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
|
#!/usr/bin/env python3
"""H8: Generate Figure 4(a)-style depth sweep plot.
4 panels (Cora/CiteSeer/PubMed/DBLP), 3 curves per panel (BP/DFA-GNN/KAFT).
x = number of layers L; y = test accuracy (%) with shaded std band.
Method distinguished by color only (per memory `feedback_viz_shape`:
shape encodes sweep axis — here L is the x-axis, so same marker for all methods).
"""
import json
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.colors import to_rgba
DATASETS = ['Cora', 'CiteSeer', 'PubMed', 'DBLP']
METHODS = ['BP', 'DFA-GNN', 'KAFT']
# Per-dataset depth grids — DBLP extends to 24, 32 from dblp_depth_scaling.
# Other datasets cover 2..20. Missing entries (e.g. DFA-GNN at L=2/3, DBLP L=10
# for BP/KAFT) will be silently skipped by lookup().
DEPTHS_DEFAULT = [2, 3, 4, 6, 8, 10, 12, 14, 16, 18, 20]
DEPTHS_DBLP = [2, 3, 4, 6, 8, 10, 12, 14, 16, 18, 20, 24, 32]
DEPTHS_BY_DS = {ds: (DEPTHS_DBLP if ds == 'DBLP' else DEPTHS_DEFAULT)
for ds in DATASETS}
# All result files we might need to consult
SOURCES = [
'results/combo_20seeds/per_seed_data.json', # L=6 BP/KAFT/stacks on Cora/CS/DBLP
'results/hero_extras_20seeds/per_seed_data.json', # L=6 on PubMed + Coauthor
'results/shallow_depth_20seeds/per_seed_data.json', # L=2,3,4 on 4ds
'results/dblp_depth_scaling_20seeds/per_seed_data.json', # DBLP L=8-32
'results/bp_kaft_depth_20seeds/per_seed_data.json', # Cora/CS/PubMed L=8-20
'results/dfagnn_depth_20seeds/per_seed_data.json', # DFA-GNN at all depths
'results/dfagnn_resgcn_20seeds/per_seed_data.json', # DFA-GNN L=6 Cora/CS/DBLP
'results/depth_extras_20seeds/per_seed_data.json', # L=14, L=18 × 4ds × 3 methods
]
# Colors — KAFT brick red (main method), BP gray, DFA-GNN complementary blue
COLORS = {
'BP': '#888888', # reference gray
'DFA-GNN': '#3B7AC2', # complementary blue
'KAFT': '#C23B3B', # brick red (our method)
}
GRID_COLOR = '#ECEFF3'
TEXT_COLOR = '#2F3437'
def load_all():
"""Load all sources into a single dict keyed by original keys."""
merged = {}
for path in SOURCES:
try:
with open(f'/home/yurenh2/graph-grape/{path}') as f:
d = json.load(f)
for k, v in d.items():
if k not in merged:
merged[k] = v
else:
# Merge seed dicts (take first available if conflict)
for sk, sv in v.items():
if sk not in merged[k]:
merged[k][sk] = sv
except FileNotFoundError:
pass
return merged
def lookup(data, ds, L, method):
"""Return (mean, std) or None if unavailable."""
# Try multiple key formats
# 1. {ds}_L{L}_{method} (depth-indexed)
# 2. {ds}_{method} (for L=6, assumed default in combo/hero files)
for key in [f'{ds}_L{L}_{method}', f'{ds}_{method}' if L == 6 else None]:
if key and key in data:
seeds = data[key]
if len(seeds) >= 15: # allow a few missing seeds
vals = np.array(list(seeds.values())) * 100
return vals.mean(), vals.std()
return None
def main():
data = load_all()
plt.rcParams.update({
'font.size': 10,
'axes.labelsize': 10,
'xtick.labelsize': 9,
'ytick.labelsize': 9,
'legend.fontsize': 9,
'pdf.fonttype': 42,
'ps.fonttype': 42,
})
fig, axes = plt.subplots(1, 4, figsize=(13.0, 3.3), sharey=False)
legend_handles = {}
for ax, ds in zip(axes, DATASETS):
depths = DEPTHS_BY_DS[ds]
for method in METHODS:
xs, means, stds = [], [], []
for L in depths:
r = lookup(data, ds, L, method)
if r is not None:
xs.append(L)
means.append(r[0])
stds.append(r[1])
if not xs:
continue
xs = np.array(xs); means = np.array(means); stds = np.array(stds)
color = COLORS[method]
line, = ax.plot(xs, means, marker='o', markersize=5,
color=color, linewidth=1.6,
markerfacecolor=to_rgba(color, alpha=0.35),
markeredgecolor=color, markeredgewidth=0.8,
zorder=3)
ax.fill_between(xs, means - stds, means + stds,
color=color, alpha=0.12, edgecolor='none', zorder=2)
if method not in legend_handles:
legend_handles[method] = line
ax.set_title(ds, fontsize=10, color=TEXT_COLOR, pad=6)
ax.set_xlabel('Number of layers $L$', fontsize=9, color=TEXT_COLOR)
ax.grid(axis='both', color=GRID_COLOR, linewidth=0.7)
ax.set_axisbelow(True)
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.spines['left'].set_color('#C9CDD3')
ax.spines['bottom'].set_color('#C9CDD3')
ax.tick_params(colors=TEXT_COLOR)
# Show every other tick for readability when grid is dense
ticks = depths if len(depths) <= 8 else depths[::2]
ax.set_xticks(ticks)
axes[0].set_ylabel('Test accuracy (%)', fontsize=10, color=TEXT_COLOR)
handles = [legend_handles[m] for m in METHODS if m in legend_handles]
labels = [m for m in METHODS if m in legend_handles]
fig.tight_layout(rect=(0.0, 0.06, 1.0, 1.0), w_pad=1.5)
fig.legend(handles, labels,
frameon=False, loc='lower center',
ncol=len(labels), bbox_to_anchor=(0.5, -0.005),
handletextpad=0.6, columnspacing=1.8)
fig.savefig('/home/yurenh2/graph-grape/kaft_depth_sweep.png', dpi=300, bbox_inches='tight')
fig.savefig('/home/yurenh2/graph-grape/kaft_depth_sweep.pdf', bbox_inches='tight')
plt.close(fig)
print('Saved /home/yurenh2/graph-grape/kaft_depth_sweep.{png,pdf}')
# Data dump
print('\nData (mean ± std):')
for ds in DATASETS:
print(f'\n{ds}:')
depths = DEPTHS_BY_DS[ds]
for method in METHODS:
row = [f'{method:<9}']
for L in depths:
r = lookup(data, ds, L, method)
row.append(f'L{L}: {r[0]:5.1f}±{r[1]:4.1f}' if r else f'L{L}: {"—":>10}')
print(' ' + ' '.join(row))
if __name__ == '__main__':
main()
|
