The definition of a replay buffer exempler

[thkimYonsei]

I also want to clarify one more thing about the rule of the competition.

It says we can only store up to 200 exemplers into replay buffer from previous experiences.
Then how are you going to count the number of exemplers in the buffer?

For example, if I store output feature from a certain convolution layer and output logit from a prediction head of a single image, will it be counted as 2 exemplers stored, or counted as 1 exempler that is a tuple of (feature, logit)?

Also, if I keep the average loss of each trained (seen) class, will it be counted as storing N (the number of seen classes) exemplers, or just counted as one exempler that is the list of mean value?

Moreover, is the previous model from the previous experience counted as a single exempler? Or does it treated as it is not in memory buffer?

I hope I can figure it out.

Thank you.

[gigug]

Hi,
just to update you: we are currently working on better defining what a replay buffer element is.
I will come back to you tomorrow with more details.

[gigug]

From the new detailed guidelines that will soon be published on the website:

• How many samples can be stored?
  We allow storage of 200 samples. The scenarios assume that the buffer size is agnostic regarding the number of classes.
• How is a sample defined?
  An individual sample stored in the replay buffer cannot exceed the size of 1024 floating-points (in vector, matrix or other format). Each sample can encompass information pertaining to one or multiple samples or classes. In a sample multiple values of the input/class (for example: class mean & standard deviation) can be stored in the form of a tuple. Raw images, as well as any straightforward derivatives of them, are not allowed to be stored. Prohibited modifications include, among others: slicing, selecting specific segments of the image, pixel-wise alterations, basic color adjustments and aggregates of multiple images.

[thkimYonsei]

Thank you for the clarification!

[thkimYonsei]

Oh I have one more confusion. It seems like we can keep the previous model (the model of the previous experience) as you guys provide LWF as a baseline. Is there any limit in the number of previous models that we can utilize in the future experience? Or does it not matter if we don't violate the gpu vram memory limit?

[Tsebeb]

Hi :),

there is no explicit limit to the number of models that you can store for your method and they are not regarded as samples of the replay buffer. They are restricted by the GPU VRAM memory limit of 10GB. Furthermore, all of the models or elements that you need for training your method or during inference have to fit into this VRAM limit at once, please do not unload copies of models to the cpu or process them in sequence to comply with the VRAM limit.