georgosgeorgos / few-shot-diffusion-models Goto Github PK

Can you provide the details about the metrics computation?
I found the code in the following run file:

Thank you for your excellent work. I would like to ask how to specify the category of the generated image during the sampling process. Is it by using the relevant seen or unseen image set as the input set of ViT to control the single-round image generation?

It seems datasets are expected to be packed in .pkl file, but the instruction on how to prepare the .pkl files are missing. Alternatively, would it be possible to provide pre-processed pkl files?

Here are errors I got due to missing files.

Traceback (most recent call last):
  File "main.py", line 112, in <module>
    main()
  File "main.py", line 49, in main
    TrainLoop(
  File "/home/roger/reproduction/few-shot-diffusion-models/model/set_diffusion/train_util.py", line 389, in run_loop
    batch = next(self.data)
  File "/home/roger/reproduction/few-shot-diffusion-models/dataset/__init__.py", line 58, in create_loader
    dataset = select_dataset(args, split)
  File "/home/roger/reproduction/few-shot-diffusion-models/dataset/__init__.py", line 26, in select_dataset
    dataset = BaseSetsDataset(**kwargs)
  File "/home/roger/reproduction/few-shot-diffusion-models/dataset/base.py", line 56, in __init__
    self.images, self.labels, self.map_cls = self.get_data()
  File "/home/roger/reproduction/few-shot-diffusion-models/dataset/base.py", line 81, in get_data
    with open(path, 'rb') as f:
FileNotFoundError: [Errno 2] No such file or directory: '/home/gigi/ns_data/cifar100/train_cifar100.pkl'

There is no run.sh script in the scripts folder. There is only one main.sh script. Is it the run.sh document? I renamed main.sh to run.sh, and then run bash script/run.sh 2num ddpm_ cifar100, there is no output, why?

It seems that the requirements.txt need some local files.
Could you please check the requirements.txt and modify its content?

Is there a plan to release pre-trained weights so that sampling/evaluation can be done without re-training everything from scratch?

Im vfsddpm.py line 223-225: x_set_tmp = batch[:, ix] yout select ns - 1 of data, but you don't select ns-1 t_emd, that cause vitset.py line 176 :t_emb = t_emb.view(b, ns, -1) error, this t_emd is 49152 contain deleted data t_emd, but x_set 's ns is 5.
I want to know this code is the origin code of your paper result?
Thank you.

few-shot-diffusion-models/dataset/base.py", line 95, in get_data
residual = self.img_cls - value.shape[0]
AttributeError: 'list' object has no attribute 'shape'

Could you please provide the data file?

There are some inconsistencies between the code and the contents of the readme file.

Thank you for open-sourcing the project. I see a lot of issues with the repo, no explicit instruction to prepare data, no working training or evaluation scripts, etc. Do you plan to publish the dataset files at least so we can experiment with it?

In vit.py in forward_set function

the transform module is being used twice in 1 forward evaluation.

First the common forward

Second during mean_patch pooling

Is this intended?