Is there data leakage in the maml-omniglot example?

[SunHaozhe]

In the maml-omniglot.py example code, net.train() is used for meta-test phases (link).

Does this not cause data leakage of meta-test data via the statistics of nn.BatchNorm2d (net contains several nn.BatchNorm2d)?

[brando90]

I think there is.

However, in my experience without it the model diverges and explodes after an adaptation step (i.e. e.g. 5 steps of the inner opt):

>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5939, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5940, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5940, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5940, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5941, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5940, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5942, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5940, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5940, grad_fn=<NormBackward1>)
>maml_old (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>>maml_old (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5939, grad_fn=<NormBackward1>)
eval_loss=0.9859228551387786, eval_acc=0.5907692521810531
args.meta_learner.lr_inner=0.01
==== in forward2
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(171440.6875, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(208426.0156, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(17067344., grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(40371.8125, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(1.0911e+11, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(21.3515, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(5.4257e+13, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(128.9109, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(3994.7734, grad_fn=<NormBackward1>)
>maml_new (before inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(9.5937, grad_fn=<NormBackward1>)
>maml_new (after inner adapt): fmodel.model.features.conv1.weight.norm(2)=tensor(1682896., grad_fn=<NormBackward1>)
eval_loss_sanity=nan, eval_acc_santiy=0.20000000298023224

though these are on mini-imagenet but I am 100% that the issue causing it. when I do mdl.train() it goes away...

[brando90]

cross posted:

• https://stackoverflow.com/questions/69845469/when-should-one-call-eval-and-train-when-doing-maml-with-the-pytorch-highe
• https://discuss.pytorch.org/t/when-should-one-call-eval-and-train-when-doing-maml-with-the-pytorch-higher-library/136022

[brando90]

related:

• https://stackoverflow.com/questions/69846779/how-does-one-use-the-mean-and-std-from-training-in-batch-norm
• How does one return an adapted model without using the context manager? #119

[brando90]

@SunHaozhe did you ever fix this...?

I think the only way to fix it (and idk if it will work) is to either

1. re-train the model but use batch stats (+ maml)
2. make sure your using the train stats...? https://stackoverflow.com/questions/69846779/how-does-one-use-the-mean-and-std-from-training-in-batch-norm

[brando90]

create a pull request: #122

[brando90]

@SunHaozhe I don't think this is an issue anymore because of this:

this is likely wrong because during training we want to use batch statistics in meta-leanring, since tasks have different distributions. But then how do we retain determinism at inference time? see: https://discuss.pytorch.org/t/how-does-one-use-the-mean-and-std-from-training-in-batch-norm/136029/5

in summary .train() uses batch statistics, yes it updates the running mean with cheating means but those are never actually used if the .train() is set. During training, the network uses batch statistics anyway. As long as you don't save the model as a checkpoint with these cheated means it doesn't matter. But your code does become less determinsitic.