single_object.py

[BernhardEgger]

Hi,

I just got started with TDW and it so far looks all great :)

I have however three questions to the single-object case - this is probably closest to what we plan to do:

• it refers to "Martin Schrimpf's and Jonas Kubilius' ImageNet transfer research" is there a paper or something that shows me what that project was?
• it renders in the highest possible quality, but then saves the files in jpg format with strong compression artifacts - is that intended?
• I tried to start with --train=1 --val=1 and it seems like those values are just ignored?

Best
Bernhard

[alters-mit]

it refers to "Martin Schrimpf's and Jonas Kubilius' ImageNet transfer research" is there a paper or something that shows me what that project was?

I think this paper has the most information https://arxiv.org/pdf/2007.04954.pdf

it renders in the highest possible quality, but then saves the files in jpg format with strong compression artifacts - is that intended?

Yes, because we found that they didn't affect the results. Use the --screen_size and --output_size parameters; if you set screen size to be higher (512) and output size lower (256) the image will be scaled down, which actually looks a lot better.

I tried to start with --train=1 --val=1 and it seems like those values are just ignored?

How many images per category does the script generate?

[BernhardEgger]

Thank you for guiding me for the paper

for the second point, I just deactivated the jpg output, since I'm generating pretty small images

it generates more than one - I have to stop it so I don't know where it would go. This is the directory structure I get after few seconds:

the meta-data file however gets the correct value
"train": 1,
"val": 1,

[alters-mit]

Closing this issue because this is the intended behavior of single_object.py

There should be one image per model. The fact that there's 2 is probably just a small rounding error.