SimCLR implementation.

At each iteration, we get for every data x two differently augmented versions, which we refer to as x_i and x_j. Both of these images are encoded into a one-dimensional feature vector, between which we want to maximize similarity which minimizes it to all other data in the batch. The encoder network is split into two parts: a base encoder network f(.), and a projection head g(.). The base network is usually a deep CNN or SCNN, and is responsible for extracting a representation vector from the augmented data examples. Let’s denote the representations obtained from the encoder h=f(x). The projection head g(.) maps the representation h into a space where we apply the contrastive loss, i.e., compare similarities between vectors. In the original SimCLR paper g(.) was defined as a two-layer MLP with ReLU activation in the hidden layer. Note that in the follow-up paper, SimCLRv2, the authors mention that larger/wider MLPs can boost the performance considerably.

After finishing the training with contrastive learning, we will remove the projection head g(.), and use f(.) as a pretrained feature extractor. The representations z that come out of the projection head g(.) have been shown to perform worse than those of the base network f(.) when finetuning the network for a new task. This is likely because the representations z are trained to become invariant to many features that can be important for downstream tasks. Thus, g(.) is only needed for the contrastive learning stage.

Now that the architecture is described, let’s take a closer look at how we train the model. As mentioned before, we want to maximize the similarity between the representations of the two augmented versions of the same image, i.e., z_i and z_j, while minimizing it to all other examples in the batch. SimCLR thereby applies the InfoNCE loss, originally proposed by Aaron van den Oord et al. for contrastive learning. In short, the InfoNCE loss compares the similarity of z_i and z_j to the similarity of z_i to any other representation in the batch by performing a softmax over the similarity values. The loss can be formally written as:

\ell_{i,j} = -\log \frac{\exp(\text{sim}(z_i,z_j)/\tau)}{

sum_{k=1}^{2N}mathbb{1}_{[kneq i]}

exp(text{sim}(z_i,z_k)/tau)}

= -text{sim}(z_i,z_j)/tau

+logleft[sum_{k=1}^{2N}mathbb{1}_{[kneq i]}

exp(text{sim}(z_i,z_k)/tau)right]

The function text{sim} is a similarity metric, and the hyperparameter tau is called temperature determining how peaked the distribution is. Since many similarity metrics are bounded, the temperature parameter allows us to balance the influence of many dissimilar image patches versus one similar patch. The similarity metric that is used in SimCLR is cosine similarity, as defined below:

\text{sim}(z_i,z_j) = \frac{z_i^\top \cdot z_j}{||z_i||\cdot||z_j||}

The maximum cosine similarity possible is 1, while the minimum is -1. In general, we will see that the features of two different images will converge to a cosine similarity around zero since the minimum, -1, would require z_i and z_j to be in the exact opposite direction in all feature dimensions, which does not allow for great flexibility.

Alternatively to performing the validation on the contrastive learning loss as well, we could also take a simple, small downstream task, and track the performance of the base network f(.) on that.

Parameters:

encoder : nn.Module

the encoder f(.) architecture.

hidden_dims : list of str

the projector g(.) MLP architecture.

lr : float

the learning rate.

temperature : float

the SimCLR loss temperature parameter.

weight_decay : float

the Adam optimizer weight decay parameter.

max_epochs : int, default=None

optionaly, use a CosineAnnealingLR scheduler.

random_state : int, default=None

setting a seed for reproducibility.

kwargs : dict

Trainer parameters.

Notes

A batch of data must contains two elements: two tensors with contrasted images, and a list of tensors containing auxiliary variables. Using the set_batch_connector method allows you to pass a function that reorganizes your batch of data according to these specifications.

Attributes

configure_optimizers()[source]¶

Choose what optimizers and learning-rate schedulers to use in your optimization. Normally you’d need one. But in the case of GANs or similar you might have multiple. Optimization with multiple optimizers only works in the manual optimization mode.

