search

cc-ai / climategan

Code and pre-trained model for the algorithm generating visualisations of 3 climate change related events: floods, wildfires and smog.
https://thisclimatedoesnotexist.com
GNU General Public License v3.0
75 stars 18 forks source link

Image logging error in depth head #101

Closed melisandeteng closed 4 years ago

melisandeteng commented 4 years ago

In this PR There are some issues in logging depth images from the real world (image is repeated) and then there's also a slight issue with colors in the mask images that are logged in comet. (see this experiment )

vict0rsch commented 4 years ago

@melisandeteng I won't have time today but I'd be happy to peer-program with you tomorrow!

alexrey88 commented 4 years ago

The weird coloring error for the masker logging has been found. The repeated image is due to the file val_r_resized_d.json, where each path for the depth task is the same ("d": "/network/tmp1/ccai/data/munit_dataset/testA_megadepth_resized/04.npy"). Working on finding why depth images are only logged once.

melisandeteng commented 4 years ago

Wow, thanks for spotting that @alexrey88 . It was a bit of a stupid error, but on the other hand, I'm glad it's not something more important !

alexrey88 commented 4 years ago

My pleasure !! The trainer is running right now on my side and depth seems to be logged more than one time, see this experiment. It's just that for now the predictions are all black but maybe that's normal since it's only been one epoch?

melisandeteng commented 4 years ago

I can't see the experiment, but I'm glad it looks promising ! What was the issue with depth being logged only once ?

alexrey88 commented 4 years ago

Oh I think you will be able to see it now, my experiments were set to private. Actually I don't know what the issue was, I just let it run for a first look and it was being logged!

melisandeteng commented 4 years ago

Predictions after one epochs are just nan tensors for both masker and depth head ...

melisandeteng commented 4 years ago

There was a mistake in the simulated data depth map decoder. Input was thought as having values between 0 and 1, but they were already pixel values between 0-255. The problem has been fixed