Skip to content

Latest commit

 

History

History
219 lines (158 loc) · 10.1 KB

README.md

File metadata and controls

219 lines (158 loc) · 10.1 KB
Raw

AIDesign-GAN Model Folder

A folder that holds the subfolders and files that an AIDesign-GAN model needs.

Documentation Files

Texts.

README.md

This file itself.

Configuration Files

Texts.

coords_config.json

Coordinators configuration.

Configuration items. Type dict[str, typing.Union[dict, list, str, bool, int, float, None]].

Configuration item descriptions are listed below.

  • training. Training coordinator configuration. Type dict.
    • mode. Training mode. Type str. Values "new", "resume".
    • algorithm. Training algorithm. Type str. Values "alt_sgd_algo", "pred_alt_sgd_algo", "fair_pred_alt_sgd_algo".
    • manual_seed. Manual random seed. Type typing.Union[None, int].
    • gpu_count. Type int. Range [0, ).
    • iter_count. Iteration count. Each iteration contains multiple epochs. Type int. Range [0, ). Compatibility alias iteration_count. Precedence iter_count > iteration_count.
    • epochs_per_iter. Epochs per iteration. Each epoch contains one complete pass of the training and validation datasets to use. Type int. Range [0, ). Compatibility alias epochs_per_iteration. Precedence epochs_per_iter > epochs_per_iteration.
    • max_rollbacks. Maximum rollbacks. Type int. Range [0, ).
    • max_early_stops. Maximum early stops. Type int. Range [0, ).
    • datasets. Datasets configuration. Type dict.
      • loader_worker_count. Type int. Range [0, ).
      • percents_to_use. Percentage of dataset to use for training. Type float. Range [0, 100].
      • images_per_batch. Type int. Range [1, ).
      • image_resolution. The resolution in pixels to resize the images used for training. Type int. Range [1, ).
      • image_channel_count. Type int. Range [1, ).
      • training_set_weight. Type float. Range [0, ].
      • validation_set_weight. Type float. Range [0, ].
    • labels. Labels configuration. Type dict.
      • real. The number used to label each real image from the real world. Type float. Range [0, 1].
      • fake. The number used to label each fake image from the generator. Type float. Range [0, 1].
    • noise_models. Model noising configuration. Type dict.
      • before_each_iter. Whether to inject noises into the model parameters before each iteration. Type bool.
      • before_each_epoch. Whether to inject noises into the model parameters before each epoch. Type bool.
      • save_noised_images. Whether to save the generated images after model noising. Type bool.
    • epoch_collapses. Epoch training collapse detection configurations. Type dict.
      • max_loss: Maximum loss allowed for a training batch to be marked as not collapsed. If the loss exceeds max_loss, the training batch will be marked as collapsed. Type float. Range [0, 100].
      • percents_of_batches. Maximum percentage of collapsed training batches allowed to mark an epoch as not collapsed. Type float. Range [0, 100].
    • retrirals. Epoch training exception retrials configurations. Type dict.
      • max_count. Maximum retrial count. Type int. Range [0, ].
      • delay_seconds. Delay in seconds, before starting a retrial. Type float. Range [0, ].
  • generation. Generation coordinator configuration. Type dict.
    • manual_seed. Manual random seed. Type typing.Union[None, int].
    • gpu_count. Type int. Range [0, ).
    • image_count. Type int. Range [0, ).
    • images_per_batch. Type int. Range [1, ).
    • grid_mode. Grid generation mode configuration. Type dict.
      • enabled. Whether to enable the grid mode. Type bool.
      • images_per_grid. Type int. Range [1, ).
      • padding. Grid padding. Type int. Range [0, ).
  • exportation. Exportation coordinator configuration. Type dict.
    • manual_seed. Manual random seed. Type typing.Union[None, int].
    • gpu_count. Type int. Range [0, ).
    • images_per_batch. Type int. Range [1, ).
    • preview_grids. Preview grids configuration. Type dict.
      • images_per_grid. Type int. Range [1, ).
      • padding. Grid padding. Type int. Range [0, ).

discriminator_struct.py

Discriminator structure.

Python code fragment. Uses information from the loaded modeler configuration at self.config to setup the targeted model structure at self.model.

self.config. Loaded modeler configuration. Type dict.

self.model. Targeted model structure to setup. type torch.nn.Module.

Warning: Code in this file can possibly make gan train execute malicious code from this file itself and somewhere else. Please be careful.

format_config.json

Format configuration.

Automatically configured. Not supposed to be edited manually. Serves as a model folder format reference.

Configuration items. Type dict[str, typing.Union[dict, list, str, bool, int, float, None]].

Configuration item descriptions are listed below.

