To Ask for Adapting Pretrained Models #239
Replies: 1 comment 1 reply
-
@kbjoon1011 Hi Jun. Any of the models here that I trained were done so by downloading the ImageNet dataset (images and annotations) in tar format to one of my local machines after agreeing to their license terms which state that it's for non-commercial/research purposes only. What that means for the use of pretrained models for commercial purposes is not clear to me. I put the note at the bottom of the README to make it clear that my code is Apache 2.0, but the pretrained weights are likely governed by the original dataset license. So I can make no additional claims as to how the weights themselves can be used, other than to say I trained them for research and do not use them for commercial purposes. I have been working on creating an OpenImages pretraining dataset with a subset of classes / images that would provide a larger, but not completely unmanageable alternative dataset for pretraining models. OpenImages has a less restrictive license for the annotations, and the images have been sorced under creative commons. |
Beta Was this translation helpful? Give feedback.
-
Hello, Mr. Ross.
I am Jun and in checking that it is legal to adapt models you have developed below,
https://rwightman.github.io/pytorch-image-models/models/
becuase ImageNet was developed only for non-commercial research and you have developed them for research.
Therefore, I would like to know the way you have developed.
When you develop pre-trained weights,
a) Did you download and save original images on your own server or database and use them?
b) Did you just use thumbnail images on ImageNet without downloading?
c) Did you use original images online through the URLs provided by ImageNet?
A lawyer wanted me to ask you about above.
I will really be grateful with your answer.
Thank you,
Regards.
Beta Was this translation helpful? Give feedback.
All reactions