slmsuite.holography.algorithms.MultiplaneHologram

class MultiplaneHologram(holograms, weights=None)

Bases: Hologram

Holography combining multiple objectives, potentially across planes of focus or color. Other Hologram subclasses are restricted to either optimizing a hologram within a fixed basis of spots or within the grid of a discrete Fourier transform at a fixed plane of focus. This MultiplaneHologram acts as a metaclass to optimize many individual holograms simultaneously—over many planes or pointsets—producing a composite phase pattern.

Note

Though the infrastructure to make this trivial is not yet in place, the idea of a ‘plane’ extends to planes of color. That is, this class MultiplaneHologram could also be used to optimize a multicolor hologram and account for how the farfield of each color scales with wavelength.

Tip

Calls to optimize() which update flags also update the flags of any child hologram.

holograms

List of sub-holograms to optimize simultaneously.

Type:

list of Hologram

weights

Weight for each hologram. This allows the user to redistribute power between holograms. Keep in mind that each hologram will normalize itself, so differences in intensity between target patterns cannot be relied upon.

Type:

list of float

Methods

get_amp	Collects the current nearfield amplitude regardless of np/cp configuration.
get_farfield	Collects the current complex DFT farfield, potentially with transformations.
get_log	Return this object's log records (only), as a list of plain-text strings.
get_mempool_limit	Helper function to get the cupy memory pool size.
get_multiplane_defocus_blur	From a stack of target (power) images at target_depths, generate a stack of images at return_depths, accounting for defocus blur.
get_padded_shape	Helper function to calculate the shape of the computational space.
get_phase	Collects the current nearfield phase from the GPU with cupy.ndarray.get().
get_weights	Returns the current weights.
load_stats	Uses load_h5() to import the statistics hierarchy from a given h5 file.
log_state	Log the current values of all tracked attributes.
measure	Raises NotImplementedError; FeedbackHologram measures the camera.
optimize	Optimizes to solve the "phase problem": approximating the nearfield phase that transforms a known nearfield source amplitude to a desired farfield target amplitude.
optimize_cg	Gradient-based iterative phase retrieval via torch autograd and a torch.optim optimizer (the "CG" method; see optimize()).
optimize_gs	GPU-accelerated Gerchberg-Saxton (GS) iterative phase retrieval.
pickle	Returns a dictionary containing selected attributes of this class.
plot_farfield	Plots an overview (left) and zoom (right) view of source.
plot_nearfield	Plots the amplitude (left) and phase (right) of the nearfield (plane of the SLM).
plot_stats	Plots the statistics contained in the given dictionary.
remove_vortices	Removes each child's farfield phase vortices; see Hologram.remove_vortices().
reset	Resets the hologram to an initial state.
reset_phase	Randomize or set the meta phase.
reset_weights	Resets the hologram weights to the target defaults.
save	Saves the dictionary returned from pickle() to a file like "path/name_id.h5".
save_stats	Uses save_h5() to export the statistics hierarchy to a given h5 file.
set_mempool_limit	Helper function to set the cupy memory pool size.
set_propagation_kernels	Set every child's propagation_kernel at once from a stacked array.
set_target	Change the target to something new.
set_weights	Sets the weights to a new value.

__init__(holograms, weights=None)

Initializes a ‘meta’ hologram consisting of several sub-holograms optimizing at the same time.

Parameters:

• holograms (list of Hologram) – List of N sub-holograms to optimize simultaneously.

• weights (array_like of float OR None) – List of N floats. If None, defaults to even power.

get_amp()

Collects the current nearfield amplitude regardless of np/cp configuration.

Returns:

Nearfield amplitude.

Return type:

numpy.ndarray

get_farfield(shape=None, propagation_kernel=None, affine=None, get=True)

Collects the current complex DFT farfield, potentially with transformations. This includes collecting the data from the GPU with cupy.ndarray.get().