  • aidesign_gan_version. Type str.
  • aidesign_gan_repo_tag. Type str.

generator_struct.py

Generator structure.

Python code fragment. Uses information from the loaded modeler configuration at self.config to setup the targeted model structure at self.model.

self.config. Loaded modeler configuration. Type dict.

self.model. Targeted model structure to setup. type torch.nn.Module.

Warning: Code in this file can possibly make gan train, gan generate, and gan export ... execute malicious code from this file itself and somewhere else. Please be careful.

modelers_config.json

Modelers configuration.

Configuration items. Type dict[str, typing.Union[dict, list, str, bool, int, float, None]].

Configuration item descriptions are listed below.

  • discriminator. Discriminator modeler only configurations. Type dict.
    • image_resolution. Input image resolution in pixels. Type int. Range [1, ).
    • image_channel_count. Input image channel count. Type int. Range [1, ).
    • label_resolution. Output label resolution in pixels. Type int. Range [1, ).
    • label_channel_count. Output label channel count. Type int. Range [1, ).
    • feature_map_count. Layer 0 (first layer) output feature map count. Type int. Range [1, ). Compatibility alias feature_map_size. Precedence feature_map_count > feature_map_size.
  • generator. Generator modeler only configurations. Type dict.
    • noise_resolution. Input noise resolution in pixels. Type int. Range [1, ).
    • noise_channel_count. Input noise channel count. Type int. Range [1, ).
    • image_resolution. Output image resolution in pixels. Type int. Range [1, ).
    • image_channel_count. Output image channel count. Type int. Range [1, ).
    • feature_map_count. Layer -1 (last layer) input feature map count. Type int. Range [1, ). Compatibility alias feature_map_size. Precedence feature_map_count > feature_map_size.
  • discriminator or generator. Discriminator and generator modelers common configurations. Type dict.
    • struct_name. Structure name. Type str.
    • state_name. Model state name. Type str.
    • optim_name. Optimizer state name. Type str.
    • adam_optimizer. Adam optimizer configuration. Type dict.
      • learning_rate. Type float. Range [0, ).
      • beta1. Momentum beta-1. Type float. Range [0, 1].
      • beta2. Momentum beta-2. Type float. Range [0, 1].
      • pred_factor. Prediction factor. Type float.
    • params_init. Parameters initialization configuration. Type dict.
      • conv. Convolution layers configuration. Type dict.
        • weight_mean. Type float.
        • weight_std. Weight standard deviation. Type float. Range [0, ).
      • batch_norm. Batch normalization layers configuration. Type dict.
        • weight_mean. Type float.
        • weight_std. Weight standard deviation. Type float. Range [0, ).
        • bias_mean. Type float.
        • bias_std. Bias standard deviation. Type float. Range [0, ).
    • params_noising. Parameters noising configuration. Type dict.
      • conv. Convolution layers configuration. Type dict.
        • delta_weight_mean. Weight incrementation mean. Type float.
        • delta_weight_std. Weight incrementation standard deviation. Type float. Range [0, ).
      • batch_norm. Batch normalization layers configuration. Type dict.
        • delta_weight_mean. Weight incrementation mean. Type float.
        • delta_weight_std. Weight incrementation standard deviation. Type float. Range [0, ).
        • delta_bias_mean. Bias incrementation mean. Type float.
        • delta_bias_std. Bias incrementation standard deviation. Type float. Range [0, ).
    • fairness. Fair loss factor configuration. Type dict.
      • dx_factor. D(X) factor. Type float.
      • dgz_factor. D(G(Z)) factor. Type float.
      • cluster_dx_factor. Cluster D(X) factor. Type float.
      • cluster_dgz_factor. Cluster D(G(Z)) factor. Type float.
      • cluster_dx_overact_slope. Cluster D(X) overact slope. Type float.
      • cluster_dgz_overact_slope. Cluster D(G(Z)) overact slope. Type float.

Result Subfolders And Files

Folders, texts, and images.

Generation-Results

Note: Not present and ready until an AIDesign-GAN generation session completes.

Generation results.

Training-Results

Note: Not present and ready until an AIDesign-GAN training session completes.

Training results.

log.txt

Note: Not present and ready until an AIDesign-GAN training, generation, or exportation session completes.

Log.

State Storage Files

Binaries.

discriminator_optim.pt

Note: Not present and ready until an AIDesign-GAN training session completes.

Discriminator optimizer.

discriminator_state.pt

Note: Not present and ready until an AIDesign-GAN training session completes.

Discriminator state.

generator_optim.pt

Note: Not present and ready until an AIDesign-GAN training session completes.

Generator optimizer.

generator_state.pt

Note: Not present and ready until an AIDesign-GAN training session completes.

Generator state.