def analysis_position(subset=None):
# get test files and labels
txtfile = './split/{}/test.txt'.format(configer.splitmode)
with open(txtfile, 'r') as f:
testfiles = f.readlines()
testfiles_sub, indexes_sub = get_sub_filenames(testfiles, subset)
labels = get_labels_from_pathlist(testfiles_sub)
labels = list(map(lambda x: x - 1, labels))
# get test outputs
npyfile = os.path.join(configer.logspath, configer.modelname,
'test_out.npy')
testout = np.load(npyfile)
for i in range(testout.shape[0]):
testout[i] = softmax(testout[i])
testout_sub = testout[indexes_sub] # ndarray(N, cls)
testout_sub = np.argmax(testout_sub, axis=1)
n_position = np.zeros(shape=7) # 统计各角度的样本数量
n_accuracy = np.zeros(shape=7) # 统计各角度的正确数量
for i in range(len(testfiles_sub)):
testfile = testfiles_sub[i]
y_true = labels[i]
y_pred = testout_sub[i]
posidx = getPos(testfile) - 1
n_position[posidx] += 1
if y_pred == y_true: n_accuracy[posidx] += 1
acc = n_accuracy / n_position
print(list(acc))
def analysis_position(subset=None):
# get test files and labels
txtfile = './dataset/{}/test.txt'.format(configer.splitmode)
with open(txtfile, 'r') as f:
testfiles = f.readlines()
testfiles_sub, indexes_sub = get_sub_filenames(testfiles, subset)
labels = get_labels_from_pathlist(testfiles_sub)
used = [i for i in range(1, 34) if (i not in notUsedSubjects)]
labels = [used.index(i) for i in labels]
# get test outputs
npyfile = os.path.join(configer.logspath, configer.modelname,
'test_output_{}.npy'.format(configer.splitmode))
testout = np.load(npyfile)
for i in range(testout.shape[0]):
testout[i] = softmax(testout[i])
testout_sub = testout[indexes_sub] # ndarray(N, cls)
testout_sub = np.argmax(testout_sub, axis=1)
n_position = np.zeros(shape=7) # 统计各角度的样本数量
n_accuracy = np.zeros(shape=7) # 统计各角度的正确数量
for i in range(len(testfiles_sub)):
testfile = testfiles_sub[i]
y_true = labels[i]
y_pred = testout_sub[i]
posidx = int(testfile.split('/')[-2].split('_')[1]) - 1
n_position[posidx] += 1
if y_pred == y_true: n_accuracy[posidx] += 1
acc = n_accuracy / n_position
print(list(acc))
def analysis(subset=None):
# get test files and labels
txtfile = './split/{}/test.txt'.format(configer.splitmode)
with open(txtfile, 'r') as f:
testfiles = f.readlines()
testfiles_sub, indexes_sub = get_sub_filenames(testfiles, subset)
labels = get_labels_from_pathlist(testfiles_sub)
labels = [i - 1 for i in labels]
# get test outputs
npyfile = os.path.join(configer.logspath, configer.modelname,
'test_out.npy')
testout = np.load(npyfile)
for i in range(testout.shape[0]):
testout[i] = softmax(testout[i])
testout_sub = testout[indexes_sub] # ndarray(N, cls)
N, cls = testout_sub.shape
C = 23
WAVELEN = [550 + i * 20 for i in range(C)]
## 20190226
# N, cls = testout_sub.shape
# C = 46
# testout_sub_ch = np.zeros(shape=(N, C, cls))
# WaveLen = [550 + 10*i for i in range(C)]
# for c in range(C):
# for n in range(N):
# if (WaveLen[c] == getWavelen(testfiles_sub[n])):
# testout_sub_ch[n, c] = testout_sub[n]
## 20190227 - AUC
from sklearn.metrics import roc_auc_score
from sklearn.preprocessing import OneHotEncoder
auc_scores = np.zeros(C)
acc_scores = np.zeros(C)
for c in range(C):
y_true = []
y_pred_prob = []
for n in range(N):
testfile = testfiles_sub[n]
testout = list(testout_sub[n])
idxch = WAVELEN.index(getWavelen(testfile))
if idxch == c: # 对应波长
y_true += [getLabel(testfile) - 1]
y_pred_prob += [testout]
y_true = np.array(y_true)
y_pred_prob = np.array(y_pred_prob)
equ = (y_true == np.argmax(y_pred_prob,
axis=1)).astype('int') # 正确记1，错误记0
y_pred_prob_bin = np.zeros(shape=y_true.shape[0])
for i in range(y_true.shape[0]):
prob = y_pred_prob[i, y_true[i]]
y_pred_prob_bin[i] = prob if equ[i] else (1 - prob
) # 统计判断正确的概率，即从多分类变为二分类
# AUC
if len(set(equ)) == 1:
auc_scores[c] = 1.
