Some questions about training

[logic1988]

@mgharbi Thank you very much for your great work, very helpful to me. Can you help me for the following questions:

1. why use tf.image.ResizeMethod.NEAREST_NEIGHBOR but not bilinear in https://github.com/mgharbi/hdrnet_legacy/blob/master/hdrnet/data_pipeline.py#L169
2. I used fivek dataset for training, 100 of which were used as eval sets. But why the eval pnsr was so low, only about 8.

3. Can you share some of the experience in training image enhancement tasks? How to adjust those super parameters, such as learning rate, batchnorm, etc.

[logic1988]

I found the reason for point 2: code in https://github.com/google/hdrnet/blob/master/hdrnet/bin/train.py#L105 was wrong.

[mgharbi]

1. We used nearest neighbor to match our fast Android implementation with the Camera2 API 2. Don’t hesitate to submit a pull request with your fix. Best, Michael From: logic1988 <notifications@github.com> Reply-To: mgharbi/hdrnet_legacy <reply@reply.github.com> Date: Sunday, November 25, 2018 at 4:25 AM To: mgharbi/hdrnet_legacy <hdrnet_legacy@noreply.github.com> Cc: Subscribed <subscribed@noreply.github.com> Subject: [mgharbi/hdrnet_legacy] Some questions about training (#35) Thank you very much for your great work, very helpful to me. Can you help me for the following questions: 1. why use tf.image.ResizeMethod.NEAREST_NEIGHBOR but not bilinear in https://github.com/mgharbi/hdrnet_legacy/blob/master/hdrnet/data_pipeline.py#L169 2. I used fivek dataset for training, 100 of which were used as eval sets. But why the eval pnsr was so low, only about 8. [Image removed by sender. a]<https://user-images.githubusercontent.com/13864726/48978956-e33b5800-f0ee-11e8-9215-11e8409e97d9.png> 3. Can you share some of the experience in training image enhancement tasks? How to adjust those super parameters, such as learning rate, batchnorm, etc. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#35>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ABib32ONsud4erppdIRybNV4k0Tf5b69ks5uyowjgaJpZM4Yx2DN>.