Evaluation code of references is slightly off

[NicolasHug]

There is a subtle known bug in the evaluation code of the classification references (and other references as well, but not all):

vision/references/classification/train.py

Lines 65 to 66 in 261cbf7

	# FIXME need to take into account that the datasets
	# could have been padded in distributed setup

It deserves some attention, because it's easy to miss and yet can impact our reported results, and those of research papers.

As the comment above describes, when computing the accuracy of the model on a validation set in a distributed setting, some images will be counted more than once if len(dataset) isn't divisible by batch_size * world_size ¹.

On top of that, since the test_sampler uses shuffle=True by default, the duplicated images aren't even the same across executions, which means that evaluating the same model on the same dataset can lead to different results every time.

Should we try to fix this, or should we just leave it and wait for the new lightning recipes to handle it? And as a follow-up question, is there a builtin way in lightning to mitigate this at all? (I'm not familiar with lightning, so this one may not make sense.)

cc @datumbox

Footnotes

1. For example if we have 10 images and 2 workers with a batch_size of 3, we will have something like:

   worker1: img1, img2, img3
   worker2: img4, img5, img6
   worker1: img7, img8, img9
   worker2: img10, **img1, img2** 
                     ^^^^^^^^^
    "padding": duplicated images which will affect the validation accuracy
   

   ↩

[datumbox]

Thanks for raising.

Here are some thoughts:

1. Reporting slightly biased val stats during the training process is less concerning. I agree that this is concerning when we report final results.
2. We could change the batch_size to 1 similar to what we do on detection. If that slows down the epochs too much we can try doing it only when --test-only is passed. This way we will be able to report correct stats for our models.
3. Setting shuffle=False for testing is IMO something worth doing anyway to get more determinism.
4. We could also try setting drop_last=True on the DataLoader during training. Not sure it improves things by much though.

I think lightning is likely to face similar issues but for now I think it's worth patching batch_size=1 when --test-only and settings shuffle=False on the validation sampler. This can be a good bootcamp task BTW. Let me know what you think.

[NicolasHug]

Thanks for sharing your thoughts @datumbox

I'm not sure that setting the batch_size to 1 will avoid the issue: we'll still get duplicated images if len(dataset) isn't divisible by world_size. For example with 10 samples, 3 workers and batch_size=1 I think we get something like this:

I'm rank 0 and I got tensor([0]))
I'm rank 1 and I got tensor([1]))
I'm rank 2 and I got tensor([2]))
I'm rank 0 and I got tensor([3]))
I'm rank 1 and I got tensor([4]))
I'm rank 2 and I got tensor([5]))
I'm rank 0 and I got tensor([6]))
I'm rank 1 and I got tensor([7]))
I'm rank 2 and I got tensor([8]))
I'm rank 0 and I got tensor([9]))
I'm rank 1 and I got tensor([0]))  <-- dup
I'm rank 2 and I got tensor([1]))  <-- dup

source:

torchrun --nproc_per_node=3 --nnodes 1 lol.py

import os
import torch

from torch.utils.data import Dataset, DataLoader

class MyDataset(Dataset):
    def __getitem__(self, index):
        return index

    def __len__(self):
        return 10


local_rank = int(os.environ["LOCAL_RANK"])
rank = int(os.environ["RANK"])
world_size = int(os.environ["WORLD_SIZE"])

torch.distributed.init_process_group(backend="gloo", rank=rank, world_size=world_size, init_method="env://")

dataset = MyDataset()
sampler = torch.utils.data.distributed.DistributedSampler(dataset, shuffle=False, drop_last=False)
dataloader = DataLoader(dataset, sampler=sampler, batch_size=1)
for e in dataloader:
    for r in range(world_size):
        torch.distributed.barrier()
        if r == rank:
            print(f"I'm rank {rank} and I got {e})")

[datumbox]

You can try setting batch_size=1 and running on single GPU if you really want to avoid the issue. You could also do drop_last=True as said above to drop the last sample, but now you will lose img10. In 99.999%, this doesn't worth the effort.

