Approach refinement
This commit is contained in:
Johnny Fernandes
+21
-188
@@ -25,12 +25,7 @@ import os
|
||||
import time
|
||||
from copy import deepcopy
|
||||
|
||||
import matplotlib
|
||||
matplotlib.use("Agg")
|
||||
import matplotlib.pyplot as plt
|
||||
import matplotlib.patches as mpatches
|
||||
import numpy as np
|
||||
from matplotlib.collections import LineCollection
|
||||
from stable_baselines3 import PPO
|
||||
from stable_baselines3.common.callbacks import BaseCallback
|
||||
from stable_baselines3.common.vec_env import (
|
||||
@@ -40,15 +35,13 @@ from stable_baselines3.common.vec_env import (
|
||||
)
|
||||
|
||||
from herding_env import HerdingEnv
|
||||
|
||||
|
||||
# ── Colours ──────────────────────────────────────────────────────────────────
|
||||
|
||||
SHEEP_COLORS = [
|
||||
"#e41a1c", "#377eb8", "#4daf4a", "#984ea3", "#ff7f00",
|
||||
"#a65628", "#f781bf", "#999999", "#66c2a5", "#fc8d62",
|
||||
]
|
||||
DOG_COLOR = "#4e342e"
|
||||
from viz import (
|
||||
run_and_record,
|
||||
plot_trajectory,
|
||||
plot_timeseries,
|
||||
plot_success_rate,
|
||||
save_episode_gif,
|
||||
)
|
||||
|
||||
|
||||
# ── Callbacks ────────────────────────────────────────────────────────────────
|
||||
@@ -198,178 +191,6 @@ def evaluate(model, vn_template, n_sheep, n_episodes, max_steps,
|
||||
return result
|
||||
|
||||
|
||||
# ── Visualization helpers ────────────────────────────────────────────────────
|
||||
|
||||
def _draw_field(ax):
|
||||
ax.set_xlim(-16, 16)
|
||||
ax.set_ylim(-16, 16)
|
||||
ax.set_aspect("equal")
|
||||
ax.set_facecolor("#dcedc8")
|
||||
ax.add_patch(mpatches.Rectangle((-15, -15), 30, 30,
|
||||
fill=False, edgecolor="#795548", lw=2))
|
||||
ax.add_patch(mpatches.Rectangle((10, -15), 3, 7,
|
||||
facecolor="#ffe082", edgecolor="#795548", lw=2))
|
||||
ax.text(11.5, -11.5, "pen", ha="center", va="center",
|
||||
fontsize=8, color="#795548")
|
||||
|
||||
|
||||
def _faded_path(ax, xs, ys, color, lw=1.5, label=None):
|
||||
n = len(xs)
|
||||
if n < 2:
|
||||
return
|
||||
points = np.array([xs, ys]).T.reshape(-1, 1, 2)
|
||||
segs = np.concatenate([points[:-1], points[1:]], axis=1)
|
||||
alphas = np.linspace(0.15, 1.0, len(segs))
|
||||
colors = [(*matplotlib.colors.to_rgb(color), a) for a in alphas]
|
||||
ax.add_collection(LineCollection(segs, colors=colors, linewidth=lw))
|
||||
if label:
|
||||
ax.plot([], [], color=color, lw=lw, label=label)
|
||||
|
||||
|
||||
def run_and_record(model, vn_template, n_sheep, max_steps,
|
||||
reward_cfg=None, seed=42):
|
||||
"""Run one deterministic episode and return full history."""
