Issues doing Multi-Image Regression using datablock api

After a few days of googling for solution, I am kinda of stuck now. I am attempting a kaggle challenge, where we supposed to predict how fast a pet will be adopted. 0 being the fastest, and 5 being the slowest.
Looking at the images, it seems the younger and cuter looking ones are usually adopted faster. Thus, I was hoping the CNN would be able to detect this. However, after using resnet as a base model, the accuracy was simply horrible. Like error rate was like 70%. :open_mouth: I am most probably missing some additional touches required for this model.

I had a few questions and have attached the notebook below. Pls do a brief run through of the notebook and let me know, where I am going amiss.
Thanks Everyone. :smiley:

Notebook Link

Qn. 1 : Should this be a classification or a regression problem ? I was thinking regression would be better solution.

Qn. 2 : How do I read in multiple images to predict per Petid. Usually, for tabular data, I would usually aggregate rows to become one row. Here it was stated that _1 is the default image, so I just choose that

Qn. 3 : I also tried unsuccessfully to use the datablock API, After a few days of trying, I was only able to read in the _1 images using the suffix as _1.jpg. Another simpler approach, would be reorganise all the images in bash, by putting the images in the labelled folder like MNIST. However, it would be nicer if this can be done in-house within the datablock API.

Qn. 4 : I also tried to run a image regression problem, but got a runtime error. I also am not able to fix it.

To simplify the qns, looking at the notebook, for image classification, I got an error rate of 65%. Can someone guide me in another direction ?