Thanks for this impressive work, I am still going through the work but I want to understand,
For the image_sample script, you apply a [0.5, 0.5, 0.5] mean and standard deviation normalization to the image. I want to understand if I need to replace that with my own norm and standard deviation.
The line in consideration is here:
https://github.com/facebookresearch/RCDM/blob/71daaf10a73bb2012864f0827c68d209fc92b0a5/scripts/image_sample.py#L35C1-L37C10
Secondly, in the following lines: https://github.com/facebookresearch/RCDM/blob/71daaf10a73bb2012864f0827c68d209fc92b0a5/scripts/image_train.py#L84C1-L88C86
You extract representations from the image using the ssl model, but in this case, I see no image transformation applied to the dataloader that aims to normalize the image to the range where the ssl model is trained on. Should I be inserting a way to normalize the image batch here using my own mean and standard deviation (I did not use the imagenet mean and standard deviation, I have pathology images so I don't use those values).
I would love to get a clarification.
Thanks for this impressive work, I am still going through the work but I want to understand,
For the image_sample script, you apply a [0.5, 0.5, 0.5] mean and standard deviation normalization to the image. I want to understand if I need to replace that with my own norm and standard deviation.
The line in consideration is here:
https://github.com/facebookresearch/RCDM/blob/71daaf10a73bb2012864f0827c68d209fc92b0a5/scripts/image_sample.py#L35C1-L37C10
Secondly, in the following lines: https://github.com/facebookresearch/RCDM/blob/71daaf10a73bb2012864f0827c68d209fc92b0a5/scripts/image_train.py#L84C1-L88C86
You extract representations from the image using the ssl model, but in this case, I see no image transformation applied to the dataloader that aims to normalize the image to the range where the ssl model is trained on. Should I be inserting a way to normalize the image batch here using my own mean and standard deviation (I did not use the imagenet mean and standard deviation, I have pathology images so I don't use those values).
I would love to get a clarification.