Compose a learner from 3 pretrained resnets in parallel?


I have 3 classifiers that I trained on my data for different tasks. Lets say resnetA, resnetB and resnetC.

I would like to create a “super classifier” that would:
get an image as input, send it separately to resnetA, resnetB and resnetC. Get the 3 features map set (before the FC prediction layer), concatenate those 3 sets of features maps, plug it to 1 to few FC layers and get a final classification.

And I would like to initialize the 3 “subparts” of that super classifier by the previous training I did for resnetA, resnetB and resnetC.

the “frozen” version of the training would have only the last FC('s) trainable.
unfreeze would make the “subparts” resnetA, resnetB and resnetC also trainable.

Is there a tutorial / page / video that could help me to achieve that?

(I am fairly new to fastai and pytorch).

Any help would be very welcome