composer.core.types

This is our reference for common types used throughout our library.

Tensor Types

class composer.core.types.Tensor: Alias for torch.Tensor

class composer.core.types.Tensors

Commonly used to represent e.g. a set of inputs, where it is unclear whether each input has its own tensor, or if all the inputs are concatenated in a single tensor.

Type: torch.Tensor, tuple of torch.Tensors, or list of torch.Tensors

Batch Types

A batch of data can be represented in several formats, depending on the application.

class composer.core.types.Batch

Union type covering the most common representations of batches.

Type: BatchPair, BatchDict, or torch.Tensor

class composer.core.types.BatchPair

Commonly used in computer vision tasks. The object is assumed to contain exactly two elements, where the first represents inputs and the second represents targets.

Type: List of Tensors

class composer.core.types.BatchDict

Commonly used in natural language processing tasks.

Type: str to Tensor dict

We provide helper functions to cast batches in situations where the batch type is known.

composer.core.types.as_batch_dict(batch: composer.core.types.Batch) → composer.core.types.BatchDict[source]

Casts a Batch as a BatchDict.

Parameters: batch (Batch) – A batch.
Raises: TypeError – If the batch is not a BatchDict.
Returns: BatchDict – The batch, represented as a BatchDict.

composer.core.types.as_batch_pair(batch: composer.core.types.Batch) → composer.core.types.BatchPair[source]

Casts a Batch as a BatchPair.

Parameters: batch (Batch) – A batch.
Returns: BatchPair – The batch, represented as a BatchPair.
Raises: TypeError – If the batch is not a BatchPair.

Dataset and Data Loader Types

class composer.core.types.Dataset[source]: Alias for torch.utils.data.Dataset[Batch]

class composer.core.types.DataLoader(*args, **kwargs)[source]

Protocol for custom DataLoaders compatible with torch.utils.data.DataLoader.

dataset

Dataset from which to load the data.

Type: Dataset

batch_size

How many samples per batch to load (default: 1).

Type: int, optional

num_workers

How many subprocesses to use for data loading. 0 means that the data will be loaded in the main process.

Type: int

pin_memory

If True, the data loader will copy Tensors into CUDA pinned memory before returning them.

Type: bool

drop_last

If len(dataset) is not evenly divisible by batch_size, whether the last batch is dropped (if True) or truncated (if False).

Type: bool

timeout

The timeout for collecting a batch from workers.

Type: float

sampler

The dataloader sampler.

Type: torch.utils.data.Sampler[int]

prefetch_factor

Number of samples loaded in advance by each worker. 2 means there will be a total of 2 * num_workers samples prefetched across all workers.

Type: int

Trainer Types

class composer.core.types.Metrics

Union type covering common formats for representing metrics.

Type: Metric or MetricCollection

class composer.core.types.Optimizer[source]: Alias for torch.optim.Optimizer

class composer.core.types.Optimizers

Union type for indeterminate amounts of optimizers.

Type: Optimizer or tuple of Optimizer

class composer.core.types.Scheduler: Alias for torch.optim.lr_scheduler._LRScheduler

class composer.core.types.Schedulers

Union type for indeterminate amounts of schedulers.

Type: Scheduler or tuple of Scheduler

class composer.core.types.Scaler: Alias for torch.cuda.amp.grad_scaler.GradScaler

class composer.core.types.Model: Alias for torch.nn.Module

Miscellaneous Types

class composer.core.types.ModelParameters

Type alias for model parameters used to initialize optimizers.

Type: List of torch.Tensor or list of str to torch.Tensor dicts

class composer.core.types.JSON: JSON data.

class composer.core.types.Precision(value)[source]

Enum class for the numerical precision to be used by the model.

AMP: Use torch.cuda.amp. Only compatible with GPUs.

FP32: Use 32-bit floating-point precision. Compatible with CPUs and GPUs.

class composer.core.types.TPrefetchFn

Type alias for prefetch functions used for initializing dataloaders.

Type: A function that takes a Batch and returns a Batch

class composer.core.types.Serializable(*args, **kwargs)[source]

Interface for serialization; used by checkpointing.

load_state_dict(state: composer.core.types.StateDict) → None[source]

Restores the state of the object.

Parameters: state (StateDict) – The state of the object, as previously returned by state_dict()

state_dict() → composer.core.types.StateDict[source]

Returns a dictionary representing the internal state.

The returned dictionary must be pickale-able via torch.save().

Returns: StateDict – The state of the object

class composer.core.types.StateDict

Class for representing serialized data.

Type: dict with str keys.

exception composer.core.types.BreakEpochException[source]

Raising this exception will immediately end the current epoch.

If you’re wondering whether you should use this, the answer is no.