There is a trick you can use for some types of models that allows you to identify padded records from real data directly from the output without storing additional info.The trick it to pad your batch to a fix size using NaN values. The prediction outputs of these records will be NaNs as well, which makes it easy to distinguish and filter from real data without additional info.

Here are types of models where this is not possible:

• Unstable models that produce NaNs randomly
• Models that have post-processing steps that filter or aggregate the CNN outputs (like in Detection)
• GANs that have to produce results on training mode instead of inference mode
• Exotic layers/techniques that keep using mini-batch stats across images even on inference mode

Here is a POC which takes all the Classification and Segmentation models of TorchVision and shows that such padding is possible:

import torch
import torchvision

# image 0 has valid data. Image 1 is used only for pad and contains only nans
batch = torch.randn((2, 3, 224, 224))
batch[1] *= float("nan")

contains_nans = lambda t: torch.isnan(t).any()
get_models = lambda module: [k for k, v in module.__dict__.items() if callable(v) and k[0].lower() == k[0] and k[0] != "_"]


print("Classification:")
for model_name in get_models(torchvision.models):
    model = torchvision.models.__dict__[model_name](pretrained=False).eval()
    try:
        with torch.no_grad():
            out = model(batch)

            assert not contains_nans(out[0]), "Found nans on the real image!"
            assert contains_nans(out[1]), "No nan data found on the padded image!"
    except Exception as e:
        print(f"[FAIL] {model_name}: {e}")
        continue
    print(f"[SUCCESS] {model_name}")


print("Segmentation:")
for model_name in get_models(torchvision.models.segmentation):
    model = torchvision.models.segmentation.__dict__[model_name](pretrained=False, pretrained_backbone=False).eval()
    try:
        with torch.no_grad():
            out = model(batch)['out']

            assert not contains_nans(out[0]), "Found nans on the real image!"
            assert contains_nans(out[1]), "No nan data found on the padded image!"
    except Exception as e:
        print(f"[FAIL] {model_name}: {e}")
        continue
    print(f"[SUCCESS] {model_name}")

Personally I find this too much of a hassle. The mitigations I described on my previous comment should be enough for most users/use-cases.

[NicolasHug]

You could also do drop_last=True as said above to drop the last sample, but now you will lose img10

drop_last=True will still drop some images, but it can drop more than just the last image. For example with len(dataset) == 10, batch_size=2 and world_size=6, the last 4 images will be dropped. Agreed, this is a weird setting, but it can still significantly influence the validation result, especially on small datasets.

As far as I can tell, the only correct way is to have batch_size == world_size == 1 as you already noted, or to have a combination such that len(dataset) % (batch_size * world_size) == 0.

I agree that in some cases, e.g. during training, we don't necessarily need to have 100% accurate results and the current behaviour is good enough.

But I think we should still aim at providing exact validation results on all models. It's very likely that some users / researchers are using this code without realizing these subtle caveats, and we need to prevent them from reporting incorrect results.

[datumbox]

I don't think there is any disagreement on how each option works or on their pros/cons. The reference scripts are there as a starting point for people to implement their own solutions and adjust them to their needs. If their need is to report precise numbers, then they would have to implement one of the solutions outlined above.

Currently the scripts are tuned to our needs which are to train our pre-trained models. In our use-case the datasets are big and the padded examples won't change noticeably the accuracy. So unless one proves that the effect on the accuracy of the current models is massive, I'm not sure implementing a solution that will increase the complexity of the scripts would be worth it.

Perhaps @fmassa can provide background info on whether this was ever a problem to any of the research work he was involved, as many of these scripts are ported from those repos.

[NicolasHug]

If their need is to report precise numbers, then they would have to implement one of the solutions outlined above.