Return:

Any of these 6 options.

• Single optimizer.

• List or Tuple of optimizers.

• Two lists - The first list has multiple optimizers, and the second has multiple LR schedulers (or multiple lr_scheduler_config).

• Dictionary, with an "optimizer" key, and (optionally) a "lr_scheduler" key whose value is a single LR scheduler or lr_scheduler_config.

• None - Fit will run without any optimizer.

The lr_scheduler_config is a dictionary which contains the scheduler and its associated configuration. The default configuration is shown below.

lr_scheduler_config = {
    # REQUIRED: The scheduler instance
    "scheduler": lr_scheduler,
    # The unit of the scheduler's step size, could also be 'step'.
    # 'epoch' updates the scheduler on epoch end whereas 'step'
    # updates it after a optimizer update.
    "interval": "epoch",
    # How many epochs/steps should pass between calls to
    # `scheduler.step()`. 1 corresponds to updating the learning
    # rate after every epoch/step.
    "frequency": 1,
    # Metric to monitor for schedulers like `ReduceLROnPlateau`
    "monitor": "val_loss",
    # If set to `True`, will enforce that the value specified 'monitor'
    # is available when the scheduler is updated, thus stopping
    # training if not found. If set to `False`, it will only produce a warning
    "strict": True,
    # If using the `LearningRateMonitor` callback to monitor the
    # learning rate progress, this keyword can be used to specify
    # a custom logged name
    "name": None,
}

When there are schedulers in which the .step() method is conditioned on a value, such as the torch.optim.lr_scheduler.ReduceLROnPlateau scheduler, Lightning requires that the lr_scheduler_config contains the keyword "monitor" set to the metric name that the scheduler should be conditioned on.

# The ReduceLROnPlateau scheduler requires a monitor
def configure_optimizers(self):
    optimizer = Adam(...)
    return {
        "optimizer": optimizer,
        "lr_scheduler": {
            "scheduler": ReduceLROnPlateau(optimizer, ...),
            "monitor": "metric_to_track",
            "frequency": "indicates how often the metric is updated",
            # If "monitor" references validation metrics, then "frequency" should be set to a
            # multiple of "trainer.check_val_every_n_epoch".
        },
    }

# In the case of two optimizers, only one using the ReduceLROnPlateau scheduler
def configure_optimizers(self):
    optimizer1 = Adam(...)
    optimizer2 = SGD(...)
    scheduler1 = ReduceLROnPlateau(optimizer1, ...)
    scheduler2 = LambdaLR(optimizer2, ...)
    return (
        {
            "optimizer": optimizer1,
            "lr_scheduler": {
                "scheduler": scheduler1,
                "monitor": "metric_to_track",
            },
        },
        {"optimizer": optimizer2, "lr_scheduler": scheduler2},
    )

Metrics can be made available to monitor by simply logging it using self.log('metric_to_track', metric_val) in your LightningModule.

Note:

Some things to know:

• Lightning calls .backward() and .step() automatically in case of automatic optimization.

• If a learning rate scheduler is specified in configure_optimizers() with key "interval" (default “epoch”) in the scheduler configuration, Lightning will call the scheduler’s .step() method automatically in case of automatic optimization.

• If you use 16-bit precision (precision=16), Lightning will automatically handle the optimizer.

• If you use torch.optim.LBFGS, Lightning handles the closure function automatically for you.

• If you use multiple optimizers, you will have to switch to ‘manual optimization’ mode and step them yourself.

• If you need to control how often the optimizer steps, override the optimizer_step() hook.

on_validation_epoch_end()[source]¶

Called in the validation loop at the very end of the epoch.

predict_step(batch: tuple[Tensor, Sequence[Tensor]] | tuple[Tensor], batch_idx: int)[source]¶

Step function called during predict(). By default, it calls forward(). Override to add any processing logic.

The predict_step() is used to scale inference on multi-devices.

