-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
149 lines (135 loc) · 4.72 KB
/
eval.py
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
# Third party import
from argparse import ArgumentParser
import yaml
import torch
import pandas as pd
from tqdm import tqdm
import numpy as np
from sklearn.metrics import roc_auc_score, roc_curve, RocCurveDisplay
import matplotlib.pyplot as plt
import matplotlib
# Local import
from kd4jets.models import MLPKD, DeepSetKD
from kd4jets import boost_jets, boost_batch
matplotlib.use("agg")
def parse_arguments():
parser = ArgumentParser()
parser.add_argument("--cfg", type = str, required = True)
args = parser.parse_args()
return args
def tocuda(x):
if isinstance(x, list):
return [y.cuda() for y in x]
else:
return x.cuda()
def plot_roc(name, df, ax):
new_df = df[df["beta"] == 0]
fpr, tpr, _ = roc_curve(np.array(new_df["labels"]), np.array(new_df["prob"]))
fpr, tpr = fpr[(tpr > 0) & (fpr > 0)], tpr[(tpr > 0) & (fpr > 0)]
ax.plot(tpr, 1/fpr, label = name, ms = 1)
def plot_boost(name, df, ax):
df["acc"] = (df["pred"] == df["labels"])
acc = df.groupby("beta").mean()
ax.plot(acc.index, acc["acc"], ms = 1, ls="-", label=name)
@torch.no_grad()
def get_predictions(model):
loader = model.test_dataloader()
dfs = []
model.eval()
for batch in tqdm(loader):
for name in batch:
batch[name] = tocuda(batch[name])
for beta in np.linspace(0, 1, 20, endpoint=False):
batched_beta = torch.tensor([beta]*model.hparams["batch_size"])
logit, latent_rep = model.student(boost_batch(batch, batched_beta))
new_df = pd.DataFrame({
"beta": beta,
"pred": logit.cpu().argmax(-1).numpy(),
"prob": torch.sigmoid(logit[:, 1] - logit[:, 0]).cpu().numpy(),
"labels": batch["is_signal"].cpu().float().numpy(),
})
dfs.append(new_df)
df = pd.concat(dfs, ignore_index = True)
return df
def main():
args = parse_arguments()
with open(args.cfg, 'r') as f:
cfg = yaml.safe_load(f)
CMS = {
"font.family": "sans-serif",
"mathtext.fontset": "custom",
"mathtext.rm": "TeX Gyre Heros",
"mathtext.bf": "TeX Gyre Heros:bold",
"mathtext.sf": "TeX Gyre Heros",
"mathtext.it": "TeX Gyre Heros:italic",
"mathtext.tt": "TeX Gyre Heros",
"mathtext.cal": "TeX Gyre Heros",
"mathtext.default": "regular",
"figure.figsize": (10.0, 10.0),
"font.size": 26,
"axes.labelsize": "medium",
"axes.unicode_minus": False,
"xtick.labelsize": "small",
"ytick.labelsize": "small",
"legend.fontsize": "small",
"legend.handlelength": 1.5,
"legend.borderpad": 0.5,
"xtick.direction": "in",
"xtick.major.size": 12,
"xtick.minor.size": 6,
"xtick.major.pad": 6,
"xtick.top": True,
"xtick.major.top": True,
"xtick.major.bottom": True,
"xtick.minor.top": True,
"xtick.minor.bottom": True,
"xtick.minor.visible": True,
"ytick.direction": "in",
"ytick.major.size": 12,
"ytick.minor.size": 6.0,
"ytick.right": True,
"ytick.major.left": True,
"ytick.major.right": True,
"ytick.minor.left": True,
"ytick.minor.right": True,
"ytick.minor.visible": True,
"grid.alpha": 0.8,
"grid.linestyle": ":",
"axes.linewidth": 2,
"savefig.transparent": False,
}
plt.style.use(CMS)
pred_dfs = {}
for name, ckpt_cfg in cfg["models"].items():
if ckpt_cfg["class"] == "MLP":
model = MLPKD.load_from_checkpoint(ckpt_cfg["ckpt"]).cuda()
elif ckpt_cfg["class"] == "DeepSet":
model = DeepSetKD.load_from_checkpoint(ckpt_cfg["ckpt"]).cuda()
else:
raise NotImplementedError(f"model {name} is not implemented")
pred_dfs[name] = get_predictions(model)
fig, (ax) = plt.subplots(figsize = (8, 8), ncols = 1, nrows = 1)
for name, df in pred_dfs.items():
plot_roc(name, df, ax)
ax.legend()
ax.grid(True, which="both", ls="--", color='0.65')
ax.set_yscale("log")
ax.set_ylim([1, 3e4])
ax.set_xlabel(r"Signal efficiency $\epsilon_s$")
ax.set_ylabel(r"Background rejection $1/\epsilon_b$")
fig.tight_layout()
fig.savefig("plots/roc.pdf")
plt.close(fig)
fig, (ax) = plt.subplots(figsize = (12, 8), ncols = 1, nrows = 1)
for name, df in pred_dfs.items():
plot_boost(name, df, ax)
ax.legend(fontsize = 'x-small')
ax.grid(True, ls="--", color='0.65')
ax.set_ylim([0.5, 1])
ax.set_xlabel(r"$\beta = v/c$")
ax.set_ylabel("Accuracy")
fig.tight_layout()
fig.savefig("plots/boost.pdf")
plt.close(fig)
if __name__ == "__main__":
main()