Questions on using sample images

[metaphorz]

You are pulling images from .github/ such as .github/cake.png. Where is this located--in your repo? In any case, I was able to insert a reference in colab's file system i.e. /content/catr/test.jpg so all is OK.

I tried two pictures and obtained peculiar results. For the attached, I got "Baby eating a donut with a spoon". Is it trained on a particular corpus so I will know what sorts of images will work better at receiving a more accurate caption?

https://www.google.com/search?q=picture+of+kid+eating+ice+cream&client=firefox-b-1-d&sxsrf=ALeKk01x0N-Cp698EaTBLihZO1tmGKxsEQ:1624046293165&tbm=isch&source=iu&ictx=1&fir=8linmDlDQlkv2M%252CaFcoZXjfh6m-RM%252C_&vet=1&usg=AI4_-kTh7sbXkQdBOOZh72YivTDSQmnZzA&sa=X&ved=2ahUKEwjd-Zvz-6HxAhWLG80KHaxGBL0Q9QF6BAgQEAE&biw=1595&bih=1126#imgrc=c2CgWDwHJ4rJtM

[saahiluppal]

This model is trained on COCO 2014 captions for 30 epochs and fine-tuned on COCO 2017 for 8-10 epochs.
Find coco data-set here

Due to attention mechanism, this model needs to be trained for at least 200 epochs directly on COCO 2017 dataset for extremely well results, but i couldn't do it because of resources constrain (providing this 30 epochs baseline took more than a week on p100 GPU)

Therefore, this model is provided as a standard baseline.

[metaphorz]

I completely understand on the training time, having spent 4 days training for StyleGAN2-ADA.
Thanks for your time and work. I guess I am now seeking a pretrained model using the COCO dataset,
so I'll scout around.