Vision: Origin of the means and stds used for preprocessing?

[pmeier]

Does anyone remember how exactly we came about the channel means and stds we use for the preprocessing?

transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

I think the first mention of the preprocessing in this repo is in #39. In that issue @soumith points to https://github.com/pytorch/examples/tree/master/imagenet for reference. If you look at the history of main.py the commit pytorch/examples@27e2a46c1d1505324032b1d94fc6ce24d5b67e97 first introduced the values. Unfortunately it contains no explanation, hence my question.

Specifically, I'm seeking answers to the following questions:

• Are these values rounded, floored, or even ceiled?
• Did we use only the images in the training set of ImageNet or additionally the images of the validation set?
• Did we perform any kind of resizing or cropping on each image before the calculations were performed?

---

I've tested some combinations and will post my results here.

| Parameters | mean | std |
| --- | --- | --- |
| train set only, no resizing / cropping| [0.4803, 0.4569, 0.4083] | [0.2806, 0.2736, 0.2877] |
| train set only, resize to 256 and center crop to 224 | [0.4845, 0.4541, 0.4025] | [0.2724, 0.2637, 0.2761] |
| train set only, center crop to 224 | [0.4701, 0.4340, 0.3832] | [0.2845, 0.2733, 0.2805] |

While the means match fairly well, the std differ significantly.

[nizhib]

You need to go deeper ;)

https://github.com/facebook/fb.resnet.torch/blob/master/datasets/imagenet.lua

-- Computed from random subset of ImageNet training images
local meanstd = {
   mean = { 0.485, 0.456, 0.406 },
   std = { 0.229, 0.224, 0.225 },
}

[pmeier]

For my project I need to know the covariances between the channels. Since they are not part of the current implementation, my hope was that I can calculate them myself if I know the necessary images and processing. Unfortunately

random subset

gives me little hope that I'm able to do that. I suppose no one remembers how this random subset was selected?

Should we investigate this further? I'm a little anxious that we simply use this normalization for all our models without being able to reproduce it.

[fmassa]

@colesbury do you have more information here to clarify on the mean / std for imagenet that we use?

[soumith]

afaik we calculated the mean / std to use by running one pass on the training set of Imagenet

[soumith]

that being said, i see that std is not matching. possibly a bug of the past or some detail that we completely forgot about :-/

[apple2373]

Can we put batch normalization layer before input so that mean/std will be computed automatically in the training time?

[pmeier]

@apple2373 We currently implementing the transforms for tensors in order to be able to use them within a model (see #1375). Whether we want to include them within the models is AFAIK still up for discussion (see #782)

[pmeier]

@fmassa @soumith

Any update on this? Do we investigate further or keep it as is?

[fmassa]

@pmeier I don't know if we will ever be able to get back those numbers, given that they seem to have been computed on a randomly-sampled part of the dataset.

If we really want to see if this has any impact, we would run multiple runs of end-to-end training with the new mean/std and see if it brings any noticeable improvement.

[pmeier]

I don't think we get significant improvement (or decline) of performance. I just think we shouldn't use numbers that are not reproducible. A change like this is of course a lot of work, BC breaking etc, but we don't know what the future brings. Maybe this becomes significant in the future and than its even harder to correct.

[fmassa]

don't think we get significant improvement (or decline) of performance. I just think we shouldn't use numbers that are not reproducible.

I agree. But given the scale of how things would break with such a change, I think we should just live with it for now, and maybe document somewhere the findings you have shown in here.

[colesbury]

It's been almost four years, so I don't remember, but I probably just used the mean / std from the previous Lua ImageNet training script:

https://github.com/soumith/imagenet-multiGPU.torch/blob/deb5466a16e54ec7a69fe027e5fbcd3c1bfb49cc/donkey.lua#L161-L187

It uses the average standard deviation of an individual image's channel instead of the an estimate of the standard deviation across the entire dataset.

I don't think we should change the mean/std, nor do I see any reproducibility issue. The scientific result here is the neural network, not mean/std values. Especially since the exact choice does not matter as long as they approximately whiten the input.

[nizhib]

A change like this is of course a lot of work, BC breaking etc, but we don't know what the future brings.

These numbers have become a standard for most neural networks created so far, it's not just a lot of work — one need to retrain hundreds of neural networks (approx. 2 gpu x week each for a model like resnet50) and create pull requests for all the pretrainedmodels/dpn/wide resnets/etc. repos all over the github just to adjust normalizing std by 0.05. What the future can justify this?

[fmassa]

Following the discussion that we had in here, I agree with @colesbury and @nizhib points above.

@pmeier would you like to send a PR adding some summary of the discussion that we had here, including @colesbury comment on how those numbers were obtained?

[pmeier]

I'm covered for the next weeks. This will take some time.

[Stannislav]

Maybe the reason why the stds don't match is that it was originally called with unbiased=False?

[pmeier]

@Stannislav in #1965 I've managed to get pretty close the the original numbers.