Quick help

psyop --help
psyop <command> --help

Model artifact

Most commands take a path to a .nc artifact produced by psyop model.

The artifact bundles:

• raw columns (for plotting & metadata),

• standardized design matrices,

• per-feature transforms & standardization stats,

• two GP heads (success probability; expected target conditional on success),

• convenience predictions and metadata (direction, seed, etc.).

Constraint syntax (used by suggest, optimal, plot2d, plot1d)

These commands accept extra CLI options that are not predeclared—they are parsed into feature constraints. Constraints are always interpreted in original units (before any internal log/standardization).

Pass any mix of:

• Fixed value (number) — clamp a feature at a single value and (for plots) remove it from the axes:

  --epochs 20
  --learning-rate 0.001

• Range (slice) — inclusive lower/upper bound:

  --dropout 0.0:0.2
  --x 0..2          # same as 0:2
  --width 64:512:64 # optional step token; used where applicable

• Choices (finite set) — list/tuple or Python-like range() call:

  --batch-size "(16, 32, 64)"
  --optimizer "[adam, sgd, adamw]"
  --layers "range(2,8,2)"        # -> (2, 4, 6, 8)

Rules:

• Unknown keys raise an error. Keys must match a feature name (case/spacing normalized) or a categorical base name (e.g., language).

• If you pass both a fixed value and a range/choices for the same feature, the fixed value wins.

• For suggest/optimal, bounds/choices are enforced strictly during optimization.

• For plot2d/plot1d, fixed features are clamped and not shown on axes; range constraints restrict the sweep domain even if historical points exist outside the range.

Tip (shells): quote lists/tuples and anything that contains commas or parentheses to avoid shell expansion.

1) Fit a model

psyop model INPUT.csv OUTPUT.nc [OPTIONS]

Arguments

• INPUT (CSV) — your experiment log.

• OUTPUT (.nc) — where to save the model artifact.

Options

• --target, -t TEXT — target column name (default: loss).

• --exclude TEXT — repeatable; columns to exclude from features.

• --direction, -d [min|max] — optimization direction for the target (default: min).

• --seed INTEGER — RNG seed (default: 0).

• --compress / --no-compress — compress numeric arrays inside the artifact (default: on).

Example

psyop model runs.csv output/trials.nc \
  --target loss --exclude run_id --exclude notes --direction min --seed 42

2) Suggest candidates (gradient + exploration)

psyop suggest MODEL.nc [OPTIONS] [EXTRA_CONSTRAINTS...]

What it does

• Optimizes inside your constraints using L-BFGS-B directly on the GP heads.

• Each suggestion is either exploit (minimize mean target with a smooth success penalty) or explore (maximize Expected Improvement gated by success probability). The choice per-take is random with probability --explore.

• Diversity among suggestions is handled internally (standardized space).

Options

• --output, -o PATH — write suggestions CSV (if omitted, prints the table).

• --count, -k INTEGER — number of suggestions to return (default: 1).

• --success-threshold FLOAT — feasibility threshold used in success penalty/gate (default: 0.8).

• --explore FLOAT — probability of using exploration (EI) for each take (default: 0.34).

• --seed INTEGER — RNG seed (default: 0).

Output CSV columns

rank, feature columns, pred_p_success, pred_target_mean, pred_target_sd.

Examples

# Fix epochs; bound dropout; ask for 12 diverse suggestions
psyop suggest output/trials.nc --epochs 20 --dropout 0.0:0.2 -k 12 -o output/suggest.csv

# Discrete numeric and categorical choices:
psyop suggest output/trials.nc \
  --batch-size "(16, 32, 64)" \
  --optimizer "[adam, sgd, adamw]" \
  -k 10

3) Optimal solution (constrained, gradient)

psyop optimal MODEL.nc [OPTIONS] [EXTRA_CONSTRAINTS...]

What it does

Finds a constraint-feasible point that optimizes the posterior mean target (min for losses; max if the model’s direction is max), with a smooth penalty that discourages low success probability. Uses L-BFGS-B and returns the best solution (one row).

Options

• --output PATH — write the row to CSV (prints table if omitted).

• --seed INTEGER — RNG seed (default: 0).

Output CSV columns

feature columns, pred_p_success, pred_target_mean, pred_target_sd.

Example

psyop optimal output/trials.nc --epochs 12 --dropout 0.0:0.2 -o output/optimal.csv

4) 2D Partial Dependence

psyop plot2d MODEL.nc [OPTIONS] [EXTRA_CONSTRAINTS...]

Options

• --output PATH — HTML file.

• --grid-size INTEGER — grid size per axis for 2D panels (default: 70).

• --use-log-scale-for-target — enable log10 colors for the target (toggle flag; default: off).

• --log-shift-epsilon FLOAT — epsilon shift for log colors (default: 1e-9).

• --colorscale TEXT — Plotly colorscale (default: RdBu).

• --show — open in a browser (defaults to True if no --output is provided).

• --n-contours INTEGER — contour levels (default: 12).

• --optimal / --no-optimal — overlay the current optimal solution (default: on).

• --suggest INTEGER — overlay up to N suggested points (default: 0).

• --width INTEGER / --height INTEGER — panel dimensions in pixels (default: 1000x1000).

• --seed INTEGER — RNG seed for overlays (default: 42).

Constraints

• Fixed features are clamped and removed from the axes.

• Ranges restrict the sweep domain for that feature.

