File size: 11,868 Bytes
44706c2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 | #!/usr/bin/env python
# -*- coding: utf-8 -*-
import argparse
import glob
import os
from typing import List, Optional, Tuple
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib as mpl
import re
ENV_LIST: List[str] = [
"AutoEncodeEasy",
"BattleShipEasy",
"BreakoutEasy",
"CartPoleEasy",
"CountRecallEasy",
"MineSweeperEasy",
"NavigatorEasy",
"NoisyCartPoleEasy",
"SkittlesEasy",
"TetrisEasy",
]
plt.rcParams['text.usetex'] = True
plt.rcParams['font.family'] = 'sans-serif'
plt.rcParams['font.sans-serif'] = ['Arial']
plt.rcParams['text.latex.preamble'] = r'\usepackage{amsmath} \usepackage{amssymb} \usepackage{amsfonts}'
def _find_csvs_for(env_name: str, partial: bool, recall_density_dir: str) -> List[str]:
pattern = os.path.join(
recall_density_dir, f"saliency_results_*_{env_name}_Partial={partial}_*.csv"
)
return sorted(glob.glob(pattern))
def _extract_pos_columns(df: pd.DataFrame) -> Tuple[np.ndarray, List[str]]:
pos_cols = [c for c in df.columns if c.startswith("pos_")]
# Sort by numeric suffix
def _key(c: str) -> float:
try:
return float(c.split("pos_")[-1])
except Exception:
return 0.0
pos_cols.sort(key=_key)
return df[pos_cols].to_numpy(dtype=float), pos_cols
def _thirds_from_distribution_rows(pos_values: np.ndarray) -> np.ndarray:
"""Given matrix [num_rows, num_positions], compute thirds per row.
Returns array [num_rows, 3] with per-third sums normalized so each row sums to 1
(if a row sums to 0, it remains zeros).
"""
if pos_values.size == 0:
return np.zeros((pos_values.shape[0], 3), dtype=float)
num_cols = pos_values.shape[1]
e1 = num_cols // 3
e2 = (num_cols * 2) // 3
thirds = np.stack(
[
pos_values[:, :e1].sum(axis=1),
pos_values[:, e1:e2].sum(axis=1),
pos_values[:, e2:].sum(axis=1),
],
axis=1,
)
row_sums = thirds.sum(axis=1, keepdims=True)
norm = np.zeros_like(thirds, dtype=float)
mask = row_sums[:, 0] > 0
if np.any(mask):
norm[mask] = thirds[mask] / row_sums[mask]
return norm
def compute_env_mode_summary(env_name: str, partial: bool, recall_density_dir: str) -> np.ndarray:
"""Aggregate all CSVs for (env, partial) across files and seeds.
Returns a vector of length 3 that sums to 1 (or zeros if nothing found).
"""
csv_files = _find_csvs_for(env_name, partial, recall_density_dir)
if not csv_files:
return np.zeros(3, dtype=float)
thirds_all = []
for path in csv_files:
try:
df = pd.read_csv(path)
except Exception:
continue
if df.empty:
continue
pos_values, _ = _extract_pos_columns(df)
thirds_rows = _thirds_from_distribution_rows(pos_values)
if thirds_rows.size == 0:
continue
thirds_all.append(thirds_rows)
if not thirds_all:
return np.zeros(3, dtype=float)
thirds_concat = np.concatenate(thirds_all, axis=0) # [num_total_rows, 3]
mean_thirds = thirds_concat.mean(axis=0)
mean_sum = mean_thirds.sum()
return mean_thirds / mean_sum if mean_sum > 0 else np.zeros(3, dtype=float)
def _load_data_from_summary_csv(summary_csv: str) -> np.ndarray:
"""Load pre-aggregated bar data with columns EnvName, Partial, third_1..third_3."""
