updates in techreport artifacts.

artifacts/tech_report_v1/appendix_training_curves/curve_plotting/plot_config.py

@@ -17,9 +17,11 @@
 from pathlib import Path
 
 # --- Paths -----------------------------------------------------------------
-# Derived from this file's location (``<repo>/logs/curve_plotting/plot_config.py``)
-# so the scripts work from any checkout without a hardcoded ``~`` path.
-REPO_ROOT = Path(__file__).resolve().parents[2]
+# Derived from this file's location (``<repo>/artifacts/tech_report_v1/
+# appendix_training_curves/curve_plotting/plot_config.py``) so the scripts
+# work from any checkout without a hardcoded ``~`` path.
+REPO_ROOT = Path(__file__).resolve().parents[4]
+REPORT_ROOT = Path(__file__).resolve().parent
 DEFAULT_LOG_DIR = REPO_ROOT / "logs" / "self_tuning"
 DEFAULT_PLOT_DIR = REPO_ROOT / "logs" / "curve_plotting"
 DEFAULT_DATA_PATH = DEFAULT_PLOT_DIR / "curve_data.csv"

artifacts/tech_report_v1/appendix_training_curves/curve_plotting/plot_curves.py

@@ -20,6 +20,7 @@
 
 import argparse
 import os
+import sys
 from functools import partial
 from multiprocessing import Pool
 from pathlib import Path
@@ -28,6 +29,9 @@
 
 matplotlib.use("Agg")  # headless-safe; avoids the Qt "no display" crash.
 
+if __name__ == "__main__" and not __package__:
+    sys.path.insert(0, str(Path(__file__).resolve().parents[4]))
+
 import matplotlib.pyplot as plt  # noqa: E402
 import numpy as np  # noqa: E402
 import pandas as pd  # noqa: E402
@@ -36,9 +40,12 @@
     ALGO_CONFIGS,
     DEFAULT_DATA_PATH,
     DEFAULT_PLOT_DIR,
-    PLOT_RC_PARAMS,
     is_higher_better,
 )
+from artifacts.tech_report_v1.report_utils import (  # noqa: E402
+    NOTEBOOK_STYLE,
+    set_plot_style,
+)
 
 GRID_POINTS = 150
 TIME_COL = "accumulated_submission_time"
@@ -134,19 +141,6 @@ def compute_ylimits(values, target_value, higher_is_better, zoom):
     return ymin, ymax
 
 
-def _style_axis(ax):
-    ax.grid(True, which="major", color="#e8e8e8", linewidth=0.8)
-    ax.spines["top"].set_visible(False)
-    ax.spines["right"].set_visible(False)
-    ax.spines["left"].set_color("#cccccc")
-    ax.spines["bottom"].set_color("#cccccc")
-
-
-def _legend(ax):
-    ax.legend(fontsize=9, frameon=True, facecolor="white", framealpha=0.9,
-              edgecolor="#e5e5e5")
-
-
 def _plot_metric(ax, submissions, wdf, x_col, x_scale):
     """Metric mean + std band per submission; returns pooled values for ylim."""
     all_vals = []
@@ -223,8 +217,7 @@ def plot_workload(workload, wdf, submissions, out_dir, dpi=200):
                        label=f"Target ({target:g})")
         ax.set_xlabel(xlab, fontweight="semibold")
         ax.set_ylabel(mlabel, fontweight="semibold")
-        _legend(ax)
-        _style_axis(ax)
+        ax.legend()
     ax_mt.set_title(f"Validation {metric} vs. wall-clock time", fontsize=12)
     ax_ms.set_title(f"Validation {metric} vs. training steps", fontsize=12)
 
@@ -235,8 +228,7 @@ def plot_workload(workload, wdf, submissions, out_dir, dpi=200):
         for ax, xlab in [(ax_gt, "Wall-clock time (hours)"), (ax_gs, "Training steps (x10³)")]:
             ax.set_xlabel(xlab, fontweight="semibold")
             ax.set_ylabel(glabel, fontweight="semibold")
-            _legend(ax)
-            _style_axis(ax)
+            ax.legend()
         ax_gt.set_title("Distance to target vs. wall-clock time", fontsize=12)
         ax_gs.set_title("Distance to target vs. training steps", fontsize=12)
     else:
@@ -253,7 +245,7 @@ def plot_workload(workload, wdf, submissions, out_dir, dpi=200):
 
 def _render_job(job, dpi):
     workload, wdf, submissions, out_dir = job
-    plt.rcParams.update(PLOT_RC_PARAMS)  # rcParams don't survive a spawn start.
+    set_plot_style(NOTEBOOK_STYLE)  # rcParams don't survive a spawn start.
     return plot_workload(workload, wdf, submissions, out_dir, dpi=dpi)
 
 
@@ -290,7 +282,7 @@ def main() -> None:
     if not args.data.is_file():
         raise SystemExit(f"No curve data at {args.data}. Run parse_logs.py first (pass 1).")
 