In general, people do need to report precise numbers. I think we do, and I think our users expect accurate, correct results as well. OTOH I don't think it's fair for us to expect our users to identify these subtle bugs or to be aware of them, so I wouldn't expect them to use anything else but what we provide them with.

We're telling users that they can rely on our reference scripts for training - why should it be any different for validation?

If we identify a bug, we owe it to our users to try to fix it. I disagree with the sentiment that this doesn't matter, however small the difference in results might be.

[fmassa]

Hi,

There are a lot of points being discussed in here. Let's split this up in parts:

Simple way of fixing this for classification

Just ask the users to evaluate on single-GPU, for any batch size. We handle this case properly already (and we keep drop_last=False in the dataloader).

Batch size 1

This is actually only necessary for detection models, as we perform padding on the images. So in order to keep results consistent we evaluate with batch size 1.

But note that the detection references are the only one that currently handle distributed properly. What we do is to return the image id together with the image / target in the dataset, and keep a map of the {image_id: acc} for every image in the dataset per process. Then we aggregate over all processes (which will remove duplicates) and then we average out for the final numbers.

IMO this is the most robust way of fixing this, but this introduces changes to the rest of the pipeline.

If we note it carefully, a key part of this approach is to return the image_id together with the data in the dataset -- the refactoring of the datasets does it natively, so it will open the door for us to get this implemented in a clean and concise way everywhere

[datumbox]

@fmassa I agree a simple approach is preferred here. Increasing further the complexity of the reference scripts is going towards the opposite direction of where we aim. The STL seems a much better place to handle such details.

@NicolasHug I think it was obvious but to avoid misinterpretations of what I said, by "precise numbers" I meant differences on the 5th decimal point. We typically report results up to 1st or 3rd digit on models, so I believe changes wont be visible. But you are welcome to quantify the change and show otherwise.

[NicolasHug]

I ran

torchrun --nproc_per_node=8 references/classification/train.py --model resnet18 --test-only --pretrained

65 times on main and here are the results. The variance affects the results around the 2nd / 3rd decimal, so IMHO it's high enough for us to address this.

Acc@1: resnet18
count    65.000000
mean     69.762923
std       0.003555
min      69.754000
25%      69.760000
50%      69.764000
75%      69.766000
max      69.770000
dtype: float64

Acc@5: resnet18
count    65.000000
mean     89.070246
std       0.002194
min      89.068000
25%      89.068000
50%      89.070000
75%      89.072000
max      89.076000
dtype: float64

Regarding the suggestion to use a single GPU and set shuffle=False, I agree this is a reasonable temporary solution, until we address this problem properly, and provided that this yields consistently the same results. However it looks like there are additional sources of randomness in the results - would you guys mind taking a look at #4587 ?

EDIT: thanks @datumbox for checking #4587. It seems clear from there that running the validation on a single GPU isn't enough to get exact and consistent results. We still need to have len(dataset) % (batch_size * world_size) == 0.

[datumbox]

Thanks for running experiments to quantify the size of the effect.

The variation is not massive but noticeable. It might be worth applying some of the temporary solutions described above. To summarize we could initialize the test_sampler with shuffle=False and when --test-only is provided, we could check if the len(dataset) % (batch_size * world_size) == 0. If that's not true, one approach is to make the necessary changes on args (such as setting batch and word size to 1) to ensure the user gets correct stats. Another way is to raise a warning about the variation and explain more visibly why this happens; I think the current comment on the code is a bit hidden.

For a permanent fix, this is probably something we should investigate on the new recipes where we will have the opportunity to inherit code. I remember @kazhang raising this in the past in one of our previous discussions. We need to find a solution that works on all models and all datasets, for example the aforementioned image_id approach is only possible on COCO.

[NicolasHug]

we could initialize the test_sampler with shuffle=True

Did you mean shuffle=False?

For the rest I agree on all fronts :)

I'll submit a PR soon.

[datumbox]

Did you mean shuffle=False?

Doh... Yes I meant False. I'll edit in place to avoid confusion of future readers. :p