Convert tensor to PIL image.

Forward pass through MLP.

Log a message to all outputs. Args: message: Message to log

Initialize custom MLP.

Run inference with a specific color. Args: save_filename (str): Base filename for saving outputs output_dir (Path): D

Create directory if not exists. Args: dir_path: Directory path Returns: Created directory path

Execute the main training loop. This method runs the main training loop, handling: - Batch processing and forward passes

Get background color based on settings.

Apply nested dropout to embedding.

Returns list of utf-8 byte and a corresponding list of unicode strings. The reversible bpe codes work on unicode strings. This means you

Calculate the loss for regular training. Args: pred (Tensor): Predicted tensor toggle_color_value (Optional[str]): Co

Calculate the loss for pretraining. Args: pred (Tensor): Predicted tensor gt (Tensor): Ground truth tensor

Save a model checkpoint. Args: is_best (bool): Whether this is the best model so far. Defaults to False. is_final (bo

Generate text embeddings.

Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length strings).

Update current training step.

Add aspect ratio encoding.

Add aspect ratio rectangles to shapes and groups.

Add color encoding.

Create background rectangle shape.

Create color toggle embedding.

Create embedding layers.

Create network layers.

Create main MLPs for points and colors.

Create shapes and shape groups for rendering.

Create stroke color with opacity.

Calculate learning rate using exponential ramp and cosine decay. Args: step (int): Current step Returns: flo

Generate network outputs.

Get aspect ratio value for current step.

Get base positional encoding.

Get color encoding based on method.

Generate suffix for color-specific output files. Args: color_value (Optional[Union[str, torch.Tensor]]): Color value

Get predictions from MLP.

Get number of shapes to process.

Get stroke color for rendering.

Calculate truncation indices for shape introduction.

Handle nested dropout and return truncation index.

Run single inference without interpolation.

Initialize and save initial sketch state.

Initialize network dimensions.

Initialize NeRF MLP model.

Initialize model parameters.

Initialize and configure the PainterNerf renderer.

Initialize color and aspect ratio toggle values.

Interpolate between two colors during inference.

Calculate learning rate using linear ramp and cosine decay. Args: step (int): Current step Returns: float: L

Load model weights from checkpoint.

Log metrics to wandb. Args: loss_dict: Dictionary of loss components

Plot attention maps if available.

Process colors tensor.

Process points tensor and apply aspect ratio.

Process a color during the pretraining phase. Args: toggle_color_value: Color value to process Returns:

Process a single color value during regular training. Args: toggle_color_value: Color value to process Returns:

Process stroke points tensor.

Process width and opacity tensors.

Initialize training parameters and state. Sets up: - Initial learning rates - Optimizer parameters - Trainin

Calculate learning rate using sinusoidal ramp and exponential decay. Args: step (int): Current step Returns:

Store configuration parameters.

Validate and process sub-layer sizes.

Returns the names of available CLIP models

Calculate aspect ratio from string.

Verify CUDA is available for training.

Apply gradient clipping to color parameters.

Apply gradient clipping to point parameters.

Configure loguru logger settings.

Convert applicable model parameters to fp16

Create experiment directory structure.

Create transform pipeline for image preprocessing.