Structures

The helper to collect data from the given Envs.

The policy is intentional to be flexible as a simple closure. We collect both the action and state data the closure provides. There may be some type conversions for what the Env expects and what the policy provides, thus, need to specify both. To make this easier to use, the placeholder type would be useful: Collector<Float, (), _, _>(envs: envs) { ... }

A pandas-inspired dataframe. Dataframe is a tabular data representation. This particular Dataframe implementation is most useful to implement data feeder pipeline. You need some transformations so some text or images can be transformed into tensors for a model to consume. Dataframe can be used to implement that pipeline.

Hint are parameters to these operations that changes shape from input to output. It given proper stride / padding parameters for these operations.

Softmax cross-entropy loss. This combines softmax with cross-entropy loss to maximize numerical stability.

Sigmoid cross-entropy loss. This combines sigmoid with binary cross-entropy loss to maximize numerical stability.

Binary cross-entropy loss.

Multi-class cross-entropy loss.

Smooth L1 loss (for object detection).

MSE loss. Currently it does reduce mean.

Basic tensor type.

A simple helper to compute numerical statistic properties (mean and standard deviations) from a given collection.

Stochastic gradient descent optimizer.

Adam optimizer.

LAMB optimizer.

AdamW optimizer.

A writer for writing model execution summaries to a tensorboard-readable file; the summaries include scalars for logging statistics, graphs for visualizing model etc.

Tensor shape (or dimensions)