To prevent an OOM error, it is possible to use BasePredictionWriter callback to write the predictions to disk or database after each batch or on epoch end.

The BasePredictionWriter should be used while using a spawn based accelerator. This happens for Trainer(strategy="ddp_spawn") or training on 8 TPU cores with Trainer(accelerator="tpu", devices=8) as predictions won’t be returned.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

Predicted output (optional).

Example

class MyModel(LightningModule):

    def predict_step(self, batch, batch_idx, dataloader_idx=0):
        return self(batch)

dm = ...
model = MyModel()
trainer = Trainer(accelerator="gpu", devices=2)
predictions = trainer.predict(model, dm)

training_step(batch: tuple[tuple[Tensor, Tensor], Sequence[Tensor]], batch_idx: int)[source]¶

Here you compute and return the training loss and some additional metrics for e.g. the progress bar or logger.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

• Tensor - The loss tensor

• dict - A dictionary which can include any keys, but must include the key 'loss' in the case of automatic optimization.

• None - In automatic optimization, this will skip to the next batch (but is not supported for multi-GPU, TPU, or DeepSpeed). For manual optimization, this has no special meaning, as returning the loss is not required.

In this step you’d normally do the forward pass and calculate the loss for a batch. You can also do fancier things like multiple forward passes or something model specific.

Example:

def training_step(self, batch, batch_idx):
    x, y, z = batch
    out = self.encoder(x)
    loss = self.loss(out, x)
    return loss

To use multiple optimizers, you can switch to ‘manual optimization’ and control their stepping:

def __init__(self):
    super().__init__()
    self.automatic_optimization = False

# Multiple optimizers (e.g.: GANs)
def training_step(self, batch, batch_idx):
    opt1, opt2 = self.optimizers()

    # do training_step with encoder
    ...
    opt1.step()
    # do training_step with decoder
    ...
    opt2.step()

Note:

When accumulate_grad_batches > 1, the loss returned here will be automatically normalized by accumulate_grad_batches internally.

validation_step(batch: tuple[tuple[Tensor, Tensor], Sequence[Tensor]], batch_idx: int)[source]¶

Operates on a single batch of data from the validation set. In this step you’d might generate examples or calculate anything of interest like accuracy.

Args:

batch: The output of your data iterable, normally a DataLoader. batch_idx: The index of this batch. dataloader_idx: The index of the dataloader that produced this batch.

(only if multiple dataloaders used)

Return:

• Tensor - The loss tensor

• dict - A dictionary. Can include any keys, but must include the key 'loss'.

• None - Skip to the next batch.

# if you have one val dataloader:
def validation_step(self, batch, batch_idx): ...

# if you have multiple val dataloaders:
def validation_step(self, batch, batch_idx, dataloader_idx=0): ...

Examples:

# CASE 1: A single validation dataset
def validation_step(self, batch, batch_idx):
    x, y = batch

    # implement your own
    out = self(x)
    loss = self.loss(out, y)

    # log 6 example images
    # or generated text... or whatever
    sample_imgs = x[:6]
    grid = torchvision.utils.make_grid(sample_imgs)
    self.logger.experiment.add_image('example_images', grid, 0)

    # calculate acc
    labels_hat = torch.argmax(out, dim=1)
    val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)

    # log the outputs!
    self.log_dict({'val_loss': loss, 'val_acc': val_acc})

If you pass in multiple val dataloaders, validation_step() will have an additional argument. We recommend setting the default value of 0 so that you can quickly switch between single and multiple dataloaders.

# CASE 2: multiple validation dataloaders
def validation_step(self, batch, batch_idx, dataloader_idx=0):
    # dataloader_idx tells you which dataset this is.
    ...

Note:

If you don’t need to validate you don’t need to implement this method.

Note:

When the validation_step() is called, the model has been put in eval mode and PyTorch gradients have been disabled. At the end of validation, the model goes back to training mode and gradients are enabled.