-    plt.rcParams.update(PLOT_RC_PARAMS)
+    set_plot_style(NOTEBOOK_STYLE)
     data = pd.read_csv(args.data)
     algos = list(ALGO_CONFIGS) if args.algo == "all" else [args.algo]
     print(f"Plotting algorithms: {algos}")

artifacts/tech_report_v1/appendix_training_curves/plot.py

@@ -0,0 +1,97 @@
+# ---
+# jupyter:
+#   jupytext:
+#     formats: ipynb,py:percent
+#     text_representation:
+#       extension: .py
+#       format_name: percent
+#       format_version: '1.3'
+#       jupytext_version: 1.19.5
+#   kernelspec:
+#     display_name: myenv
+#     language: python
+#     name: python3
+# ---
+
+# %% [markdown]
+# # Curve Plotter
+#
+# This notebook imports the helper functions from `plot_helper.py` to generate and display publication-grade comparison plots for self-tuning algorithms (e.g., `sfadamw`, `muon`, `ademamix`, etc.).
+# You can customize the styling parameters below and select which algorithms/workloads to plot.
+
+# %%
+import sys
+from pathlib import Path
+
+# Ensure the scratch directory is in python path
+scratch_dir = Path('./').resolve()
+if str(scratch_dir) not in sys.path:
+    sys.path.append(str(scratch_dir))
+
+import plot_helper
+from plot_helper import (
+    ALGO_CONFIGS,
+    configure_styling,
+    plot_workload,
+    find_workloads,
+    find_repo_root,
+)
+
+# %% [markdown]
+# ## Customize Styling
+#
+# You can override any Matplotlib RC parameters in the dictionary below. Call `configure_styling(custom_params)` to apply them.
+
+# %%
+# configure_styling()'s defaults already match score_submissions.py's theme
+# (Section 4 leaderboards); add overrides here only if you want to deviate
+# from that shared look for this notebook.
+custom_styles = {}
+
+# Apply styling configurations
+configure_styling(custom_styles)
+
+# %% [markdown]
+# ## Select Algorithm & Generate Plots
+#
+# Generates plots for every algorithm in `ALGO_CONFIGS` (set `ALGO_NAMES` to a
+# subset to limit it) and the given zoom type (`'log'` or `'percentile'`).
+# Outputs go to `out/<algo>/<workload>_curves.png`, next to this notebook —
+# one subdirectory per algorithm, mirroring each algorithm's `sub_dir`.
+
+# %%
+# Parameters - feel free to change these
+ALGO_NAMES = list(ALGO_CONFIGS.keys())  # e.g. ['sfadamw', 'muon'] to limit to a subset
+ZOOM_TYPE = 'log'                       # Options: 'log', 'percentile'
+
+# Configure directories (repo root found by walking up from plot_helper.py,
+# not hardcoded to any user's home directory). base_save_dir is anchored to
+# plot_helper.py's own location (reliable even when this notebook's own
+# __file__ isn't set) so outputs land in out/ next to this notebook.
+REPO_ROOT = find_repo_root()
+base_log_dir = REPO_ROOT / 'logs' / 'self_tuning'
+base_save_dir = Path(plot_helper.__file__).resolve().parent / 'out'
+
+for algo_name in ALGO_NAMES:
+    if algo_name not in ALGO_CONFIGS:
+        print(f"Algorithm '{algo_name}' not found in ALGO_CONFIGS. Available: {list(ALGO_CONFIGS.keys())}")
+        continue
+
+    config = ALGO_CONFIGS[algo_name]
+    submissions = config['submissions']
+
+    # Find all workloads for this algorithm
+    workloads = find_workloads(base_log_dir, submissions)
+    print(f"\n=== {algo_name}: found workloads {workloads} ===")
+
+    for workload in workloads:
+        print(f"Processing workload: {workload}")
+        plot_workload(
+            workload=workload,
+            algo_name=algo_name,
+            submissions=submissions,
+            base_log_dir=base_log_dir,
+            base_save_dir=base_save_dir,
+            zoom=ZOOM_TYPE,
+            show_plot=True,  # displays every figure inline, in addition to saving it
+        )