File size: 6,515 Bytes
aab061c | 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 | # === Global Variables ===
BASE_FONTSIZE = 16
SMALL_FONTSIZE = 14
MEDIUM_FONTSIZE = 15
LARGE_FONTSIZE = 20
LEGEND_FONTSIZE = 17
ANNOTATION_FONTSIZE = 15
LINEWIDTH = 6
AXIS_LINEWIDTH = 1.0
GRID_LINEWIDTH = 0.5
HLINE_LINEWIDTH = AXIS_LINEWIDTH * 2
MARKERSIZE = 10
MARKEREDGEWIDTH = AXIS_LINEWIDTH * 2
# =========================
import csv
import os
import re
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
from matplotlib.ticker import NullFormatter
mpl.rcParams["font.family"] = "serif"
mpl.rcParams["font.serif"] = ["Times New Roman", "DejaVu Serif"]
mpl.rcParams["mathtext.fontset"] = "cm"
mpl.rcParams["axes.linewidth"] = AXIS_LINEWIDTH
mpl.rcParams["axes.labelsize"] = BASE_FONTSIZE
mpl.rcParams["xtick.labelsize"] = MEDIUM_FONTSIZE
mpl.rcParams["ytick.labelsize"] = MEDIUM_FONTSIZE
mpl.rcParams["legend.fontsize"] = LEGEND_FONTSIZE
mpl.rcParams["figure.dpi"] = 100
mpl.rcParams["savefig.dpi"] = 300
EXP_ROOT = os.path.dirname(os.path.abspath(__file__))
blue_color = "#3498db"
red_color = "#e74c3c"
colors = [blue_color, red_color]
# (exp_dir_name): (ckpt, optimal_gh_string, kind)
# kind is "LDM-only" (variational, separate VAE) or "Joint" (e2e training).
EXPS = {
"sit-xl-1-f16d32-e2e-novariational-logvar0.0": (
"0250000", "0.35", "Joint",
),
"sit-xl-1-f16d64-e2e-novariational-logvar0.0": (
"0250000", "0.55", "Joint",
),
"sit-xl-1-f16d32-ldm-imagenet256-f16d32-50e-250k": (
"0250000", "0.75", "LDM-only",
),
"sit-xl-1-f16d64-ldm-imagenet256-f16d64-50e-250k": (
"0250000", "0.825", "LDM-only",
),
"sit-xl-1-dinov2-b-enc8-repae-f16d32-scratch-0.5-0.5-400k": (
"0400000", "0.8", "LDM-only",
),
"sit-xl-1-dinov2-f16d32-e2e-novariational-logvar0.0-repae-coeff0.5": (
"0400000", "0.525", "Joint",
),
}
PAIRS = [
("repa", [
"sit-xl-1-dinov2-b-enc8-repae-f16d32-scratch-0.5-0.5-400k",
"sit-xl-1-dinov2-f16d32-e2e-novariational-logvar0.0-repae-coeff0.5",
], {1}, " + REPA"),
("f16d32", [
"sit-xl-1-f16d32-e2e-novariational-logvar0.0",
"sit-xl-1-f16d32-ldm-imagenet256-f16d32-50e-250k",
], set(), ""),
("f16d64", [
"sit-xl-1-f16d64-e2e-novariational-logvar0.0",
"sit-xl-1-f16d64-ldm-imagenet256-f16d64-50e-250k",
], set(), ""),
]
def load_curve(exp):
"""Return (cfgs, fids) sorted by cfg, restricted to the per-exp optimal gh."""
