It gives a warning page saying that the file is too big that google cannot perform virus scan, and it asks you to confirm if you still want to download the file.

Seems like we still need to find another way to bypass the Google Drive Virus Scanning.

PS: there's a workaround though: just manually download the datasets into ../Datasets_CATs/. (Datasets_CATs/ needs to be in the parent folder of Volumetric-Aggregation-Transformer)

Hi, First, thank you for sharing your work.

I tried the following command to start training. python train.py --config "config/pascal_resnet50/pascal_resnet50_fold0/config.yaml"

However, strangely training with base config file is not possible because of out of memory even though I use 3090 GPU.

I don't know the reason but somewhere takes lots of GPU memory. Training only started when I reduce batchsize to 4 from 16.

and..., This command doesn't work. Could you check ? conda env create -f environment.yaml

Congratulations on this cool work! VAT achieves SOTA on multiple datasets including PF-WILLOW and Spair-71k. I am very interested in this task. However, I can not find training and test codes related to this task in the repo. I would appreciate it if you could include them in this repo. Thank you in advance.

I'm trying to visualize some test results by using /utils_training/utils.flow2kps method. and I figured your model is trying to map keypoints from target images to src images, i.e. making predictions of keypoints in src images. (if im not mistaken). but the Estimated_kps Im getting seem off and sometimes out of range (out side of the image regions). I also noticed both the source and target image in a pair are resized to a same image size. this is prbly something I missed

I wonder how i can rescale the estimated_kps so it will correctly mark out the keypoints in the original images. (Orwhat size should i rescale the original image to use the predicted keypoints)

Hello.

Thank you for this amazing work.

I want to retrain your model, but when I traid I got this error. Do you have any experience with this error?

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [2, 128, 32, 32, 8, 8]], which is output 0 of ReluBackward0, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

Hi, Seokju-Cho

Good work! This is a quite surprising job for few-shot segmentation.

I am trying to re-train VAT on PASCAL, but I find that it may take about 5 days on 4 Tesla V100 GPUs for 300 epochs.

• I wonder how long did it use in your experiments (with 3090 or other GPU)?
• And could you share a copy of the training log of the experiment for reference?

Thanks a lot!

Hi @Seokju-Cho

Thank you for the code.

I have several questions that would be great If you answer.

As I see in the code, you are saving the best model by validating on validation data of the current fold(Pascal dataset) in every epoch of training. Am I right? In pascal dataset validation data are used for reporting the MIOU and should not be used during training.

For reporting the model accuracy, most of the papers in few-shot segmentation are using Kfold experimenting. As I see in your code you report the accuracy just by one try. Is it right?

The final question is image sizes for training and validation. As I see in your code, you are using 473 which is different than HSnet (400) and REPRI (417) net. As it is mentioned in Hsnet GitHub issues " https://github.com/juhongm999/hsnet/issues/6 ", The author claimed, "In our experiments, we found that larger image sizes typically (> 417) result in better mIoU results". You are better than HSnet by a small margin which can be due to image sizes. Am I right?

Best

Got this error on running on the images folder

Traceback (most recent call last): File "test.py", line 94, in test_miou, test_fb_iou = test(model, dataloader_test, args.nshot) File "test.py", line 44, in test average_meter.write_result('Test', 0) File "/content/Volumetric-Aggregation-Transformer/common/logger.py", line 51, in write_result loss_buf = torch.stack(self.loss_buf) RuntimeError: stack expects a non-empty TensorList