[bugfix] Logging only on not should_accumulate() during training

[tchaton]

What does this PR do?

This PR fixes visual logging with accumulated_grad_batches > 1.

Solution: We can't assume the metrics can be averaged, so we will log only when optimizer_step will be called.

Previously

Now.

Code used to generate the visualization

def test_logging_with_accumulate_grad_batches(tmpdir):
    class LitClassifier(pl.LightningModule):
        def __init__(self, hidden_dim=128, learning_rate=1e-3):
            super().__init__()
            self.save_hyperparameters()

            self.l1 = torch.nn.Linear(28 * 28, self.hparams.hidden_dim)
            self.l2 = torch.nn.Linear(self.hparams.hidden_dim, 10)

            self.train_acc = pl.metrics.Accuracy()

        def forward(self, x):
            x = x.view(x.size(0), -1)
            x = torch.relu(self.l1(x))
            return self.l2(x)

        def training_step(self, batch, batch_idx):
            x, y = batch
            y_hat = self(x)
            loss = F.cross_entropy(y_hat, y)
            self.log('train_acc', self.train_acc(y_hat, y), on_step=True, on_epoch=True, prog_bar=True)
            self.log("train_loss",loss)
            return loss

        def validation_step(self, batch, batch_idx):
            x, y = batch
            y_hat = self(x)
            loss = F.cross_entropy(y_hat, y)

        def configure_optimizers(self):
            return torch.optim.Adam(self.parameters(), lr=self.hparams.learning_rate)

    def run_test(model, max_epochs, accumulate_grad_batches, batch_size, num_workers=4):
        dataset = MNIST('', train=True, download=True, transform=transforms.ToTensor())
        mnist_train, mnist_val = random_split(dataset, [55000, 5000])
        train_loader = DataLoader(mnist_train,batch_size)
        val_loader = DataLoader(mnist_val,batch_size)

        trainer = pl.Trainer(
            logger=WandbLogger(name="bug", project='.....', save_dir=".", log_model=False),
            accumulate_grad_batches=accumulate_grad_batches,
            limit_train_batches=100,
            log_every_n_steps=5,
            max_epochs=max_epochs
            )
        trainer.fit(model, train_loader, val_loader)

    model = LitClassifier()

    run_test(model, 3, 1, 32)
    run_test(model, 3, 8, 32)

Fixes #5405 <- this links related issue to this PR

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified
• Check that target branch and milestone match!

Did you have fun?

Make sure you had fun coding 🙃

[codecov]

Codecov Report

Merging #5417 (20b1abf) into master (f2e99d6) will not change coverage.
The diff coverage is 100%.

@@          Coverage Diff           @@
##           master   #5417   +/-   ##
======================================
  Coverage      93%     93%           
======================================
  Files         134     134           
  Lines        9996    9996           
======================================
  Hits         9313    9313           
  Misses        683     683           

[carmocca]

Not familiar with W&B. Can you see the raw data for the epoch graph? Doesn't look right, does it?

[tchaton]

Not familiar with W&B. Can you see the raw data for the epoch graph? Doesn't look right, does it?

It is contracted to log on number of optimizer_step.

Here are the other options which I didn't prefer.

[teddykoker]

Great fix :)

[Borda]

lgtm, just check back-compatibity and chlog

[Borda]

is it used by a user, right? then let's hold back compatibility with API and add a warning...

[tchaton]

It is internal. LoggerConnector class

[tchaton]

Not it is not.