Conventions

The rules that let datasets, models, losses and metrics be combined freely.

Notation

Symbol	Meaning
C	Coordinate axes, the in_features of an INR
D	Output channels, the out_features of an INR
N	Points in a flat batch; (N, C) is such a batch and (..., C) allows any leading dimensions
K_c, K	Basis functions on axis c, or on every axis
R	CP rank
H, W	Image height and width
channels, depth	An image's channels and a volume's first grid axis, spelled out because C and D mean the above

Coordinates

Coordinates live in \([-1, 1]\) on every axis, both endpoints included. inro.create_coordinates((H, W)) returns an (H, W, 2) grid whose column \(c\) spans axis \(c\) (indexing="ij"), and the datasets build their inputs with it.

FUTON's bases and the hash and feature grids are defined on this domain, so do not pre-normalize to \([0, 1]\). Models accept any leading shape; the taps of FUTON's local bases need a flat (N, C) batch, which FUTON.forward produces internally.

Values

Targets are in the model's output range, and the reference signal keeps its native form.

Dataset	input	target	_original
ImageCoordinateDataset	(H, W, 2)	(H, W, channels) float in \([-1, 1]\)	(channels, H, W) uint8
OccupancyCoordinateDataset	(depth, H, W, 3)	(depth, H, W, 1) in \(\{-1, 1\}\)	(1, depth, H, W) float in \(\{0, 1\}\)
nerf.BlenderDataset	rays (H, W, 3) origins and directions	(H, W, 4) RGBA in \([0, 1]\)

• Subsampling (subsample, max_samples) flattens input and target to (N, ...) with shared random indices.
• preprocess maps native values to targets: uint8 to \([-1, 1]\), occupancy \(\{0, 1\}\) to \(\{-1, 1\}\).
• postprocess maps model output back: clipping to \([-1, 1]\) before the uint8 cast, thresholding occupancy at zero.

Because targets are in \([-1, 1]\), the benchmark models end in torch.tanh, and the default training loss reports PSNR with a peak-to-peak range of 2.

Metrics

In inro.train and inro.evaluate, metrics are called as fn(reconstruction, original) after postprocess, so they see uint8 images and {0, 1} volumes.

Metric	Input	Notes
inro.psnr, inro.ssim, inro.ms_ssim	two (channels, H, W) images	data_range from the target: 255 for uint8, 1 for floating point
inro.lpips	two (3, H, W) images	VGG by default, net="alex" for AlexNet; its weights are downloaded once
inro.iou	two volumes	Thresholds both at zero, so it works on the {0, 1} original and on raw model output alike

Widths, depths and activations

hidden_layers=h means h hidden layers and h + 1 weight layers. output_activation is applied after the last layer of every model; inside inro.nerf.RadianceField the density network must have none, because the field applies trunc_exp to the density and a sigmoid to the color.

Checkpoints

• inro.train saves the model's state dict as checkpoint.pt in log_dir and returns it as model_dict.
• inro.nerf.train saves the field's state dict, including the aabb buffer.
• MLP.layers holds only weight layers (layers.0 to layers.h). FUTON keeps its head under head.* (head.layers.* for an MLP), and the grid models their decoder under decoder.layers.*.

Seeds and devices

• seed in inro.train and inro.nerf.train seeds PyTorch's global RNG before training, after the model has been built. Construct the model after torch.manual_seed to also fix its initialization, as the scripts do.
• A dataset's seed reseeds Python's RNG only; subsampling draws from PyTorch's RNG.
• device=None selects CUDA when available. The training loop moves the model and, through their to method, the datasets to that device, and _original stays on the CPU.

Experiment files

A benchmark run lives in logs/<task>/<signal>/<model>/ with log.txt, log.json, checkpoint.pt, the final reconstruction (or two rendered test views for NeRF), and results.json: the training configuration, the model's setup from the YAML, num_params, train_time, the final metrics, and the history. Failed runs leave error.txt. The reproduction page describes the scripts that write and read them.