utils

A collection of utility functions, mostly used for internal purposes.

Modules

keras_utils
logging
numpy_utils

Functions

add_metric(ax[, metric_text, metric_value, ...])	TODO: docstring
add_titles_and_labels(axes[, num_row, ...])	Wrapper function for configuring labels for both axes.
batched_call(f, batch_shape[, args, kwargs, ...])	Map f over the given batch shape with a for loop, preserving randomness unlike the keras built-in map apis.
calibration_curve(targets, estimates, *[, ...])	Compute true and predicted probabilities for a calibration curve.
check_estimates_prior_shapes(estimates, ...)	Checks requirements for the shapes of estimates and prior_samples draws as necessitated by most diagnostic functions.
check_lengths_same(*args)
concatenate_valid(tensors[, axis])	Concatenate multiple tensors along axis, ignoring None values.
confusion_matrix(targets, estimates[, ...])	Compute confusion matrix to evaluate the accuracy of a classification or model comparison setting.
convert_args(f, *args, **kwargs)	Convert positional and keyword arguments to just positional arguments for f
convert_kwargs(f, *args, **kwargs)	Convert positional and keyword arguments qto just keyword arguments for f
deserialize_value_or_type(config, name)	Deserialize an object that can be either a value or a type and add it to the supplied dictionary.
detailed_loss_callback()	Provides detailed callback for loss trajectory for every training step.
devices()	Returns a list of available GPU devices.
expand(x, n, side)
expand_as(x, y, side)
expand_left(x, n)	Expand x to the left n times
expand_left_as(x, y)	Expand x to the left, matching the dimension of y
expand_left_to(x, dim)	Expand x to the left, matching dim
expand_right(x, n)	Expand x to the right n times
expand_right_as(x, y)	Expand x to the right, matching the dimension of y
expand_right_to(x, dim)	Expand x to the right, matching dim
expand_tile(x, n, axis)	Expand and tile x along the given axis n times
expand_to(x, dim, side)
fill_triangular_matrix(x[, upper, positive_diag])	Reshapes a batch of matrix elements into a triangular matrix (either upper or lower).
filter_kwargs(kwargs, f)	Filter keyword arguments for f
find_batch_size(sample[, memory_budget, ...])	Returns an estimation of an optimal batch size based on memory budget and sample memory footprint.
find_distribution()
find_inference_network()
find_memory_budget([device_type])	Returns an estimation of available memory in bytes for the given device type.
find_network()
find_permutation()
find_pooling()
find_recurrent_net()
find_summary_network()
format_bytes(b[, precision, si])	Format a number of bytes as a human-readable string in the format '{value} {prefix}{unit}'.
integrate(fn, state[, start_time, ...])
jacobian(f, x[, return_output])	Compute the Jacobian matrix of f with respect to x.
jacobian_trace(f, x[, max_steps, ...])	Compute or estimate the trace of the Jacobian matrix of f.
jvp(f, x, tangents[, return_output])	Compute the Jacobian-vector product of f at x with tangents.
layer_kwargs(kwargs)	Filter keyword arguments for keras.Layer
make_quadratic(ax, x_data, y_data)	Utility to make a subplots quadratic in order to avoid visual illusions in, e.g., recovery plots.
model_kwargs(kwargs)	Filter keyword arguments for keras.Model
optimal_transport(x1, x2[, method, ...])	Matches elements from x2 onto x1, such that the transport cost between them is minimized, according to the method and cost matrix used.
pad(x, value, n, axis[, side])	Pad x with n values along axis on the given side.
parse_bytes(s)	Parse a string in the format '{value} {prefix}{unit}' and return the number of bytes, flooring to the nearest integer.
pickle_load(filepath)	Small utility to safely load a pickle object.
prepare_plot_data(estimates, targets[, ...])	Procedural wrapper that encompasses all preprocessing steps, including shape-checking, parameter name generation, layout configuration, figure initialization, and collapsing of axes.
prettify_subplots(axes, num_subplots[, ...])	TODO: docstring
searchsorted(sorted_sequence, values[, side])	Find indices where elements should be inserted to maintain order.
sequential_kwargs(kwargs)	Filter keyword arguments for keras.Sequential
serialize_value_or_type(config, name, obj)	Serialize an object that can be either a value or a type and add it to a copy of the supplied dictionary.
simultaneous_ecdf_bands(num_estimates[, ...])	Computes the simultaneous ECDF bands through simulation according to the algorithm described in Section 2.2 of
size_of(x)
split_arrays(data[, axis])	Split tensors in the dictionary along the given axis.
split_tensors(data[, axis])	Split tensors in the dictionary along the given axis.
squeeze_inner_estimates_dict(estimates)	If a dictionary has only one key-value pair and the key is "value", return only its value.
stack_valid(tensors[, axis])	Stack multiple tensors along axis, ignoring None values.
tile_axis(x, n, axis)	Tile x along the given axis n times
tree_concatenate(structures[, axis, numpy])	Concatenate all tensors in the given sequence of nested structures.
tree_stack(structures[, axis, numpy])	Like tree_concatenate(), except tensors are stacked instead of concatenated.
vjp(f, x[, return_output])	Compute the vector-Jacobian product of f at x.
weighted_mean(elements[, weights])	Compute the (optionally) weighted mean of the input tensor.