effectiveness for pretraining

[knazeri]

effectiveness for pretraining

[knazeri]

@1900zyh Technically Places2 dataset is a huge dataset that captures most of the variation in natural images. That means you can use the pre-trained model on Places for other datasets; however, unlike most classification tasks, do not freeze early layers while training. Using the pre-trained model significantly reduces your training time.
Another trick that can help you reduce training time is to pre-train the model on a smaller input size (say 128) and use the weights to train larger images later!

[zengyh1900]

Hi @knazeri , really appreciate for your instant response and helpful suggestions!
One more question, which standard do you usually use to judge the performance of the model during training?
I mean, will you carefully check each output image or do you think the quantitative evaluation numbers is more important during training?

[zengyh1900]

I wonder,
as there are a lot of solutions for the missing regions, do you think the quantitative evaluation is valuable for image inpainting task? higher number means higher performance?

also I wonder why don't you use Inception Score in your paper?

Thanks.

[knazeri]

@1900zyh Quantitative measures do not necessarily mean better output quality. I can refer you to this paper: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
Having said that, these measures might give you an estimate of how your model is performing. For better quantitative measures we use FID which has been shown to correlate with human perception! You can use the fid_score we have provided to find the Frechet Distance between ground truth and model results.
Inception score is good with class conditional GANs (say creating imagenet images). It shows how good a model creates images per class. FID is good when comparing any two distributions.

[zengyh1900]

@knazeri I saw existing image inpainting papers usually use Inception Score instead of FID, do you think Inception Score can help judge the fidelity of inpainting results (as it is usually applied in general GAN task)?

[knazeri]

FID is a more recent metric than Inception Score and most GAN papers are reporting it as a better quantitative measure. You can read more here.

I don't think Inception Score is a good metric for conditional GANs (like inpaiting problem) because it only takes into account how diverse the generated images are from a class conditional distribution, while FID measures the distance between two distributions without considering the labels!

[zengyh1900]

@knazeri thanks for your kind and patient answer!
Thanks a lot.