Parameters:

• shape ((int, int)) – Shape of the DFT. Useful to change the resolution of the farfield. If None, defaults to shape, and falls back to slm_shape.

• propagation_kernel (array_like) – Used to check the result of the hologram at different depths. See propagation_kernel. If None, defaults to propagation_kernel if one is present. Otherwise, no kernel is applied. Zeroing can force no kernel to be applied and yield the raw DFT.

• affine (dict) – Affine transformation to apply to farfield data (in the form of a dictionary with keys "M" and "b"). If None, no transformation is applied.

• get (bool) – Whether or not to convert the cupy array to a numpy array if cupy is used. This is ignored if numpy is used.

Returns:

Current farfield expected from the current phase.

Return type:

numpy.ndarray

get_log(verbose=False)

Return this object’s log records (only), as a list of plain-text strings.

Parameters:

verbose (bool, optional) – If True, also print the records to the console.

static get_mempool_limit(device=0)

Helper function to get the cupy memory pool size.

Parameters:

device (int) – Which GPU to set the limit on. Passed to cupy.cuda.Device().

Returns:

Current memory pool limit in bytes

Return type:

int

static get_multiplane_defocus_blur(cameraslm, targets, target_depths, return_depths=None, sharp_focus=True)

From a stack of target (power) images at target_depths, generate a stack of images at return_depths, accounting for defocus blur. Power is summed as if all depths were transparent; i.e. objects do not block objects further behind. This is a partial farfield implementation of realistic defocus blur.

Warning

This feature seems to lead to less stable holography without, perhaps, some additional optimizations.

Parameters:

• cameraslm (FourierSLM) –

  Hardware to implement blur for. Calibrations are necessary to determine how much to blur.

  Tip

  Right now, the blurring is Gaussian and analytic, but in the future, the measurement of the point spread function should be used.

• targets (array_like) – Stack of images of shape (image_count, h, w).

• target_depths (list of float) – Depths in "kxy" focal power units corresponding to the targets.

• return_depths (list of float) – Depths to return images at. If None, use target_depths.

• sharp_focus (bool) – If False, depths at focal planes are blurred by the point spread radius of a focused spot. If True, all the blurring is reduced by the focused point spread radius, keeping images that are in focus sharp.

static get_padded_shape(slm_shape, padding_order=1, square_padding=True, precision=inf, precision_basis='kxy')

Helper function to calculate the shape of the computational space. For a given base slm_shape, pads to the user’s requirements. If the user chooses multiple requirements, the largest dimensions for the shape are selected. By default, pads to the smallest square power of two that encapsulates the original slm_shape.

See also

Hologram

Note

Under development: a parameter to pad based on available memory (see _calculate_memory_constrained_shape()).

Parameters:

• slm_shape ((int, int) OR FourierSLM OR SLM) – The original shape of the SLM in numpy (h, w) form. The user can pass a FourierSLM or SLM instead, and should pass this when using the precision parameter.

• padding_order (int) – Scales to the padding_order th larger power of 2. A padding_order of zero does nothing. For instance, with square_padding=False, an SLM with shape (720, 1280) would yield (720, 1280) for padding_order=0, (1024, 2048) for padding_order=1, and (2048, 4096) for padding_order=2.

• square_padding (bool) – If True, sets the smaller shape dimension to that of the larger, yielding a square.

• precision (float) – Returns the shape that produces a computational k-space with resolution smaller than precision. The default, infinity, requests a padded shape larger than zero, so padding_order will dominate.

• precision_basis (str) – Basis for the precision. Can be "ij" (camera) or "kxy" (normalized blaze).

Returns:

Shape of the computational space which satisfies the above requirements.

Return type:

(int, int)

get_phase(include_propagation=False)

Collects the current nearfield phase from the GPU with cupy.ndarray.get(). Also offsets by \(\pi\), which shifts the \([-\pi, \pi]\) range of numpy.arctan2() to \([0, 2\pi]\) for faster writing to the SLM (see set_phase()). The result is not wrapped, so a phase from CG, a quadratic initialization, or an included propagation_kernel can fall outside this range.

