-
Notifications
You must be signed in to change notification settings - Fork 3
/
roc_cuda.py
149 lines (137 loc) · 5.9 KB
/
roc_cuda.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
from numba import cuda
import numpy as np
import os
from sklearn.preprocessing import normalize
import time
import multiprocessing as mp
from math import ceil
import sys
import atexit
from scipy.interpolate import interp1d
from ctypes import cdll
import argparse
@cuda.jit
def calc_ROC(feature, label, subfeature, sublabel, out):
i, j = cuda.grid(2)
if i < j and j < feature.shape[0] and i < sublabel.shape[0]:# and sublabel[i] * label[j] < 0:
tmp = 0.
for k in range(feature.shape[1]):
tmp += subfeature[i, k] * feature[j, k]
index_dis = int((tmp + 1) * 1000)
# if index_dis < 0 or index_dis > 2000:
# print("out range: " + str(index_dis))
if sublabel[i] == label[j]:
#if sublabel[i] + label[j] == 0:
cuda.atomic.add(out, 2*index_dis, 1)
else:
cuda.atomic.add(out, 2*index_dis+1, 1)
def gpu_job_producer(size, batch_size, target_size, _in_queue):
index = np.arange(size)
for i in range(ceil(target_size/batch_size)):
min_up = min((i+1)*batch_size, target_size)
_in_queue.put((index[i*batch_size : min_up], i*batch_size))
while True:
time.sleep(0.1)
def gpu_job_consumer(feature, label, _in_queue, _out_queue, device):
cuda.select_device(device)
blockdim = (32, 32)
while True:
index, start = _in_queue.get()
feature_cuda = cuda.to_device(feature[start:, :].astype(np.float32))
label_cuda = cuda.to_device(label[start:].astype(np.int32))
gridnum1 = (feature_cuda.shape[0] + blockdim[0] - 1) // blockdim[0]
gridnum2 = (len(index) + blockdim[0] - 1) // blockdim[0]
griddim = (gridnum2, gridnum1)
subfeature = cuda.to_device(feature[index, :].astype(np.float32))
sublabel = cuda.to_device(label[index].astype(np.int32))
out = cuda.to_device(np.zeros(2001*2, dtype=np.float64))
calc_ROC[griddim, blockdim](feature_cuda, label_cuda, subfeature, sublabel, out)
out_h = out.copy_to_host().astype(np.int64)
_out_queue.put(out_h)
def plot_ROC(data, output_dir,epoch, target_label):
data = np.cumsum(data, axis=0)
TPR = [1.0] #TPR = TP / (TP + FN)
FPR = [1.0] #FPR = FP / (FP + TN)
for i in range(data.shape[0]):
TPR.append((data[-1,0] - data[i, 0]) / data[-1,0])
FPR.append((data[-1,1] - data[i, 1]) / data[-1,1])
TPR = np.array(TPR)
FPR = np.array(FPR)
idx = np.argsort(FPR)
# with open(save_ROC_path, 'w') as pf:
# for i in range(TPR.shape[0]):
# pf.write("%f %f\n"%(FPR[i], TPR[i]))
ROC = interp1d(FPR[idx], TPR[idx])
result = [float('%.2f'%(100*ROC(10**(i)))) for i in range(-1,-7,-1)]
print('-'*80)
print('Target label from %d to %d'%(target_label[0],target_label[-1]))
print('Epoch %d, TPR (-1 to -6) = %r'%(epoch,result))
print('-'*80)
# plt.figure()
# plt.plot(FPR, TPR, label='ROC curve')
# plt.xlim([1e-7, 1e-2])
# plt.ylim([0.9, 1.0])
# plt.xscale('log')
# plt.xticks([1e-7, 1e-6, 1e-5, 1e-4, 1e-3, 1e-2])
# plt.grid(True)
# plt.savefig(save_ROC_path + ".png")
save_path = os.path.join(output_dir,'local_log.txt')
with open(save_path, "a") as pf:
pf.write('Target label from %d to %d\n'%(target_label[0],target_label[-1]))
pf.write('Epoch %d, TPR (-1 to -6) = %r\n'%(epoch,result))
class multiGPU(object):
def __init__(self, feature, label, batch_size, target_size, workers):
num = feature.shape[0]
self._manager = mp.Manager()
self._in_queue = self._manager.Queue(10)
self._producer = mp.Process(target=gpu_job_producer,
args=(num, batch_size, target_size, self._in_queue,))
self._producer.start()
self._out_queue = self._manager.Queue(10)
#gpu_job_consumer(feature, label, self._in_queue, self._out_queue, 0)
self._consumer = [mp.Process(target=gpu_job_consumer,
args=(feature, label, self._in_queue, self._out_queue, device))
for device in range(workers)]
for cons in self._consumer:
cons.start()
atexit.register(self._terminate)
def _terminate(self):
self._producer.terminate()
for cons in self._consumer:
cons.terminate()
if __name__ == '__main__':
parser = argparse.ArgumentParser('')
parser.add_argument('--feat_path',type=str)
parser.add_argument('--label_path',type=str)
parser.add_argument('--output_dir',type=str)
parser.add_argument('--workers',type=int,default=2)
parser.add_argument('--batch_size',type=int,default=800)
parser.add_argument('--ID_s_e',type=int,nargs='+')
parser.add_argument('--epoch',type=int,default=0)
args = parser.parse_args()
feature = np.load(args.feat_path).astype('float32')
label = np.load(args.label_path).astype('int32').reshape(-1)
workers = args.workers
batch_size = args.batch_size
start_ID,end_ID = args.ID_s_e
target_label = list(range(start_ID,end_ID))
t_idx = label==target_label[0]
for i in range(1,len(target_label)):
t_idx = t_idx | (label==target_label[i])
target_size = sum(t_idx)
feature = np.concatenate([feature[t_idx],feature[~t_idx]],axis=0)
label = np.concatenate([label[t_idx],label[~t_idx]])
proccer = multiGPU(feature, label, batch_size, target_size, workers)
out_sum = np.zeros(2001*2, dtype=np.int64)
start = time.time()
for _ in range(ceil(target_size / batch_size)):
out = proccer._out_queue.get()
out_sum += out
print('Total pair :',out_sum.sum())
# print('total use {:.2f}s'.format(time.time() - start))
cuda.profile_stop()
cuda.close()
out_sum = out_sum.reshape([-1, 2])
start = time.time()
plot_ROC(out_sum, args.output_dir,args.epoch,target_label)
# print('total use {:.2}s'.format(time.time() - start))