When should one call .eval() and .train() when doing MAML with the PyTorch higher library?

I was going through the omniglot maml example and saw that they have net.train() at the top of their testing code. This seems like a mistake since that means the stats from each task at meta-testing is shared:

def test(db, net, device, epoch, log):
    # Crucially in our testing procedure here, we do *not* fine-tune
    # the model during testing for simplicity.
    # Most research papers using MAML for this task do an extra
    # stage of fine-tuning here that should be added if you are
    # adapting this code for research.
    net.train()
    n_test_iter = db.x_test.shape[0] // db.batchsz

    qry_losses = []
    qry_accs = []

    for batch_idx in range(n_test_iter):
        x_spt, y_spt, x_qry, y_qry = db.next('test')


        task_num, setsz, c_, h, w = x_spt.size()
        querysz = x_qry.size(1)

        # TODO: Maybe pull this out into a separate module so it
        # doesn't have to be duplicated between `train` and `test`?
        n_inner_iter = 5
        inner_opt = torch.optim.SGD(net.parameters(), lr=1e-1)

        for i in range(task_num):
            with higher.innerloop_ctx(net, inner_opt, track_higher_grads=False) as (fnet, diffopt):
                # Optimize the likelihood of the support set by taking
                # gradient steps w.r.t. the model's parameters.
                # This adapts the model's meta-parameters to the task.
                for _ in range(n_inner_iter):
                    spt_logits = fnet(x_spt[i])
                    spt_loss = F.cross_entropy(spt_logits, y_spt[i])
                    diffopt.step(spt_loss)

                # The query loss and acc induced by these parameters.
                qry_logits = fnet(x_qry[i]).detach()
                qry_loss = F.cross_entropy(
                    qry_logits, y_qry[i], reduction='none')
                qry_losses.append(qry_loss.detach())
                qry_accs.append(
                    (qry_logits.argmax(dim=1) == y_qry[i]).detach())

    qry_losses = torch.cat(qry_losses).mean().item()
    qry_accs = 100. * torch.cat(qry_accs).float().mean().item()
    print(
        f'[Epoch {epoch+1:.2f}] Test Loss: {qry_losses:.2f} | Acc: {qry_accs:.2f}'
    )
    log.append({
        'epoch': epoch + 1,
        'loss': qry_losses,
        'acc': qry_accs,
        'mode': 'test',
        'time': time.time(),
    })

however whenever I do eval instead I get that my MAML model diverges (though my test is on mini-imagenet):

>>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

So what is one suppose to do to avoid this divergence?

note:

• retraining is really expensive. Takes 18 days to train a 5cnn with maml for me. A Distributed soln would really help here https://github.com/learnables/learn2learn/issues/170
• perhaps just using train during training (even if evaluating during training might be a good idea so that the batch stats are saved in the checkpoint)
• or next time train stuff with batch stats from the beginning

related:

• https://github.com/facebookresearch/higher/issues/107
• https://discuss.pytorch.org/t/when-should-one-call-eval-and-train-when-doing-maml-with-the-pytorch-higher-library/136022
• How to use have batch norm not forget batch statistics it just used in Pytorch?
• https://discuss.pytorch.org/t/how-does-pytorch-s-batch-norm-know-if-the-forward-pass-its-doing-is-for-inference-or-training/16857/10
• https://stats.stackexchange.com/questions/544048/what-does-the-batch-norm-layer-for-maml-model-agnostic-meta-learning-do-for-du/551153#551153
• https://github.com/tristandeleu/pytorch-maml/issues/19

TLDR: Use mdl.train() since that uses batch statistics (but inference will not be deterministic anymore). You probably won't want to use mdl.eval() in meta-learning.

BN intended behaviour:

• Importantly, during inference (eval/testing) running_mean, running_std is used - that was calculated from training(because they want a deterministic output and to use estimates of the population statistics).
• During training the batch statistics is used but a population statistic is estimated with running averages. I assume the reason batch_stats is used during training is to introduce noise that regularizes training (noise robustness)
• in meta-learning I think using batch statistics is the best during testing (and not calculate the running means) since we are supposed to be seeing new /tasksdistribution anyway. Price we pay is loss of determinism. Could be interesting just out of curiosity what the accuracy is using population stats estimated from meta-trian.

This is likely why I don't see divergence in my testing with the mdl.train().

So just make sure you use mdl.train() (since that uses batch statistics https://pytorch.org/docs/stable/generated/torch.nn.BatchNorm2d.html#torch.nn.BatchNorm2d) but that either the new running stats that cheat aren't saved or used later.