Pretrained-Show-and-Tell-model icon indicating copy to clipboard operation
Pretrained-Show-and-Tell-model copied to clipboard

Published 20 hours ago • verified publisher icon KranthiGV

Poor performance

Open ericktokuda opened this issue 7 years ago • 5 comments

First, thanks for sharing your work. Results using these models though seems to be worse than Karpathy's work.

ericktokuda avatar Jun 18 '17 20:06 ericktokuda

@tokudaek I will evaluate the results and get back to you.

KranthiGV avatar Jun 19 '17 15:06 KranthiGV

@tokudaek The graphs seems to be okay. I accept that this model doesn't perform well on all types of images. The issue can also be with the 2M checkpoint (I can run it at least until 3M) in which Inception weights are also updated. So, the loss is still high at 2M iteration. (I observed a peak in loss after 1M iteration when I started updating weights of CNN and LSTM together.) Can you try to use 1M checkpoint file? You might get better results.

Let me know if I'm missing anything. :smile:

KranthiGV avatar Jun 20 '17 05:06 KranthiGV

Nice, I ll try testing the 1M checkpoint. Thanks a lot.

ericktokuda avatar Jun 20 '17 18:06 ericktokuda

Hey @KranthiGV I'm trying to train the im2txt model and no custom code written. capture

This is the loss curve during training. The model so far trained only for 310,000 iterations. The loss has not been changing much over the last 200,000 iterations. It is around 2-2.3. The captions generated for validation images seem good. But I suspect the loss is high. Is this the same behavior you see in your training?

peri044 avatar Jul 24 '17 23:07 peri044

For me the 2M model works better than the 1M, but it's still far from perfect (takes an accordeon for a tie for instance). Is there any way to improve the model? Would going to 3M help?

stephanedebove avatar Oct 29 '17 17:10 stephanedebove