DeepSDFStruct.deep_sdf.training_latent_field#

Functions

append_parameter_magnitudes(param_mag_log, model)

build_template_spline(latent_dim, tiling, bounds)

Builds a splinepy.BSpline in [mins,maxs]^d with a control point grid that matches tiling.

clip_logs(loss_log, lr_log, timing_log, ...)

export_training_latent_fields_to_stl(...[, ...])

Export one STL per training scene using the trained spline latent fields (control points), and the trained decoder weights at checkpoint.

get_mean_spline_param_magnitude(latent_fields)

get_spec_with_default(specs, key, default)

load_latent_fields(experiment_directory, ...)

load_logs(experiment_directory)

load_optimizer(experiment_directory, ...)

make_latent_fields(num_scenes, latent_dim, ...)

Create one SplineParametrization (with learnable control points) per scene.

save_latent_fields(experiment_directory, ...)

save_logs(experiment_directory, loss_log, ...)

save_model(experiment_directory, filename, ...)

save_optimizer(experiment_directory, ...)

train(experiment_directory[, data_source, ...])

Train decoder + spline latent fields.

Classes

ClampedL1Loss([clamp_val])

ConstantLearningRateSchedule(value)

CosineAnnealingLRSchedule(initial, final, ...)

LearningRateSchedule()

StepLearningRateSchedule(initial, interval, ...)

WarmupLearningRateSchedule(initial, ...)

class DeepSDFStruct.deep_sdf.training_latent_field.ClampedL1Loss(clamp_val=0.1)#

Bases: torch.nn.modules.module.Module

forward(input, target)#

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class DeepSDFStruct.deep_sdf.training_latent_field.ConstantLearningRateSchedule(value)#

Bases: DeepSDFStruct.deep_sdf.training_latent_field.LearningRateSchedule

get_learning_rate(epoch)#
class DeepSDFStruct.deep_sdf.training_latent_field.CosineAnnealingLRSchedule(initial, final, total_epochs)#

Bases: DeepSDFStruct.deep_sdf.training_latent_field.LearningRateSchedule

get_learning_rate(epoch)#
class DeepSDFStruct.deep_sdf.training_latent_field.LearningRateSchedule#

Bases: object

get_learning_rate(epoch)#
class DeepSDFStruct.deep_sdf.training_latent_field.StepLearningRateSchedule(initial, interval, factor)#

Bases: DeepSDFStruct.deep_sdf.training_latent_field.LearningRateSchedule

get_learning_rate(epoch)#
class DeepSDFStruct.deep_sdf.training_latent_field.WarmupLearningRateSchedule(initial, warmed_up, length)#

Bases: DeepSDFStruct.deep_sdf.training_latent_field.LearningRateSchedule

get_learning_rate(epoch)#
DeepSDFStruct.deep_sdf.training_latent_field.append_parameter_magnitudes(param_mag_log, model)#
DeepSDFStruct.deep_sdf.training_latent_field.build_template_spline(latent_dim, tiling, bounds, degrees=None)#

Builds a splinepy.BSpline in [mins,maxs]^d with a control point grid that matches tiling.

DeepSDFStruct.deep_sdf.training_latent_field.clip_logs(loss_log, lr_log, timing_log, lat_mag_log, param_mag_log, epoch)#
DeepSDFStruct.deep_sdf.training_latent_field.export_training_latent_fields_to_stl(experiment_directory, checkpoint='latest.pth', out_dir=None, N_base=12, device=None, overwrite=True, max_scenes=3)#

Export one STL per training scene using the trained spline latent fields (control points), and the trained decoder weights at checkpoint.

Uses the same geometry pipeline as test_reconstruction.py:

SDFfromDeepSDF -> LatticeSDFStruct -> create_3D_mesh -> export_surface_mesh

Parameters:

  • experiment_directory (str)

  • checkpoint (str)

  • out_dir (str | None)

  • N_base (int)

  • device (str | None)

  • overwrite (bool)

  • max_scenes (int)

DeepSDFStruct.deep_sdf.training_latent_field.get_mean_spline_param_magnitude(latent_fields)#
DeepSDFStruct.deep_sdf.training_latent_field.get_spec_with_default(specs, key, default)#
DeepSDFStruct.deep_sdf.training_latent_field.load_latent_fields(experiment_directory, filename, latent_fields, device)#
DeepSDFStruct.deep_sdf.training_latent_field.load_logs(experiment_directory)#
DeepSDFStruct.deep_sdf.training_latent_field.load_optimizer(experiment_directory, filename, optimizer)#
DeepSDFStruct.deep_sdf.training_latent_field.make_latent_fields(num_scenes, latent_dim, tiling, bounds, device, init_std=0.01, degrees=None)#

Create one SplineParametrization (with learnable control points) per scene. All splines share identical topology (degrees/knot vectors/control point count).

DeepSDFStruct.deep_sdf.training_latent_field.save_latent_fields(experiment_directory, filename, latent_fields, epoch, num_scenes, latent_dim, device)#
DeepSDFStruct.deep_sdf.training_latent_field.save_logs(experiment_directory, loss_log, lr_log, timing_log, lat_mag_log, param_mag_log, epoch)#
DeepSDFStruct.deep_sdf.training_latent_field.save_model(experiment_directory, filename, decoder, epoch)#
DeepSDFStruct.deep_sdf.training_latent_field.save_optimizer(experiment_directory, filename, optimizer, epoch)#
DeepSDFStruct.deep_sdf.training_latent_field.train(experiment_directory, data_source=None, continue_from=None, device=None, use_mlflow=False, mlflow_run_name=None, mlflow_tags=None, mlflow_log_every_n_batches=50)#

Train decoder + spline latent fields.

Parameters:

  • use_mlflow (bool)

  • mlflow_run_name (str | None)

  • mlflow_tags (dict | None)

  • mlflow_log_every_n_batches (int)