Configuration¶

This page describes how to configure the model that RAITAP will assess.

Options¶

Name	Allowed	Default	Description
`source`	`string, null`	`null`	Path to a local model file, or a built-in model name such as `"resnet50"`. Refer to Using your own model or built-in models for more details.
`arch`	`string, null`	`null`	torchvision architecture name (e.g. `"resnet18"`) used to instantiate the model when `source` points at a state-dict (a `.pt` / `.pth` file produced by `torch.save(model.state_dict(), path)`). Required alongside `num_classes` for state-dict loading; ignored for full pickled modules and TorchScript archives.
`num_classes`	`int, null`	`null`	Output classes used to instantiate the architecture before `load_state_dict`. Required together with `arch` for state-dict loading.
`pretrained`	`bool`	`false`	If `true`, construct the architecture with ImageNet pretrained weights before loading the state-dict (`weights="DEFAULT"`). Usually `false` since the state-dict already supplies the weights.

YAML example¶

# Option A: single model file.
# `source` is the only key. The format is inferred from the extension
# (.pt/.pth, .onnx, .ubj). Per-format details and caveats (state-dict vs
# TorchScript vs pickled, the unsafe-pickle consent, the `--extra xgboost`
# requirement for .ubj) are on the "Using your own model" page (linked
# from the `source` option above).
model:
  source: "path/to/model.<ext>"

# Option B: state-dict file.
# A state-dict carries no architecture, so add `arch` + `num_classes` to
# rebuild the model before loading the weights.
model:
  source: "weights.pth"
  arch: "resnet18"
  num_classes: 2

# Option C: built-in torchvision model — `source` is the model name, not a
# path. Loaded with torchvision's `weights="DEFAULT"` (latest pretrained
# weights, not configurable). For demos / quick testing; load your own
# weights via a file path (Options A/B).
model:
  source: "resnet50"

CLI override example¶

uv run raitap model.source=resnet50

pip

raitap model.source=resnet50