Parameters:

include_propagation (bool) – Whether to include the propagation_kernel, if available.

Returns:

Current nearfield phase of the optimization.

Return type:

numpy.ndarray

get_weights()

Returns the current weights. Collects the current weights from the GPU if applicable.

load_stats(file_path, include_state=True)

Uses load_h5() to import the statistics hierarchy from a given h5 file.

Parameters:

• file_path (str) – Full path to the file to read the data from.

• include_state (bool) – If True, also overwrite all other attributes of Hologram except for dtype and amp_ff.

log_state(level=10)

Log the current values of all tracked attributes.

Parameters:

level (int, optional) – Logging level. Defaults to logging.DEBUG.

measure(basis='ij')

Raises NotImplementedError; FeedbackHologram measures the camera.

optimize(method='GS', maxiter=20, verbose=True, callback=None, feedback=None, stat_groups=None, **kwargs)

Optimizes to solve the “phase problem”: approximating the nearfield phase that transforms a known nearfield source amplitude to a desired farfield target amplitude. Supported optimization methods include:

• Gerchberg-Saxton (GS) phase retrieval.

  • 'GS'

    An iterative algorithm for phase retrieval, accomplished by moving back and forth between the imaging and Fourier domains, with amplitude corrections applied to each. This is usually implemented using fast discrete Fourier transforms, potentially GPU-accelerated.

• Weighted Gerchberg-Saxton (WGS) phase retrieval algorithms of various flavors. Improves the uniformity of GS-computed focus arrays using weighting methods and techniques from literature. The method keywords are:

  • 'WGS-Leonardo'

    The original WGS algorithm. Weights the target amplitudes by the ratio of mean amplitude to computed amplitude, which amplifies weak spots while attenuating strong spots. Uses the following weighting function:

    \[\mathcal{W} = \mathcal{W}\left(\frac{\mathcal{T}}{\mathcal{F}}\right)^p\]

    where \(\mathcal{W}\), \(\mathcal{T}\), and \(\mathcal{F}\) are the weight amplitudes, target (goal) amplitudes, and feedback (measured) amplitudes, and \(p\) is the power passed as "feedback_exponent" in flags (see kwargs). The power \(p\) defaults to .8 if not passed. In general, smaller \(p\) will lead to slower yet more stable optimization.

  • 'WGS-Kim'

    Improves the convergence of WGS-Leonardo by fixing the farfield phase strictly after a desired number of net iterations specified by "fix_phase_iteration" or after exceeding a desired efficiency (fraction of farfield energy at the desired points) specified by "fix_phase_efficiency"

  • 'WGS-Nogrette'

    Weights target intensities by a tunable gain factor.

    \[\mathcal{W} = \mathcal{W}\Big/\left(1 - f\left(1 - \frac{\mathcal{F}/\mathcal{T}}{\langle\mathcal{F}/\mathcal{T}\rangle}\right)\right)\]

    where \(\langle\cdot\rangle\) is the mean and \(f\) is the gain factor passed as "feedback_factor" in flags (see kwargs). The factor \(f\) defaults to .1 if not passed.

    Note that while Nogrette et al. compare powers, this implementation compares amplitudes for speed. These are identical to first order.

  • 'WGS-Wu'

    Weights using an exponential function, which is less sensitive to near-zero values of \(\mathcal{F}\) or \(\mathcal{T}\).

    \[\mathcal{W} = \mathcal{W}\exp\left( p (\mathcal{T} - \mathcal{F}) \right)\]

    The speed of correction is controlled by \(p\), the power passed as "feedback_exponent".

  • 'WGS-tanh'

    Weights by hyperbolic tangent, commonly used as an activation function in machine learning.

    \[\mathcal{W} = \mathcal{W}\left[1 + f\text{tanh}\left( p (\mathcal{T} - \mathcal{F}) \right) \right]\]

    This weighting limits each update to a relative change of \(\pm f\), passed as "feedback_factor", which is useful to prevent large changes. The speed of correction is controlled by \(p\), the power passed as "feedback_exponent".

