a-PyTorch-Tutorial-to-Image-Captioning icon indicating copy to clipboard operation
a-PyTorch-Tutorial-to-Image-Captioning copied to clipboard

Published 20 hours ago • verified publisher icon sgrvinod

what is the lowest loss we can achieve with this architecture?

Open laptopmutia opened this issue 7 years ago • 3 comments

I tried to overfitting the model but it seems so hard to achieve

I've minimize the data into 100 image with 1 captions per image and the lowest loss I can achieve when training in 500 epochs is around 0.9xxx and the loss value is stagnant till 1000 epoch

what is the lowest loss you've achieve @sgrvinod ?

laptopmutia avatar Oct 11 '18 15:10 laptopmutia

Hey, sorry, I've been really busy and I'll respond to your email (and here) in the next day or two.

I don't remember the loss values, especially the training loss - but I also don't recall noticing that it was stagnant or slowing down. I wouldn't think that the model couldn't overfit.

Is this on (a small part of) MSCOCO or another dataset?

sgrvinod avatar Oct 12 '18 08:10 sgrvinod

the data is from flickr8k images and karpathy's json file but I cut it to 100 training image and 20 validation image here are the json in my gist

laptopmutia avatar Oct 12 '18 18:10 laptopmutia

Hi,

When train on coco, mine minimum is about 3.x. Is it normal? My generated captions seems very bad also...

guantinglin avatar Apr 16 '19 05:04 guantinglin