Dense Tensors

Documentation status: reference — see Maturity and evidence.

Dense tensors store values in regular numerical arrays optimized for conventional tensor operations.

When to use them

Dense representations are appropriate when:

• the shape is regular;
• most cells have meaningful values;
• vectorized numerical kernels are desirable;
• the computation is already expressed as matrices or multidimensional arrays.

Initialization and trainable values

A dense tensor can be initialized from explicit values or supported initialization strategies. Trainable tensors participate in the computation graph and can receive gradients when the selected execution path supports learning.

Computation graphs

Dense tensors can be chained through arithmetic, linear algebra, activations, losses, and other operations. Forward evaluation computes values; backward evaluation propagates gradients where supported.

Relationship with conceptual tensors

A conceptual tensor preserves semantic axes and sparse conceptual structure. A dense tensor emphasizes efficient numerical layout. Convert explicitly when moving from one representation to the other.