Questions on Code Details: Image Size, Loss, Virtual Batch Multiplier

[weimingwill]

"REMEMBER to set your image size to 3x299x299 for both test and validation" in https://github.com/tstandley/taskgrouping/blob/dc6c89c269021597d222860406fa0fb81b02a231/model_definitions/xception_taskonomy_small.py#L12

Is it still applicable? I didn't find the hints to set such image size in source code or README.

[tstandley]

Oh boy. That comment needs to be removed. It's from the original xception on imagenet. Unfortunately, I'm unable to login to GitHub at the moment.

On Fri, Dec 3, 2021, 11:02 PM Zhuang Weiming @.***> wrote:

"REMEMBER to set your image size to 3x299x299 for both test and validation" in https://github.com/tstandley/taskgrouping/blob/dc6c89c269021597d222860406fa0fb81b02a231/model_definitions/xception_taskonomy_small.py#L12

Is it still applicable? I didn't find the hints to set such image size in source code or README.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tstandley/taskgrouping/issues/6, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEDABXTD7DZUZD2JKFT3CALUPG4HPANCNFSM5JLIEBQA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

[weimingwill]

Noted. Thank you for your clarification. I have several other questions regarding the code details and hope to seek your advice.

1. "Not taking the average of loss", Is it applicable only to xception_taskonomy_new? Or also applicable to other models such as xception_taskonomy_small? https://github.com/tstandley/taskgrouping/blob/dc6c89c269021597d222860406fa0fb81b02a231/taskonomy_losses.py#L163

2. How to set `virtual-batch-multipler? What's the value that is set in the paper? 1? https://github.com/tstandley/taskgrouping/blob/dc6c89c269021597d222860406fa0fb81b02a231/train_taskonomy.py#L88

3. Are the results on paper evaluated on validation set or test set?

4. What is the value of relative performance in Table 1-7? Loss? Accuracy?

Thank you.

[tstandley]

1. The sum, not average, should be used whenever the TrevorRepFun is used in the model to distribute backpropagation to tasks. Looking over the code, this is only done for xception_taskonomy_new. For xception_taskonomy_small the gradients are added from all the losses, so the average would probably lead to more reproducible models. I don't remember what I did for that part in the paper, but it only matters if you are trying to train your own.

2. The vba should be set to 1 unless you are trying to reproduce results from a multi-gpu machine with a single gpu machine. For example, if you train one model on a machine with 2 gpus, and another model on a machine with 1 gpu, and you want to compare them, and you set the batch size to be the maximum that will fit into ram, you should use -vba=2 for the single gpu machine.

What it does is just run the forward-backward part multiple times before doing an optimizer update.

On Sun, Dec 5, 2021 at 1:05 AM Zhuang Weiming @.***> wrote:

Noted. Thank you for your clarification. I have two other questions regarding the code details and hope to seek your advice.

1.

"Not taking the average of loss", Is it applicable only to xception_taskonomy_new? Or also applicable to other models such as xception_taskonomy_small? https://github.com/tstandley/taskgrouping/blob/dc6c89c269021597d222860406fa0fb81b02a231/taskonomy_losses.py#L163 2.

How to set `virtual-batch-multipler? What's the value that is set in the paper? 1?

https://github.com/tstandley/taskgrouping/blob/dc6c89c269021597d222860406fa0fb81b02a231/train_taskonomy.py#L88

Thank you.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/tstandley/taskgrouping/issues/6#issuecomment-986191998, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEDABXXTTLFU4WLEAY2Y3C3UPMTPPANCNFSM5JLIEBQA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

[weimingwill]

Thank you.