Generally on medecine papers there is some sort of data leakage (like they do data augmentation before splitting to train, val and test)

What are the techniques to male such large models run on low ressources ?

Ok thanks !

Yes the error is just in a figure that is meant to illustrate the pre-processing, it changes nothing about neither the arguments nor the conclusion. By revision, you mean after the paper is accepted ? ( it is my first paper so bare with me) I just fear that it will refused because of a silly mistake.

Yes tou need to model the word to make an AI work in any financial application

Yes ! You only need to change the last classifier layer (and initialize the added weights) to add more outputs and then further train the model on data containing all the classes (including the new ones)

Also if use mixed precision. And if you have a TPU use Xla

+1 but I think he already knows that

You can train/pretrain only on ImageNet1k and compare with swinv2 and ConvNextV2 trained only on ImageNet1k (both share their results on this setting)

There are : check the ConvNext V2 paper

None of us is a native english speaker. We tried our best to not make mistake and be clear. But I only want an external opinion only for this paper as it is my first.

Have any paper in mind that speaks about this stuff ?

Accurate 😂

Can anyone send me the main papers describing the instructed and action driven models pleaseee ?

But there is no normalization in dropout right ?

Good idea 😂 Will do so :)

Isn’t that learning rate hight for finetuning ?

In this case I think only the difference between dates matters

I have couples of dates and procedures/tests ans their results. So having the date is important (per example a patient had cancer 5 years ago and was treated using conization)

Yes I by embedding I meant transforming each number of months to a vector, like nn.embedding in pytorch (knowing that the difference between dates can’t be more than 5 years so 60 months) Thanks for the answer !