|
||||
raw = DummyVecEnv([make_env(n_sheep, seed, max_steps, reward_cfg)])
|
||||
vn = VecNormalize(raw, norm_obs=True, norm_reward=False, training=False)
|
||||
vn.obs_rms = deepcopy(vn_template.obs_rms)
|
||||
vn.ret_rms = deepcopy(vn_template.ret_rms)
|
||||
|
||||
obs = vn.reset()
|
||||
inner = vn.envs[0]
|
||||
done = False
|
||||
|
||||
dog_xs, dog_ys = [], []
|
||||
sheep_xs = [[] for _ in range(n_sheep)]
|
||||
sheep_ys = [[] for _ in range(n_sheep)]
|
||||
radii = []
|
||||
pen_dists = [[] for _ in range(n_sheep)]
|
||||
action_mags = []
|
||||
rewards = []
|
||||
penned_at = [None] * n_sheep
|
||||
step = 0
|
||||
|
||||
while not done:
|
||||
action, _ = model.predict(obs, deterministic=True)
|
||||
obs, reward, dones, infos = vn.step(action)
|
||||
done = dones[0]
|
||||
step += 1
|
||||
|
||||
dog_xs.append(float(inner.dog_pos[0]))
|
||||
dog_ys.append(float(inner.dog_pos[1]))
|
||||
com, radius, _ = inner._flock_stats()
|
||||
radii.append(radius)
|
||||
rewards.append(float(reward[0]))
|
||||
action_mags.append(float(np.linalg.norm(action[0])))
|
||||
|
||||
for i in range(n_sheep):
|
||||
sheep_xs[i].append(float(inner.sheep_pos[i][0]))
|
||||
sheep_ys[i].append(float(inner.sheep_pos[i][1]))
|
||||
pen_dists[i].append(
|
||||
float(np.linalg.norm(inner.sheep_pos[i] - inner.PEN_CENTER)))
|
||||
if inner.penned[i] and penned_at[i] is None:
|
||||
penned_at[i] = step
|
||||
|
||||
n_penned = infos[0].get("n_penned", 0)
|
||||
vn.close()
|
||||
|
||||
return dict(
|
||||
dog_xs=dog_xs, dog_ys=dog_ys,
|
||||
sheep_xs=sheep_xs, sheep_ys=sheep_ys,
|
||||
radii=radii, pen_dists=pen_dists,
|
||||
action_mags=action_mags, rewards=rewards,
|
||||
penned_at=penned_at,
|
||||
n_penned=n_penned, n_sheep=n_sheep,
|
||||
success=n_penned == n_sheep, steps=step,
|
||||
)
|
||||
|
||||
|
||||
def plot_trajectory(hist, out_path):
|
||||
fig, ax = plt.subplots(figsize=(7, 7))
|
||||
_draw_field(ax)
|
||||
for i in range(hist["n_sheep"]):
|
||||
c = SHEEP_COLORS[i % len(SHEEP_COLORS)]
|
||||
xs, ys = hist["sheep_xs"][i], hist["sheep_ys"][i]
|
||||
_faded_path(ax, xs, ys, c, lw=1.2, label=f"sheep {i+1}")
|
||||
ax.plot(xs[0], ys[0], "o", color=c, ms=7, zorder=4)
|
||||
end = hist["penned_at"][i] if hist["penned_at"][i] is not None else -1
|
||||
ax.plot(xs[end], ys[end], "*", color=c, ms=11, zorder=5)
|
||||
_faded_path(ax, hist["dog_xs"], hist["dog_ys"], DOG_COLOR, lw=2.0,
|
||||
label="dog")
|
||||
ax.plot(hist["dog_xs"][0], hist["dog_ys"][0], "s", color=DOG_COLOR,
|
||||
ms=10, zorder=5)
|
||||
ax.plot(hist["dog_xs"][-1], hist["dog_ys"][-1], "D", color=DOG_COLOR,
|
||||
ms=10, zorder=5)
|
||||
result = ("SUCCESS" if hist["success"]
|
||||
else f"FAIL ({hist['n_penned']}/{hist['n_sheep']})")
|
||||
ax.set_title(f"n={hist['n_sheep']} {result} {hist['steps']} steps",
|
||||
fontsize=12)
|
||||
ax.legend(loc="upper left", fontsize=8)
|
||||
plt.tight_layout()
|
||||
fig.savefig(out_path, dpi=120)
|
||||
plt.close(fig)
|
||||
|
||||
|
||||
def plot_timeseries(hist, out_path):
|
||||
t = np.arange(hist["steps"])
|
||||
fig, axes = plt.subplots(4, 1, figsize=(12, 10), sharex=True)
|
||||
|
||||
axes[0].plot(t, hist["radii"], color="steelblue")
|
||||
axes[0].axhline(5.0, color="orange", ls="--", lw=1, label="compact (5m)")
|
||||
axes[0].set_ylabel("flock radius (m)")
|
||||
axes[0].legend(fontsize=8)
|
||||
axes[0].set_title("Flock radius")
|
||||
|
||||
for i in range(hist["n_sheep"]):
|
||||
c = SHEEP_COLORS[i % len(SHEEP_COLORS)]
|
||||
axes[1].plot(t, hist["pen_dists"][i], color=c, lw=1,
|
||||
label=f"sheep {i+1}")
|
||||
if hist["penned_at"][i] is not None:
|
||||
axes[1].axvline(hist["penned_at"][i], color=c, ls=":", lw=1)
|
||||
axes[1].set_ylabel("dist to pen (m)")
|
||||
axes[1].legend(fontsize=7, ncol=min(hist["n_sheep"], 5))
