Why the evaluation result on IJB-B is different from paper&your test(almost same code)?

First of all, thank you for your excellent works, I learn a lot from your repository.
here is my puzzle：
my code is :
```
import os
import matplotlib
matplotlib.use('Agg')
from pathlib import Path
import pandas as pd
import numpy as np
from prettytable import PrettyTable

import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve, auc
from menpo.visualize.viewmatplotlib import sample_colours_from_colourmap

def read_template_pair_list(path):
# pairs = np.loadtxt(path, dtype=str)
pairs = pd.read_csv(path, sep=' ', header=None).values
# print(pairs.shape)
# print(pairs[:, 0].astype(np.int))
t1 = pairs[:, 0].astype(np.int)
t2 = pairs[:, 1].astype(np.int)
label = pairs[:, 2].astype(np.int)
return t1, t2, label

if __name__ == "__main__":

score_save_file = r"I:\Dataset\IJB_release\IJBB\result\MS1MV2-ResNet100-ArcFace-TestMode(N1D1F1).npy"
#MS1MV2-ResNet100-Arcface_IJBB_N1D1F1

image_path = "./IJBB"
pdf_name = "MS1MV2-ResNet100-ArcFace"
target = "IJBB"
p1, p2, label = read_template_pair_list(
os.path.join(r"I:\Dataset\IJB_release\IJBB\meta", #'%s/meta' % image_path,
'%s_template_pair_label.txt' % target.lower()))

files = [score_save_file]
methods = []
scores = []
for file in files:
methods.append(Path(file).stem)
scores.append(np.load(file))

methods = np.array(methods)
scores = dict(zip(methods, scores))
colours = dict(
zip(methods, sample_colours_from_colourmap(methods.shape[0], 'Set2')))
x_labels = [10 ** -6, 10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
#x_labels = [10 ** -5, 10 ** -4, 10 ** -3, 10 ** -2, 10 ** -1]
tpr_fpr_table = PrettyTable(['Methods'] + [str(x) for x in x_labels])
fig = plt.figure()
for method in methods:
fpr, tpr, _ = roc_curve(label, scores[method])
roc_auc = auc(fpr, tpr)
fpr = np.flipud(fpr)
tpr = np.flipud(tpr) # select largest tpr at same fpr
plt.plot(fpr,
tpr,
color=colours[method],
lw=1)
tpr_fpr_row = []
tpr_fpr_row.append("%s-%s" % (method, target))
for fpr_iter in np.arange(len(x_labels)):
_, min_index = min(
list(zip(abs(fpr - x_labels[fpr_iter]), range(len(fpr)))))
tpr_fpr_row.append('%.2f' % (tpr[min_index] * 100))
tpr_fpr_table.add_row(tpr_fpr_row)
plt.xlim([10 ** -6, 0.1])
plt.ylim([0.30, 1.0])
plt.grid(linestyle='--', linewidth=1)
plt.xticks(x_labels)
plt.yticks(np.linspace(0.30, 1.0, 8, endpoint=True))
plt.xscale('log')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('ROC on IJB')
plt.legend(loc="lower right")
fig.savefig(os.path.join("./", '%s.pdf' % pdf_name))
print(tpr_fpr_table)

```

trying to load your "MS1MV2-ResNet100-ArcFace-TestMode(N1D1F1).npy" and the label from ijbb_template_pair_label.txt to get the test result and draw the ROC curve.

But, I got result like this：
| Methods | 1e-06 | 1e-05 | 0.0001 | 0.001 | 0.01 | 0.1 |
|:------------------------------------------|---------:|---------:|---------:|---------:|---------:|---------:|
| MS1MV2-ResNet100-ArcFace-TestMode(N1D1F1) | 0.409056 | 0.908082 | 0.947712 | 0.963583 | 0.975463 | 0.986271 |

which is different from [Reformat and speed up IJB evaluation](https://github.com/deepinsight/insightface/pull/1349)：
| Methods | 1e-06 | 1e-05 | 0.0001 | 0.001 | 0.01 | 0.1 |
|:-------------------------------------|---------:|---------:|---------:|---------:|---------:|---------:|
| MS1MV2-ResNet100-Arcface_IJBB_N1D1F1 | 0.408861 | 0.899513 | 0.946349 | 0.964167 | 0.976144 | 0.98666 |

especially when FAR=1e-05，TAR increase almost 2%.
I'm confused, am I missing something?

BTW, when I use [ijb_11.py](https://github.com/deepinsight/insightface/blob/master/recognition/_evaluation_/ijb/ijb_11.py), the above mentioned problem still occurs.

hoping you can help me to figure it out, thank you in advance.