• Conjugate Gradient (CG) phase retrieval.

  • 'CG'

    (This feature is experimental.)

    Some holography—especially that with more complicated holographic objectives—can be better treated with gradient-based methods. In these cases, the phase is guided to an optimized state by following the back-propagated gradients (with respect to phase) of given objective loss which is passed as one of the flags to optimize(). Weighting different components of the objective leads to tradeoffs between those components: for instance a tradeoff between power guided into a given pattern and the uniformity of the realized pattern. slmsuite uses torch as a backend for gradient computation. Notably, memory is still owned and initialized by cupy, but gradients can be calculated by using torch-cupy interoperability.

    The objective loss is expected to be a torch.nn.Module and defaults to a complex variant of torch.nn.MSELoss() that skips undefined (nan) target pixels, such as an MRAF noise region, but normalizes the farfield over the whole plane. loss is called in the style of torch, using (as arguments) the computed farfield (with gradient tree intact) and the target values for the farfield. Internally, this looks like:

    result = loss(      # The user provides this nn.Module to .optimize()
        farfield,       # The farfield (with gradients), calculated from `phase` by slmsuite
        target          # The target, initialized by the user and processed by slmsuite
    )
    result.backward()   # Gradients are back-propagated to the input `phase`.
    

    For FeedbackHologram and subclasses, the gradients are computed computationally, but the computational values are then replaced with the experimental results. This allows optimization of the experimental results using the computational gradients (correct to first order) as a guide. Pixels the camera does not see are passed as undefined target pixels, so a null_region there is not enforced, and a target pixel the camera reads as zero passes no gradient: start from a computational solution. Currently, feedback is not supported for spot arrays with "experimental_spot" or "computational_spot" feedback (WGS probably works better for such spot array objectives anyway).

    Creating a custom objective is as simple as making a custom torch.nn.Module.forward() method. These methods can be as simple as a single expression or as complicated as a full neural network operating on the input parameters. However, remember to use torch methods because the arguments are of type torch.Tensor. Here’s an example of a custom torch.nn.Module.forward() which implements the Huber loss:

    # Define the loss as a class.
    class HuberLoss(nn.Module):
        def __init__(self, delta=1.0):
            super(HuberLoss, self).__init__()
            self.delta = delta
    
        def forward(self, farfield, target):
            residual = torch.abs(farfield - target)
            quadratic = torch.clamp(residual, max=self.delta)
            linear = residual - quadratic
            loss = 0.5 * quadratic ** 2 + self.delta * linear
    
            return torch.mean(loss)
    
    # Initialize the class. Remember that we can pass arguments (delta) here.
    loss = HuberLoss(delta=2.0)
    
    # Pass the loss to the hologram by one of two methods:
    hologram.optimize(..., loss=loss)       # 1. Pass as **kwarg.
    hologram.flags["loss"] = loss           # 2. Set directly.
    

    MRAF (next section), if desired, needs to be handled by the loss function. MRAF information is encoded in the target, with the noise region being nan.

• The option for Mixed Region Amplitude Freedom (MRAF) feedback. In standard iterative algorithms, the entire Fourier-domain unpatterned field is replaced with zeros. This is disadvantageous because a desired farfield pattern might not be especially compatible with a given nearfield amplitude, or otherwise. MRAF enables “noise regions” where some fraction of the given farfield is not replaced with zeros and instead is allowed to vary. In practice, MRAF is enabled by setting parts of the target to nan; these regions act as the noise regions. The "mraf_factor" flag in flags allows the user to partially attenuate the noise regions. A factor of 0 fully attenuates the noise region (normal WGS behavior). A factor of 1 does not attenuate the noise region at all (the default). Middle ground is recommended, but is application-dependent as a tradeoff between improving pattern fidelity and maintaining pattern efficiency. Setting the "mraf" flag to False disables MRAF even when the target contains nan. With MRAF enabled, a nonzero "zero_factor" flag \(f_0\) replaces the zeroing of zero-target pixels with an accumulated correction \(\mathcal{W}_0 \leftarrow \mathcal{W}_0 - f_0|E|E\), where \(E\) is the complex farfield there (default 0, zeroing).