|
||||
axes[1].set_title("Per-sheep distance to pen")
|
||||
|
||||
axes[2].plot(t, hist["action_mags"], color="tomato", lw=1)
|
||||
axes[2].axhline(1.0, color="gray", ls="--", lw=1, label="max")
|
||||
axes[2].set_ylabel("action ||(vx,vy)||")
|
||||
axes[2].set_ylim(0, 1.5)
|
||||
axes[2].set_title("Dog action magnitude")
|
||||
axes[2].legend(fontsize=8)
|
||||
|
||||
axes[3].plot(t, hist["rewards"], color="purple", lw=1, alpha=0.7)
|
||||
axes[3].axhline(0, color="black", lw=0.5)
|
||||
axes[3].set_ylabel("reward")
|
||||
axes[3].set_xlabel("step")
|
||||
axes[3].set_title("Reward per step")
|
||||
|
||||
result = ("SUCCESS" if hist["success"]
|
||||
else f"FAIL ({hist['n_penned']}/{hist['n_sheep']})")
|
||||
fig.suptitle(f"n_sheep={hist['n_sheep']} {result} {hist['steps']} steps",
|
||||
fontsize=13)
|
||||
plt.tight_layout()
|
||||
fig.savefig(out_path, dpi=120)
|
||||
plt.close(fig)
|
||||
|
||||
|
||||
def plot_success_rate(stage_results, out_path):
|
||||
fig, ax = plt.subplots(figsize=(8, 4))
|
||||
ns = [r["n_sheep"] for r in stage_results]
|
||||
srs = [r["sr"] * 100 for r in stage_results]
|
||||
bars = ax.bar(ns, srs, color="steelblue", edgecolor="white")
|
||||
ax.set_xlabel("Sheep count")
|
||||
ax.set_ylabel("Success rate (%)")
|
||||
ax.set_ylim(0, 105)
|
||||
ax.axhline(90, color="orange", ls="--", lw=1, label="90% target")
|
||||
for bar, sr in zip(bars, srs):
|
||||
ax.text(bar.get_x() + bar.get_width() / 2,
|
||||
bar.get_height() + 1, f"{sr:.0f}%",
|
||||
ha="center", fontsize=9)
|
||||
ax.legend()
|
||||
ax.set_title("Evaluation success rate per sheep count")
|
||||
plt.tight_layout()
|
||||
fig.savefig(out_path, dpi=120)
|
||||
plt.close(fig)
|
||||
|
||||
|
||||
# ── CLI ──────────────────────────────────────────────────────────────────────
|
||||
|
||||
@@ -400,6 +221,11 @@ def parse_args():
|
||||
p.add_argument("--max-steps", type=int, default=2500)
|
||||
p.add_argument("--eval-episodes", type=int, default=30)
|
||||
p.add_argument("--run-dir", type=str, default=None)
|
||||
p.add_argument("--no-gif", action="store_true",
|
||||
help="Skip per-stage GIF rendering (PNGs still produced).")
|
||||
p.add_argument("--gif-fps", type=int, default=20)
|
||||
p.add_argument("--gif-skip", type=int, default=3,
|
||||
help="Keep every Nth frame (smaller GIF; default 3).")
|
||||
return p.parse_args()
|
||||
|
||||
|
||||
@@ -441,13 +267,15 @@ def main():
|
||||
vn = VecNormalize(train_env, norm_obs=True, norm_reward=True,
|
||||
clip_obs=10.0)
|
||||
|
||||
# Model
|
||||
# Model — force CPU (PPO with MLP runs faster on CPU than GPU; SB3 warns
|
||||
# about this otherwise).
|
||||
model = PPO(
|
||||
"MlpPolicy", vn,
|
||||
learning_rate=3e-4, n_steps=2048, batch_size=256, n_epochs=10,
|
||||
gamma=0.995, gae_lambda=0.95, clip_range=0.2,
|
||||
ent_coef=cfg.get("ent_coef", 0.02), vf_coef=0.5, max_grad_norm=0.5,
|
||||
policy_kwargs=dict(net_arch=[256, 256]),
|
||||
device="cpu",
|
||||
verbose=0,
|
||||
)
|
||||
|
||||
@@ -488,7 +316,7 @@ def main():
|
||||
print(f" reward/step: " + " ".join(
|
||||
f"{k}={v:+.4f}" for k, v in rps.items()))
|
||||
|
||||
# Episode visualization
|
||||
# Episode visualisation: trajectory + timeseries + animated GIF
|
||||
hist = run_and_record(model, vn, n, args.max_steps, rcfg,
|
||||
seed=1000 + n)
|
||||
tag = "success" if hist["success"] else "fail"
|
||||
@@ -498,6 +326,11 @@ def main():
|
||||
plot_timeseries(
|
||||
hist,
|
||||
os.path.join(eval_dir, f"ts_{n}s_{tag}.png"))
|
||||
if not args.no_gif:
|
||||
save_episode_gif(
|
||||
hist,
|
||||
os.path.join(eval_dir, f"ep_{n}s_{tag}.gif"),
|
||||
fps=args.gif_fps, skip=args.gif_skip)
|
||||
|
||||
r["n_sheep"] = n
|
||||
stage_results.append(r)
|
||||
|
||||
Reference in New Issue
Block a user