-
Notifications
You must be signed in to change notification settings - Fork 6
/
train_classification_net.py
executable file
·161 lines (126 loc) · 5.48 KB
/
train_classification_net.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
150
151
152
153
154
155
156
157
158
159
160
#!/usr/bin/python
import numpy as np
import os
import sys
import h5py
import matplotlib.pyplot as plt
caffe_root = os.getenv('CAFFE_ROOT', './')
sys.path.insert(0, caffe_root + '/python')
import caffe
TRAIN_DATASET_DIR = os.path.join(os.getcwd(), './train_dataset_classification_500K_100batch')
TEST_DATASET_DIR = os.path.join(os.getcwd(), './test_dataset_classification_200K_100batch')
BATCH_FILENAME_FORMAT = 'dataset_batch%d.hdf5'
DATASET_MEAN_FILENAME = 'dataset_mean.hdf5'
SOLVER_PROTO_FILENAME = 'lenet_classification_solver.prototxt'
MODEL_FILENAME = 'balltracker_classification.caffemodel'
BATCH_SIZE = 100
NUM_TRAINING_ITERATIONS = 5000
TEST_INTERVAL = 200
TEST_NUM_SAMPLES = 10000
FIRST_LAYER = 'conv1'
def loadBatch(datasetDir, batch_size, n, mean = None):
data_arr = np.zeros((batch_size, 1, 100, 100))
label_arr = np.zeros((batch_size))
hdf5 = os.path.join(datasetDir, BATCH_FILENAME_FORMAT % n)
f = h5py.File(hdf5, "r")
images = f.keys()
# load arrays from HDF5 file
for idx, i in enumerate(images):
if idx < batch_size:
data_arr[idx, 0, ...] = f[i][...]
label_arr[idx] = np.int32(f[i].attrs['HAS_SPHERE'])
data_arr /= 256.0 # normalize to [0, 1)
# subtract mean
if mean is not None:
data_arr[:, 0, ...] -= mean
f.close()
return data_arr, label_arr
# setup solver
caffe.set_device(0)
caffe.set_mode_gpu()
solver = None
solver = caffe.SGDSolver(SOLVER_PROTO_FILENAME)
train_loss = np.zeros(NUM_TRAINING_ITERATIONS)
train_loss_averaged = []
test_acc = np.zeros(int(np.ceil(NUM_TRAINING_ITERATIONS / TEST_INTERVAL) + 1))
output = np.zeros((NUM_TRAINING_ITERATIONS, 8, 2))
testData, testLabel = loadBatch(TEST_DATASET_DIR, BATCH_SIZE, 0)
# auroc statistics
auroc_thresholds = np.linspace(0, 1, BATCH_SIZE)
auroc_stats = np.zeros((len(auroc_thresholds), 4), dtype='uint32') # [ TP, FN, FP, TN ]
# test the trained net over test_dataset
def testNet(i, test_dataset_dir, solver, numTestSamples, auroc = False):
print ('Iteration %d testing...' % i)
correct = 0
global auroc_thresholds
global auroc_stats
auroc_stats = np.zeros((len(auroc_thresholds), 4), dtype='uint32')
for test_it in range(numTestSamples / BATCH_SIZE):
# load new test batch
d, l = loadBatch(test_dataset_dir, BATCH_SIZE, test_it)
solver.test_nets[0].blobs['data'].data[...] = d
solver.test_nets[0].blobs['label'].data[...] = l
solver.test_nets[0].forward(start=FIRST_LAYER)
correct += sum(solver.test_nets[0].blobs['prob'].data.argmax(1) == solver.test_nets[0].blobs['label'].data)
test_acc[i // TEST_INTERVAL] = float(correct) / numTestSamples * 100.0
if auroc:
for p in range(BATCH_SIZE):
for idx, threshold in enumerate(auroc_thresholds):
label = int(solver.test_nets[0].blobs['label'].data[p])
predicted = solver.test_nets[0].blobs['prob'].data[p][1]
if label == 1 and (predicted > threshold): # true positive
auroc_stats[idx][0] += 1
elif label == 1 and (predicted < threshold): # false negative
auroc_stats[idx][1] += 1
elif label == 0 and (predicted > threshold): # false positive
auroc_stats[idx][2] += 1
elif label == 0 and (predicted < threshold): # true negative
auroc_stats[idx][3] += 1
print('Test accuracy: %f' % test_acc[i // TEST_INTERVAL])
# load mean
meanHdf = h5py.File(os.path.join(TRAIN_DATASET_DIR, DATASET_MEAN_FILENAME), 'r')
#mean = np.zeros(meanHdf['mean'][...].shape, meanHdf['mean'][...].dtype)
#mean[...] = meanHdf['mean'][...]
mean = np.zeros(meanHdf)
meanHdf.close()
moving_window = NUM_TRAINING_ITERATIONS / 50
tmp_loss_average = 0.0
# main training loop
for it in range(NUM_TRAINING_ITERATIONS):
# load batch
trainData, trainLabel = loadBatch(TRAIN_DATASET_DIR, BATCH_SIZE, it, mean)
solver.net.blobs['data'].data[...] = trainData
solver.net.blobs['label'].data[...] = trainLabel
# update weights
solver.step(1)
train_loss[it] = solver.net.blobs['loss'].data
# compute moving average of training loss
tmp_loss_average += train_loss[it] / moving_window
if (it + 1) % moving_window == 0:
train_loss_averaged.append(tmp_loss_average)
tmp_loss_average = 0.0
if it % TEST_INTERVAL == 0:
testNet(it, TEST_DATASET_DIR, solver, TEST_NUM_SAMPLES)
# last test
testNet(NUM_TRAINING_ITERATIONS, TEST_DATASET_DIR, solver, TEST_NUM_SAMPLES, True)
# save model
solver.net.save(MODEL_FILENAME)
# plot auroc
fpr = ((auroc_stats[:, 2]).astype(float) / (auroc_stats[:, 2] + auroc_stats[:, 3]))
tpr = ((auroc_stats[:, 0]).astype(float) / (auroc_stats[:, 0] + auroc_stats[:, 1]))
z = np.linspace(min(fpr), max(fpr))
plt.plot(z, z, '--')
plt.plot(fpr, tpr, 'r')
plt.fill_between(fpr, tpr, 0, color='blue', alpha=0.3)
plt.show()
# plot trainloss and test accuracy over iterations
_, ax1 = plt.subplots()
ax2 = ax1.twinx()
ax1.plot(np.arange(NUM_TRAINING_ITERATIONS), train_loss)
ax1.plot(np.arange(moving_window / 2, NUM_TRAINING_ITERATIONS + moving_window / 2, moving_window), train_loss_averaged, 'y', linewidth=2)
ax2.plot(TEST_INTERVAL * np.arange(len(test_acc)), test_acc, 'r')
ax1.set_xlabel('iteration')
ax1.set_ylabel('train loss')
ax2.set_ylabel('test accuracy')
ax2.set_title('Test Accuracy: {:.2f}'.format(test_acc[-1]))
plt.show()