  As examples, consider two cases where MRAF can be useful:

  • Sloping a top hat. Suppose we want very flat amplitude on a beam. Requesting a sharp edge to this beam can lead to fringing effects at the boundary which mitigate flatness both inside and outside the beam. If instead a noise region is defined in a band surrounding the beam, the noise region will be filled with whatever slope best enables the desired flat beam.

  • Mitigating diffractive orders. Without MRAF, spot patterns with high crystallinity often have “ghost” diffractive orders which continue the pattern past the edges of requested spots. Even though these orders are attenuated during each phase retrieval iteration, they remain part of the best solution for the recovered phase. With MRAF, a noise region can help solve for retrieved phase which does not generate these undesired orders.

Caution

Requesting stat_groups will slow the speed of optimization due to the overhead of processing and saving statistics, especially in the case of GPU-accelerated optimization where significant time cost is incurred by moving these statistics to the CPU. This is especially apparent in the case of fully-computational holography, where this effect can slow what is otherwise a fully-GPU-contained loop by an order of magnitude.

Tip

This function uses a parameter naming convention borrowed from scipy.optimize.minimize() and other functions in scipy.optimize. The parameters method, maxiter, and callback have the same functionality as the equivalently-named parameters in scipy.optimize.minimize().

Parameters:

• method (str) – Optimization method to use. See the list of optimization methods above.

• maxiter (int) – Number of iterations to optimize before terminating.

• verbose (bool OR int) – Whether to display tqdm progress bars. These bars are also not displayed for maxiter <= 1.

• callback (callable OR None) – Same functionality as the equivalently-named parameter in scipy.optimize.minimize(). callback must accept a Hologram or Hologram subclass as the single argument. If callback returns True, then the optimization exits. Ignored if None. When applying "external_spot" feedback, this callback is the place to update the "external_spot_amp" attribute.

• feedback (str OR None) –

  Type of feedback to use during optimization, for instance when weighting in "WGS". For direct instances of Hologram, this can only be "computational" feedback. Subclasses support more types of feedback. Supported feedback options include the following:

  • "computational" Uses the projected farfield pattern (transform of the complex nearfield) as feedback.

  • "experimental" Uses a camera contained in a passed cameraslm as feedback. Specific to subclasses of FeedbackHologram.

  • "computational_spot" Takes the computational result (the projected farfield pattern) and integrates regions around the expected positions of spots in an optical focus array. More stable than "computational" for spots. Specific to subclasses of SpotHologram.

  • "experimental_spot" Takes the experimental result (the image from a camera) and integrates regions around the expected positions of spots in an optical focus array. More stable than "experimental" for spots. Specific to subclasses of SpotHologram.

  • "external_spot" Uses some external user-provided metric for spot feedback. See SpotHologram.external_spot_amp. Usually applied during callback. Specific to subclasses of SpotHologram.

• stat_groups (list of str OR None) – Strings representing types of feedback (data gathering) upon which statistics should be derived. These strings correspond to valid types of feedback (see above). For instance, if "experimental" is passed as a stat group, statistics on the pixels in the experimental feedback image will automatically be computed and stored for each iteration of optimization. However, this comes with overhead (see above warning).

• **kwargs (dict, optional) – Various weight keywords and values to pass depending on the weight method. These are passed into flags. See options documented in the constructor. Flags read by reset_phase(), such as quadratic_phase, are stored too, but have no effect until reset_phase() is next called. The exception is name, which labels the tqdm progress bar and is not stored.

optimize_cg(iterations, callback)

Gradient-based iterative phase retrieval via torch autograd and a torch.optim optimizer (the "CG" method; see optimize()).

