search

facebookresearch / higher

higher is a pytorch library allowing users to obtain higher order gradients over losses spanning training loops rather than individual training steps.
Apache License 2.0
1.58k stars 122 forks source link

Is there data leakage in the maml-omniglot example? #107

Open SunHaozhe opened 3 years ago

SunHaozhe commented 3 years ago

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 commented 2 years ago

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 commented 2 years ago

cross posted:

brando90 commented 2 years ago

related:

brando90 commented 2 years ago

@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 commented 2 years ago

create a pull request: https://github.com/facebookresearch/higher/pull/122

brando90 commented 2 years ago

@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.