Open qibao77 opened 3 years ago
Hi, qibao77, Thanks for your attention to HINet! And yes, we use the default parameters in "train/REDS/HINet.yml" to train HINet. Little disturbance (ie. ± 0.02 in psnr) is possible but in our experience, it won't be large in the REDS dataset.
The validation results should be like (near):
iterations | val-300 psnr |
---|---|
40k | 28.0741 |
80k | 28.3048 |
120k | 28.4166 |
160k | 28.5189 |
200k | 28.6355 |
240k | 28.6918 |
280k | 28.7476 |
320k | 28.7961 |
360k | 28.8173 |
400k | 28.8267 |
And the training log could be downloaded at here, it might help you to address the problem. And we will check the reproducibility again.
Thank you for your reply!! What's more, I found that your NTIRE2021 competition result is 29.2533 on the official website and 28.8267 here, Is there any difference during the test?
Thank you for your reply!! What's more, I found that your NTIRE2021 competition result is 29.2533 on the official website and 28.8267 here, Is there any difference during the test?
Hi, qibao77, We extend the HINet as we describe in the paper section 4.4, ie. deeper, wider, ensemble, test-time augmentation and etc. to achieve 29.25 PSNR in the challenge : )
Hi, this is a nice job. However, Do you use the default parameters in "train/REDS/HINet.yml" to train HINet? I can not reproduce the results of the pre-trained model you provided.
Hi, qibao77,
We check the reproducibility of REDS, and get a result of 28.82 , by train the HINet with default parameters in "train/REDS/HINet.yml". The result is very close to our pretrained model (28.83).
Could you please share your results, or training log, etc. ? If the gap is marginal, ie. < 0.02, might just be a turbulence : )
Hi, Thank you for your reply again! I reproduced your model with a margin of 0.04, which may be reasonable.
Hi, this is a nice job. However, Do you use the default parameters in "train/REDS/HINet.yml" to train HINet? I can not reproduce the results of the pre-trained model you provided.