(This feature is experimental.)

Solves the “phase problem”: approximates the nearfield phase that transforms a known nearfield source amplitude to a desired farfield target amplitude.

Caution

This function should be called through optimize() and not called directly. It is left as a public function exposed in documentation to clarify how the internals of optimize() work.

Caution

This method does not currently support "experimental_spot" or "external_spot" feedback.

Parameters:

• iterations (iterable) – Number of loop iterations to run. Is an iterable to pass a tqdm iterable.

• callback (callable OR None) – See optimize().

optimize_gs(iterations, callback)

GPU-accelerated Gerchberg-Saxton (GS) iterative phase retrieval.

Solves the “phase problem”: approximates the nearfield phase that transforms a known nearfield source amplitude to a desired farfield target amplitude.

Caution

This function should be called through optimize() and not called directly. It is left as a public function exposed in documentation to clarify how the internals of optimize() work.

Parameters:

• iterations (iterable) – Number of loop iterations to run. Is an iterable to pass a tqdm iterable.

• callback (callable OR None) – See optimize().

pickle(attributes=True, metadata=True)

Returns a dictionary containing selected attributes of this class.

Parameters:

• attributes (bool OR list of str) – If False, pickles only baseline attributes, usually single floats. If True, also pickles ‘heavy’ attributes such as large images and calibrations. If list of str, pickles the keys in the given list. By default, the chosen attributes should be things that can be written to .h5 files: scalars and lists of scalars.

• metadata (bool) – If True, package the dictionary as the "__meta__" value of a superdictionary which also contains: "__version__", the current slmsuite version, "__time__", the time formatted as a date string, and "__timestamp__", the time formatted as a floating point timestamp. This information is used as standard metadata for calibrations and saving.

plot_farfield(*args, **kwargs)

Plots an overview (left) and zoom (right) view of source.

Parameters:

• source (array_like OR None) – Should have shape equal to shape. If None, defaults to amp_ff.

• title (str) – Title of the plots. If "phase" is a substring of title, then the source is treated as a phase.

• limits (((float, float), (float, float)) OR None) – \(x\) and \(y\) limits for the zoom plot in "knm" space. If None, limits are autocomputed as the smallest bounds that show all non-zero values (plus limit_padding). Note that autocomputing on target will perform well, as zero values are set to actually be zero. However, doing so on computational or experimental outputs (e.g. amp_ff) will likely perform poorly, as values in the field deviate slightly from zero and artificially expand the limits.

• units (str) – Far-field units for plots (see convert_vector() for options). If units requiring a SLM are desired, the attribute cameraslm must be filled.

• limit_padding (float) – Fraction of the width and height to expand the limits of the zoom plot by, only if the passed limits is None (autocompute).

• figsize (tuple) – Size of the plot.

• cbar (bool) – Whether to add colorbars to the plots. Defaults to False.

• axs (matplotlib.axes.Axes OR None) – If provided, uses these axes instead of creating new ones.

Returns:

Used limits, which may be autocomputed. Autocomputed limits are returned as integers.

Return type:

((float, float), (float, float))

plot_nearfield(source=None, title='', padded=False, figsize=(8, 4), cbar=False)

Plots the amplitude (left) and phase (right) of the nearfield (plane of the SLM). The amplitude is assumed (whether uniform, assumed, or measured) while the phase is the result of optimization.

Parameters:

• source (array_like OR None) – Source to plot. If None, defaults to the current amplitude and phase.

• title (str) – Title of the plots.

• padded (bool) – If True, shows the full computational nearfield of shape shape. Otherwise, shows the region at the center of the computational space of size slm_shape corresponding to the unpadded SLM.

• figsize (tuple) – Size of the plot.

