Function Reference

This section explains the package-owned public functions and classes in gpi_pack. The entries are grouped by module so that you can find the lower-level function used by each workflow. Underscore-prefixed modules and names are implementation details and are not part of the user-facing API.

Text Generation

  • get_instruction: creates the built-in system instruction for creating or repeating text, or returns a custom instruction.

  • generate_text: generates text with non-sampling decoding and saves a pooled LLM hidden state for every prompt.

  • save_generated_texts: saves generated texts and their prompts in a pickle file.

  • extract_and_save_hidden_states: runs the complete text-generation, hidden-state extraction, and output-saving workflow.

Image Generation

  • pad_to_multiple_of_8: pads a PIL image so that both spatial dimensions are divisible by eight.

  • StableDiffusionImg2ImgExtractor: loads the Stable Diffusion components and provides preprocessing, encoding, transformation, decoding, and saving methods.

  • extract_images: processes one or more image-prompt pairs and saves the final diffusion latents and optional generated images.

Video Processing

  • CosmosVideoExtractor: loads the Cosmos VAE and provides in-memory encoding, reconstruction, representation extraction, and file-processing methods.

  • VideoExtractionResult: stores the tensors and preprocessing metadata returned for an in-memory video clip.

  • VideoSegmentOutput: stores the output paths and segment identity returned by file processing.

  • extract_videos: discovers, segments, and processes video files and saves one Cosmos representation payload for every selected segment.

Static Inference

  • TarNetBase: implements the treatment-conditioned neural outcome model and learned deconfounder.

  • TarNet: trains, validates, and predicts with TarNetBase.

  • SpectralNormClassifier: estimates class probabilities with a spectrally normalized neural network.

  • dml_score: calculates the doubly robust influence score for an average treatment effect.

  • estimate_psi_split: cross-fits the propensity model within a sample and returns influence scores and propensity predictions.

  • estimate_k_ate: estimates an average treatment effect and its standard error with k-fold cross-fitting.

  • TarNetHyperparameterTuner: tunes the static TarNet outcome model with Optuna and can refit the best configuration.

  • load_hiddens: loads saved .pt representations in a requested order and returns a NumPy array.

Dynamic Inference

  • TextMLPEncoder: maps each vector or text representation to a fixed-width segment embedding.

  • Video3DEncoder: maps each video latent volume to a fixed-width segment embedding with 3D convolutions.

  • DynamicTarNetBase: implements the masked sequential representation and outcome networks used by Dynamic GPI.

  • mse_loss: calculates the scalar mean squared error for two compatible tensors.

  • DynamicTarNet: trains, validates, and predicts one scalar outcome with the dynamic sequence model in vector-only or multimodal mode.

  • DynamicGPIHyperparameterTuner: tunes and refits the scalar-outcome DynamicTarNet with Optuna. DynamicTarNetHyperparameterTuner is an alias of the same class.

  • estimate_k_ipsi: estimates cross-fitted longitudinal incremental-intervention curves and uncertainty for scalar or repeated outcomes.