tf.compat / v1 / v1.space_to_depth

View source on GitHub

A conditional accumulator for aggregating sparse gradients.

Inherits From: ConditionalAccumulatorBase

tf.compat.v1.SparseConditionalAccumulator(
    dtype, shape=None, shared_name=None,
    name='sparse_conditional_accumulator', reduction_type='MEAN'
)

Sparse gradients are represented by IndexedSlices.

Up-to-date gradients (i.e., time step at which gradient was computed is equal to the accumulator's time step) are added to the accumulator.

Extraction of the average gradient is blocked until the required number of gradients has been accumulated.

Methods

apply_grad

View source

apply_grad(
    grad_indices, grad_values, grad_shape=None, local_step=0, name=None
)

Attempts to apply a sparse gradient to the accumulator.

The attempt is silently dropped if the gradient is stale, i.e., local_step is less than the accumulator's global time step.

A sparse gradient is represented by its indices, values and possibly empty or None shape. Indices must be a vector representing the locations of non-zero entries in the tensor. Values are the non-zero slices of the gradient, and must have the same first dimension as indices, i.e., the nnz represented by indices and values must be consistent. Shape, if not empty or None, must be consistent with the accumulator's shape (if also provided).

Example:

A tensor [[0, 0], [0, 1], [2, 3]] can be represented indices: [1,2] values: [[0,1],[2,3]] shape: [3, 2]

apply_indexed_slices_grad

View source

apply_indexed_slices_grad(
    grad, local_step=0, name=None
)

Attempts to apply a gradient to the accumulator.

The attempt is silently dropped if the gradient is stale, i.e., local_step is less than the accumulator's global time step.

num_accumulated

View source

num_accumulated(
    name=None
)

Number of gradients that have currently been aggregated in accumulator.

set_global_step

View source

set_global_step(
    new_global_step, name=None
)

Sets the global time step of the accumulator.

The operation logs a warning if we attempt to set to a time step that is lower than the accumulator's own time step.

take_grad

View source

take_grad(
    num_required, name=None
)

Attempts to extract the average gradient from the accumulator.

The operation blocks until sufficient number of gradients have been successfully applied to the accumulator.

Once successful, the following actions are also triggered:

• Counter of accumulated gradients is reset to 0.
• Aggregated gradient is reset to 0 tensor.
• Accumulator's internal time step is incremented by 1.

take_indexed_slices_grad

View source

take_indexed_slices_grad(
    num_required, name=None
)

Attempts to extract the average gradient from the accumulator.

The operation blocks until sufficient number of gradients have been successfully applied to the accumulator.

Once successful, the following actions are also triggered:

• Counter of accumulated gradients is reset to 0.
• Aggregated gradient is reset to 0 tensor.
• Accumulator's internal time step is incremented by 1.