• cbar (bool) – Whether to add colorbars to the plots. Defaults to False.

plot_stats(*args, **kwargs)

Plots the statistics contained in the given dictionary.

Parameters:

• stats_dict (dict OR None) – Stats to plot in dictionary form. If None, defaults to stats.

• stat_groups (list of str OR None) – Which statistics groups to plot. If empty or None is provided, defaults to all groups present in stats.

• ylim ((int, int) OR None) – Allows the user to pass in desired y limits. If None, the default y limits are used.

• show (bool) – Whether or not to immediately show the plot. Defaults to False.

remove_vortices(plot=False)

Removes each child’s farfield phase vortices; see Hologram.remove_vortices().

reset(reset_phase=True, reset_flags=False, reset_weights=True)

Resets the hologram to an initial state. Does not restore the preconditioned phase that may have been passed to the constructor (as this information is lost upon optimization). Instead, phase is randomized if reset_phase=True. Also uses the current target rather than the target that may have been passed to the constructor (e.g. includes current refine_offset() changes, etc).

Parameters:

• reset_phase (bool) – Whether to additionally call reset_phase().

• reset_flags (bool) – Whether to erase the information (including passed kwargs) stored in flags.

• reset_weights (bool) – Whether to additionally call reset_weights(). Pass False when the caller has already re-derived the weights from the current target.

reset_phase(*args, **kwargs)

Randomize or set the meta phase.

reset_weights()

Resets the hologram weights to the target defaults.

save(path='.', name=None, **kwargs)

Saves the dictionary returned from pickle() to a file like "path/name_id.h5".

Parameters:

• path (str) – Path to directory to save in. Default is current directory.

• name (str OR None) – Name of the save file. If None, will use the object’s name + '-pickle'.

• **kwargs – Passed to pickle() to customize how and what data is saved.

Returns:

The file path that the pickled data was saved to.

Return type:

str

save_stats(file_path, include_state=True)

Uses save_h5() to export the statistics hierarchy to a given h5 file.

Tip

Enabling the "raw_stats" flag will save feedback data from each iteration instead of only derived statistics. Consider enabling this to save more detailed information.

Parameters:

• file_path (str) – Full path to the file to write the data to.

• include_state (bool) – If True, also includes all other attributes of Hologram except for dtype (cannot pickle) and amp_ff (can regenerate). These attributes are converted to numpy if necessary. Note that the intent is not to produce a runnable Hologram by default (as this would require pickling hardware interfaces), but rather to provide extra information for debugging.

static set_mempool_limit(device=0, size=None, fraction=None)

Helper function to set the cupy memory pool size.

Parameters:

• device (int) – Which GPU to set the limit on. Passed to cupy.cuda.Device().

• size (int) – Desired number of bytes in the pool. Passed to cupy.cuda.MemoryPool.set_limit().

• fraction (float) – Fraction of available memory to use. Passed to cupy.cuda.MemoryPool.set_limit().

set_propagation_kernels(kernels)

Set every child’s propagation_kernel at once from a stacked array.

The batched fast path works from a single (S, slm_h, slm_w) tensor internally, so assigning the children one at a time forces it to restack them – an allocation, S scatter-assignments, and a dtype cast – every time. Handing the stack over directly skips all of that. Each child’s propagation_kernel is rebound to the corresponding slice, so the per-child and non-batched code paths keep reading what they did before.

Parameters:

kernels (array_like) – Phase kernels of shape (S, slm_h, slm_w), or anything broadcastable to it (e.g. a single (slm_h, slm_w) array applied to every plane).

set_target(*args, **kwargs)

Change the target to something new. This method handles cleaning and normalization.

Parameters:

• new_target (array_like OR None) – New target to optimize towards. Should be of shape shape. If None, target is zeroed (used internally, but probably should not be used by a user).

• reset_weights (bool) – Whether to update the weights to this new target.

set_weights(new_weights)

Sets the weights to a new value. Handles moving data onto the GPU if applicable.