ckpt, gh, _label = EXPS[exp]
folder = os.path.join(EXP_ROOT, exp)
pat = re.compile(
rf"{re.escape(exp)}_{ckpt}_cfg(\d+(?:\.\d+)?)-0\.0-{re.escape(gh)}-labelsampling-equal\.csv$"
)
rows = []
for fn in os.listdir(folder):
m = pat.match(fn)
if not m:
continue
cfg = float(m.group(1))
with open(os.path.join(folder, fn)) as f:
r = list(csv.DictReader(f))
if r and "FID" in r[0]:
rows.append((cfg, float(r[0]["FID"])))
rows.sort()
if not rows:
return [], []
cfgs, fids = zip(*rows)
return list(cfgs), list(fids)
def plot_pair(pair_name, exp_list, output_dir, flip_below=frozenset(), label_suffix=""):
fig, ax = plt.subplots(figsize=(7, 6), dpi=300)
all_fids = []
for i, exp in enumerate(exp_list):
_ckpt, gh, kind = EXPS[exp]
label = f"{kind}{label_suffix} (Intv. [0.0, {gh}])"
cfgs, fids = load_curve(exp)
if not cfgs:
print(f"[Warn] No curve points for {exp} at gh={gh}")
continue
all_fids.extend(fids)
ax.plot(
cfgs, fids,
label=label,
marker="o",
markersize=MARKERSIZE,
linewidth=LINEWIDTH,
linestyle="-",
color=colors[i],
markerfacecolor="white",
markeredgecolor=colors[i],
alpha=0.9,
markeredgewidth=MARKEREDGEWIDTH,
)
# Mark + annotate the minimum (best CFG point) with a horizontal reference.
min_idx = int(np.argmin(fids))
best_cfg, best_fid = cfgs[min_idx], fids[min_idx]
ax.axhline(
y=best_fid,
color=colors[i],
linestyle="--",
linewidth=HLINE_LINEWIDTH,
alpha=0.7,
)
if i in flip_below:
ax.text(
best_cfg, best_fid * 0.985,
f"{best_fid:.2f}",
ha="center", va="top",
fontsize=ANNOTATION_FONTSIZE, color=colors[i],
zorder=5,
)
else:
ax.text(
best_cfg, best_fid * 1.02,
f"{best_fid:.2f}",
ha="center", va="bottom",
fontsize=ANNOTATION_FONTSIZE, color=colors[i],
zorder=5,
)
if not all_fids:
plt.close(fig)
print(f"[Warn] Skipping {pair_name}: no data.")
return
ax.legend(
fontsize=LEGEND_FONTSIZE,
framealpha=0.95,
loc="best",
frameon=False,
)
ax.grid(True, alpha=0.3, linestyle="--", linewidth=GRID_LINEWIDTH)
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
ax.set_xlabel("CFG Scale", fontsize=SMALL_FONTSIZE)
ax.set_ylabel("⟵ FID Score", fontsize=LARGE_FONTSIZE)
# Auto: log scale when the dynamic range is large.
fid_min, fid_max = float(np.nanmin(all_fids)), float(np.nanmax(all_fids))
if fid_max / max(fid_min, 1e-6) > 3:
ax.set_yscale("log")
tick_candidates = np.array([1.5, 2, 3, 4, 5, 6, 8, 10, 15, 20, 25, 30, 40, 50])
yticks = tick_candidates[
(tick_candidates >= fid_min * 0.9) & (tick_candidates <= fid_max * 1.1)
]
if len(yticks):
ax.set_yticks(yticks)
ax.set_yticklabels([f"{t:g}" for t in yticks])
ax.yaxis.set_minor_locator(plt.NullLocator())
ax.yaxis.set_minor_formatter(NullFormatter())
plt.tight_layout()
os.makedirs(output_dir, exist_ok=True)
out_png = os.path.join(output_dir, f"{pair_name}.png")
out_pdf = os.path.join(output_dir, f"{pair_name}.pdf")
plt.savefig(out_png)
plt.savefig(out_pdf)
plt.close(fig)
print(f"Wrote {out_png} and {out_pdf}")
if __name__ == "__main__":
output_dir = os.path.join(EXP_ROOT, "cfg_plots")
for pair_name, exp_list, flip_below, label_suffix in PAIRS:
plot_pair(pair_name, exp_list, output_dir, flip_below=flip_below, label_suffix=label_suffix)
|