ShiftExperiment CLI Builder

Weights Mode

Random weights is the default and emits no subcommand. Choose ImageNet or Custom to emit a subcommand.

mode

No-subcommand behavior: random weights selected when mode is “Random”.

Dataset Configuration

--source_dir (required)

Root directory for the source distribution.

--target_dir (required)

Root directory for the target distribution.

--source_list_path

List file for source images (defaults to CULane train.txt).

--target_list_path (required)

List file for target images.

Sampling & Test Controls

Controls sample size, run count, and statistical-test settings.

--sample_size

Images per run.

--num_runs

Number of times calibration is ran. Also, number of times the test is ran.

--block_idx

Index for block sampling of dataset.

--batch_size

Dataloader batch size.

--image_size

Input image size (pixels).

--latent_dim

Autoencoder latent dimensions.

--alpha

Significance level (0–1).

--seed_base

Base seed for runs.

--permutation_test_iterations

1 disables permutation p-values

Significance Preview

Percentile: 95.0th

Do not assume a Gaussian distribution. Gaussian shown for illustrative purposes only. The red line marks the (1 - alpha) * 100 percentile threshold.

Logging Output

Where to write JSON experiment logs (directory must exist).

--file_location

Directory path; must already exist.

--file_name

Output JSON filename.

Required: --source_dir, --target_dir, --target_list_path. If custom weights, also --model_weights_path.

Distribution Shift Experiment CLI Builder

Weights Mode

Dataset Configuration

Sampling & Test Controls

Logging Output