df = pd.read_csv(summary_csv)
if df.empty:
raise ValueError(f"Summary CSV is empty: {summary_csv}")
required_columns = {"EnvName", "Partial", "third_1", "third_2", "third_3"}
missing = required_columns.difference(df.columns)
if missing:
raise ValueError(
f"Summary CSV is missing required columns: {sorted(missing)}"
)
partial_series = df["Partial"].map(
lambda x: x if isinstance(x, bool) else str(x).strip().lower() == "true"
)
data = np.zeros((len(ENV_LIST), 2, 3), dtype=float)
partial_modes = [False, True]
for i, env in enumerate(ENV_LIST):
for j, part in enumerate(partial_modes):
match = df[(df["EnvName"] == env) & (partial_series == part)]
if match.empty:
continue
row = match.iloc[0]
thirds = np.array([row["third_1"], row["third_2"], row["third_3"]], dtype=float)
total = thirds.sum()
data[i, j] = thirds / total if total > 0 else np.zeros(3, dtype=float)
return data
def _load_data_from_recall_density_dir(recall_density_dir: str) -> np.ndarray:
"""Aggregate raw recall-density CSVs into plot-ready thirds data."""
partial_modes = [False, True]
data = np.zeros((len(ENV_LIST), len(partial_modes), 3), dtype=float)
for i, env in enumerate(ENV_LIST):
for j, part in enumerate(partial_modes):
data[i, j] = compute_env_mode_summary(env, part, recall_density_dir)
return data
def _plot_mode(env_values: np.ndarray, colors: list[str], mode_title: str, output_path: str, dpi: int) -> None:
fig, ax = plt.subplots(figsize=(min(20, 1.8 * len(ENV_LIST)), 7))
x = np.arange(len(ENV_LIST))
# Stacked thirds for single mode
left = np.zeros(len(ENV_LIST))
for k in range(3):
values = env_values[:, k]
ax.bar(
x,
values,
0.6,
bottom=left,
color=colors[k],
edgecolor="white",
linewidth=0.6,
alpha=0.9,
)
left += values
# x ticks and fonts
ax.set_xticks(x)
ax.set_xticklabels([e.replace("Easy", "") for e in ENV_LIST], rotation=30, ha="right", fontsize=30)
ax.tick_params(axis="both", labelsize=30)
ax.set_ylim(0, 1.05)
ax.set_ylabel(r"Saliency mass per third", fontsize=14)
# Legend (thirds only)
third_labels = [r"$[0,\frac{1}{3})$", r"$[\frac{1}{3},\frac{2}{3})$", r"$[\frac{2}{3},1)$"]
legend_handles = [
plt.Rectangle((0, 0), 1, 1, color=colors[0]),
plt.Rectangle((0, 0), 1, 1, color=colors[1]),
plt.Rectangle((0, 0), 1, 1, color=colors[2]),
]
ax.legend(legend_handles, third_labels, loc="center left", bbox_to_anchor=(1.02, 0.5), fontsize=30, frameon=True, fancybox=True, handlelength=1.2, handletextpad=0.6)
ax.set_title(f"Aggregate Recall Density — {mode_title}", fontsize=40)
plt.tight_layout(rect=[0, 0, 0.82, 1])
out_dir = os.path.dirname(output_path)
if out_dir and not os.path.exists(out_dir):
os.makedirs(out_dir, exist_ok=True)
fig.savefig(output_path, dpi=dpi)
# If the user didn't request a PDF explicitly, save a copy as PDF too
base, ext = os.path.splitext(output_path)
if ext.lower() != ".pdf":
fig.savefig(base + ".pdf")
plt.close(fig)
def _plot_bars_on_ax(ax, env_values: np.ndarray, colors: list[str], show_left_axis: bool, ylabel_math: str | None = None):
x = np.arange(len(ENV_LIST))
left = np.zeros(len(ENV_LIST))
for k in range(3):
vals = env_values[:, k]
ax.bar(
x,
vals,
0.6,
bottom=left,
color=colors[k],
edgecolor="white",
linewidth=0.6,
alpha=0.9,
)
left += vals
ax.set_xticks(x)
# Use single-line labels
single_line_labels = [e.replace("Easy", "") for e in ENV_LIST]
ax.set_xticklabels(single_line_labels, rotation=30, ha="right", fontsize=30)
# ax.tick_params(axis="x", pad=15)
ax.set_ylim(0, 1.05)
# Beautify: remove top/right spines
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
if show_left_axis:
ax.tick_params(axis="y", labelsize=35, left=True, labelleft=True)
if ylabel_math:
ax.set_ylabel(ylabel_math, fontsize=35, rotation=90, labelpad=10)
else:
ax.tick_params(axis="y", left=False, labelleft=False)
def plot_summary(
output_path: str,
dpi: int = 300,
recall_density_dir: Optional[str] = None,
summary_csv: Optional[str] = None,
):
if summary_csv is not None:
data = _load_data_from_summary_csv(summary_csv)
elif recall_density_dir is not None:
data = _load_data_from_recall_density_dir(recall_density_dir)
else:
raise ValueError("Provide either recall_density_dir or summary_csv.")