Examples

# Clamp epochs; restrict dropout domain
psyop plot2d output/trials.nc --epochs 20 --dropout 0.0:0.2 --show

# Discrete choices for batch size
psyop plot2d output/trials.nc --batch-size "(16,32,64)" -o pairplot.html

5) 1D Partial Dependence (per-feature)

psyop plot1d MODEL.nc [OPTIONS] [EXTRA_CONSTRAINTS...]

Options

• --output PATH — HTML file.

• --csv-out PATH — tidy CSV export of PD values.

• --grid-size INTEGER — sweep resolution per panel (default: 300).

• --line-color TEXT — Plotly color string for mean/band (default: blue).

• --band-alpha FLOAT — fill alpha for ±2σ (default: 0.25).

• --show — open in a browser (defaults to True if no --output is provided).

• --log-y / --no-log-y — log scale for target axis (default: log).

• --log-y-eps FLOAT — clamp for log-Y (default: 1e-9).

• --optimal / --no-optimal — overlay the current optimal solution (default: on).

• --suggest INTEGER — overlay up to N suggested points (default: 0).

• --width INTEGER / --height INTEGER — figure dimensions in pixels (default: 1000x1000).

• --seed INTEGER — RNG seed for overlays (default: 42).

Constraints

Same as Constraint syntax. Fixed features are not plotted; ranges clip the sweep domain.

Examples

psyop plot1d output/trials.nc --epochs 20 --dropout 0.0:0.2 \
  --csv-out output/pd.csv -o output/pd.html --show

Notes

• Colorscales are Plotly names (e.g. RdBu, Viridis, Inferno).

• For plots, historical points are drawn even if outside your specified range, but the sweep domain (and axes) respect your bounds.

• All constraint parsing is printed once as Constraints: ... for sanity checking.

Examples at a glance

# Fit
psyop model runs.csv output/trials.nc -t loss --exclude run_id --seed 0

# Suggest inside bounds, with discrete choices
psyop suggest output/trials.nc \
  --epochs 20 \
  --dropout 0.0:0.2 \
  --batch-size "(16,32,64)" \
  -k 12 -o output/suggest.csv

# Optimal (mean-based) with a minimum constraint set
psyop optimal output/trials.nc --epochs 12 --dropout 0.0:0.2 -o output/optimal.csv

# Pairwise PD conditioned on epochs
psyop plot2d output/trials.nc --epochs 20 --show

# 1D PD with CSV export
psyop plot1d output/trials.nc --csv-out output/pd.csv -o output/pd.html

Build a model

build_model(
    input=Path("runs.csv"),
    output=Path("output/trials.nc"),
    target="loss",
    exclude=["run_id", "notes"],
    direction="min",        # "min" or "max"
    seed=42,
    compress=True,          # compress numeric arrays within the .nc
)

Load a model

ds = xr.load_dataset("output/trials.nc")

Suggest candidates

suggestions = suggest(
    model=ds,
    output=None,
    count=12,
    success_threshold=0.8,
    explore=0.5,          # probability of EI per take
    seed=0,
    epochs=20,                          # fixed
    dropout=slice(0.0, 0.2),            # range
    batch_size=(16, 32, 64),            # choices
)
print(suggestions.head())

Optimal solution

best = optimal(
    model=ds,
    output=None,
    seed=0,
    epochs=12,
    dropout=slice(0.0, 0.2),
)
print(best)

2D Partial Dependence

# Fixed features are clamped and removed from axes.
# Ranges clip the sweep domain even if historical points exist outside the range.
plot2d(
    model=ds,                    # xarray.Dataset
    output=Path("pairplot.html"),
    use_log_scale_for_target=False,
    log_shift_epsilon=1e-9,
    colorscale="RdBu",
    show=False,
    n_contours=12,
    optimal=True,                # overlay current optimal solution
    suggest=5,                   # overlay top-N suggestions
    width=1000,
    height=1000,
    epochs=20,
    dropout=slice(0.0, 0.2),
)

1D Partial Dependence

plot1d(
    model=ds,
    output=Path("pd.html"),
    csv_out=Path("pd.csv"),
    grid_size=300,
    line_color="blue",
    band_alpha=0.25,
    show=False,
    use_log_scale_for_target_y=True,
    log_y_epsilon=1e-9,
    optimal=True,
    suggest=3,
    width=1000,
    height=1000,
    epochs=20,
    dropout=slice(0.0, 0.2),
)

Return types and side effects

• build_model(...) → None (writes a .nc file).

• suggest(...) → pandas.DataFrame (and optionally writes a CSV if output is provided).

• optimal(...) → pandas.DataFrame (and optionally writes a CSV if output is provided).

• plot2d(...) → None (writes HTML if output is provided; may open a browser if show=True).

• plot1d(...) → None (writes HTML/CSV if paths are provided; may open a browser if show=True).

Constraint objects in Python

• Fixed: epochs=20 or learning_rate=1e-3.

• Range: dropout=slice(0.0, 0.2) (inclusive ends).

• Choices: batch_size=(16, 32, 64) (tuple/list of finite values).

• Integer grids: layers=tuple(range(2, 9, 2)) → (2, 4, 6, 8).

All constraints are interpreted in original units of your data. Bounds are enforced for optimization and sweep ranges; fixed values remove the feature from PD axes.

Determinism

Randomness is dataset-aware. When you pass --seed S, Psyop may mix that seed with a checksum of the model artifact so that:

• Same dataset + same seed → same suggestions.

• Different dataset + same seed → different, but reproducible suggestions.