File size: 4,910 Bytes
78be280 | 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 | import json
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
RUNS = Path("/workspace/runs")
TMAX_VALUES = [1, 2, 6, 10, 15]
STRATEGY = "strategy_3"
def run_dir(tmax: int) -> Path:
return (
RUNS
/ f"Segformer_B0_simplified_tmax_{tmax}"
/ "repeated_holdout"
/ "stratified_holdout_v1"
/ "phase_001"
/ "pct_100"
/ "repeat_01"
/ STRATEGY
/ "final"
)
def load_json(p: Path):
with open(p) as f:
return json.load(f)
records = []
for tmax in TMAX_VALUES:
d = run_dir(tmax)
evaluation = load_json(d / "evaluation.json")
summary = load_json(d / "summary.json")
metrics = evaluation["metrics"]
timing = evaluation["timing"]
records.append(
{
"tmax": tmax,
# Segmentation quality on the held-out test set (mean +/- std)
"biou_contour_mean": metrics["biou_contour"]["mean"],
"biou_contour_std": metrics["biou_contour"]["std"],
"dice_mean": metrics["dice"]["mean"],
"dice_std": metrics["dice"]["std"],
"iou_mean": metrics["iou"]["mean"],
"iou_std": metrics["iou"]["std"],
# Training cost
"train_total_s": summary["elapsed_seconds"],
"train_to_best_s": summary["time_to_best_seconds"],
"sec_per_epoch": summary["seconds_per_epoch_measured_mean"],
"sec_per_epoch_std": summary["seconds_per_epoch_measured_std"],
"best_epoch": summary["best_epoch"],
# Inference cost (per image, test set)
"infer_ms_mean": timing["mean_per_image_inference_ms"],
"infer_ms_std": timing["std_per_image_inference_ms"],
}
)
tmax = [r["tmax"] for r in records]
print(f"{'tmax':>5} {'biou_contour':>18} {'dice':>18} {'iou':>18} {'train_total_s':>14} {'sec/epoch':>12} {'best_ep':>8} {'infer_ms/img':>16}")
for r in records:
print(
f"{r['tmax']:>5} "
f"{r['biou_contour_mean']:.4f}+/-{r['biou_contour_std']:.4f} "
f"{r['dice_mean']:.4f}+/-{r['dice_std']:.4f} "
f"{r['iou_mean']:.4f}+/-{r['iou_std']:.4f} "
f"{r['train_total_s']:>14.1f} "
f"{r['sec_per_epoch']:>12.3f} "
f"{r['best_epoch']:>8} "
f"{r['infer_ms_mean']:>10.3f}+/-{r['infer_ms_std']:.3f}"
)
import numpy as np
def add_labels(ax, bars, fmt="{:.2f}"):
for b in bars:
h = b.get_height()
ax.annotate(
fmt.format(h),
xy=(b.get_x() + b.get_width() / 2, h),
xytext=(0, 3),
textcoords="offset points",
ha="center",
va="bottom",
fontsize=8,
)
x = np.arange(len(tmax))
labels = [str(t) for t in tmax]
fig, axes = plt.subplots(1, 3, figsize=(17, 5.5))
fig.suptitle(
"Segformer-B0 simplified (strategy 3, phase 1) - metrics vs T_max",
fontsize=14,
fontweight="bold",
)
# ---- 1) Segmentation quality (test set) : grouped bars ----
ax = axes[0]
w = 0.27
b1 = ax.bar(x - w, [r["dice_mean"] for r in records], w, yerr=[r["dice_std"] for r in records], capsize=3, color="tab:blue", label="Dice")
b2 = ax.bar(x, [r["iou_mean"] for r in records], w, yerr=[r["iou_std"] for r in records], capsize=3, color="tab:green", label="IoU")
b3 = ax.bar(x + w, [r["biou_contour_mean"] for r in records], w, yerr=[r["biou_contour_std"] for r in records], capsize=3, color="tab:orange", label="bIoU contour")
add_labels(ax, b1)
add_labels(ax, b2)
add_labels(ax, b3)
ax.set_title("Segmentation quality (test set, mean +/- std)")
ax.set_xlabel("T_max")
ax.set_ylabel("score")
ax.set_xticks(x)
ax.set_xticklabels(labels)
ax.set_ylim(0, 1.05)
ax.grid(True, axis="y", alpha=0.3)
ax.legend()
# ---- 2) Training time : grouped bars ----
ax = axes[1]
w = 0.4
b1 = ax.bar(x - w / 2, [r["train_total_s"] for r in records], w, color="tab:red", label="total training time")
b2 = ax.bar(x + w / 2, [r["train_to_best_s"] for r in records], w, color="tab:orange", label="time to best checkpoint")
add_labels(ax, b1, fmt="{:.0f}")
add_labels(ax, b2, fmt="{:.0f}")
ax.set_title("Training time")
ax.set_xlabel("T_max")
ax.set_ylabel("seconds")
ax.set_xticks(x)
ax.set_xticklabels(labels)
ax.grid(True, axis="y", alpha=0.3)
ax.legend()
# ---- 3) Inference time (per image) : bars ----
ax = axes[2]
b1 = ax.bar(x, [r["infer_ms_mean"] for r in records], 0.6, yerr=[r["infer_ms_std"] for r in records], capsize=4, color="tab:purple", label="per-image inference (mean +/- std)")
add_labels(ax, b1)
ax.set_title("Inference time (test set)")
ax.set_xlabel("T_max")
ax.set_ylabel("ms / image")
ax.set_xticks(x)
ax.set_xticklabels(labels)
ax.grid(True, axis="y", alpha=0.3)
ax.legend()
fig.tight_layout(rect=[0, 0, 1, 0.96])
out = RUNS.parent / "tmax_biou_training_inference.png"
fig.savefig(out, dpi=150)
print(f"\nSaved figure to {out}")
|