# Colors
mdp_colors = ["#C6DBEF", "#6BAED6", "#2171B5"]
pomdp_colors = ["#FDD0A2", "#FDAE6B", "#E6550D"]
fig = plt.figure(figsize=(max(32, 1.7 * len(ENV_LIST) * 2), 8.0))
gs = fig.add_gridspec(1, 3, width_ratios=[2.2, 0.9, 2.2], wspace=0.12)
ax_left = fig.add_subplot(gs[0, 0])
ax_center = fig.add_subplot(gs[0, 1])
ax_right = fig.add_subplot(gs[0, 2])
ylabel_math = r"$\mathbb{E}_{\pi, f}[\,\delta(Q(\mathbf{x},\tau))\,]$"
_plot_bars_on_ax(ax_left, data[:, 0, :], mdp_colors, show_left_axis=True, ylabel_math=ylabel_math)
_plot_bars_on_ax(ax_right, data[:, 1, :], pomdp_colors, show_left_axis=False)
# Center block: title + legend (combined 6 items)
ax_center.axis("off")
ax_center.set_title("Aggregate Recall Density", fontsize=40, pad=10)
mdp_handles = [
plt.Rectangle((0, 0), 1, 1, color=mdp_colors[0]),
plt.Rectangle((0, 0), 1, 1, color=mdp_colors[1]),
plt.Rectangle((0, 0), 1, 1, color=mdp_colors[2]),
]
pomdp_handles = [
plt.Rectangle((0, 0), 1, 1, color=pomdp_colors[0]),
plt.Rectangle((0, 0), 1, 1, color=pomdp_colors[1]),
plt.Rectangle((0, 0), 1, 1, color=pomdp_colors[2]),
]
mdp_labels = [
r"MDP $0<\tau<0.33$",
r"MDP $0.33\leq\,\tau<0.66$",
r"MDP $0.66\leq\,\tau<1.0$",
]
pomdp_labels = [
r"POMDP $0<\tau<0.33$",
r"POMDP $0.33\leq\,\tau<0.66$",
r"POMDP $0.66\leq\,\tau<1.0$",
]
common_legend_kwargs = dict(frameon=True, fancybox=True, fontsize=30, handlelength=1.8, handletextpad=0.8)
combined_handles = mdp_handles + pomdp_handles
combined_labels = mdp_labels + pomdp_labels
ax_center.legend(
combined_handles,
combined_labels,
loc="center",
bbox_to_anchor=(0.44, 0.5),
ncol=1,
borderaxespad=0.0,
labelspacing=0.6,
**common_legend_kwargs,
)
plt.subplots_adjust(left=0.07, right=0.985, bottom=0.20, top=0.92, wspace=0.12)
out_dir = os.path.dirname(output_path)
if out_dir and not os.path.exists(out_dir):
os.makedirs(out_dir, exist_ok=True)
fig.savefig(output_path, dpi=dpi)
base, _ = os.path.splitext(output_path)
if not output_path.lower().endswith(".pdf"):
fig.savefig(base + ".pdf")
plt.close(fig)
def main():
parser = argparse.ArgumentParser(description="Plot aggregated saliency thirds per env and mode")
parser.add_argument(
"--recall_density_dir",
type=str,
default=None,
help="Directory containing per-weight density CSVs",
)
parser.add_argument(
"--summary_csv",
type=str,
default=None,
help="Pre-aggregated bar-summary CSV generated by density_analysis_{pqn,ppo}.py",
)
parser.add_argument(
"--output",
type=str,
default="your_output_pdf",
help="Path to save the summary figure (PNG & PDF)",
)
parser.add_argument("--dpi", type=int, default=300)
args = parser.parse_args()
if args.summary_csv is None and args.recall_density_dir is None:
raise SystemExit("Please provide either --summary_csv or --recall_density_dir")
plot_summary(
output_path=args.output,
dpi=args.dpi,
recall_density_dir=args.recall_density_dir,
summary_csv=args.summary_csv,
)
print(f"Saved summary to {args.output}")
if __name__ == "__main__":
main() |