else:
auc_scores[c] = roc_auc_score(equ, y_pred_prob_bin)
# ACC
acc_scores[c] = np.mean(y_pred_prob_bin, axis=0)
subset = 'all' if subset is None else subset
npyfile = os.path.join(configer.logspath, configer.modelname,
'analy_{}_auc.npy'.format(subset))
np.save(npyfile, auc_scores)
npyfile = os.path.join(configer.logspath, configer.modelname,
'analy_{}_acc.npy'.format(subset))
np.save(npyfile, acc_scores)
plt.figure("testout_sub_{}_auc".format(subset))
plt.bar(np.arange(C),
auc_scores,
alpha=0.9,
width=0.35,
facecolor='lightblue',
edgecolor='white')
plt.figure("testout_sub_{}_acc".format(subset))
plt.bar(np.arange(C),
acc_scores,
alpha=0.9,
width=0.35,
facecolor='lightblue',
edgecolor='white')
def analysis_similarity(subset=None, mode='euc'):
# get test files and labels
txtfile = './split/{}/test.txt'.format(configer.splitmode)
with open(txtfile, 'r') as f:
testfiles = f.readlines()
testfiles_sub, indexes_sub = get_sub_filenames(testfiles, subset)
labels = get_labels_from_pathlist(testfiles_sub)
labels = list(map(lambda x: x - 1, labels))
# get test outputs
npyfile = os.path.join(configer.logspath, configer.modelname,
'test_out.npy')
testout = np.load(npyfile)
for i in range(testout.shape[0]):
testout[i] = softmax(testout[i])
testout_sub = testout[indexes_sub] # ndarray(N, cls)
if mode == 'euc': # 欧氏距离
f = lambda x, y: np.linalg.norm(x - y)
elif mode == 'cos': # 余弦相似度
f = lambda x, y: x.dot(y) / (np.linalg.norm(x) * np.linalg.norm(y))
# get similarity_matirx, 每两个样本间的相似度
subset = 'all' if subset is None else subset
matrixfile = os.path.join(
configer.logspath, configer.modelname,
'similarity_sample_{}_{}.npy'.format(configer.splitmode, subset, mode))
if os.path.exists(matrixfile):
similarity_matrix = np.load(matrixfile)
else:
similarity_matrix = np.zeros(shape=(testout_sub.shape[0],
testout_sub.shape[0]))
for i in range(testout_sub.shape[0]):
out_i = testout_sub[i]
out_i /= np.linalg.norm(out_i) # 单位化
for j in range(i, testout_sub.shape[0]):
out_j = testout_sub[j]
out_j /= np.linalg.norm(out_j) # 单位化
similarity_matrix[i, j] = f(out_i, out_j)
np.save(matrixfile, similarity_matrix)
# 计算类间相似度, 每两个类间的相似度
matrixfile = os.path.join(
configer.logspath, configer.modelname,
'similarity_class_{}_{}.npy'.format(configer.splitmode, subset, mode))
if os.path.exists(matrixfile):
similarity = np.load(matrixfile)
else:
similarity = np.zeros(shape=(63, 63))
for i in range(similarity.shape[0]): # 类 i
idxi = [l for l in range(len(labels)) if labels[l] == i]
for j in range(similarity.shape[1]): # 类 j
idxj = [l for l in range(len(labels)) if labels[l] == j]
sum_similarity = 0
cnt = 0
for ii in idxi:
for jj in idxj:
cnt += 1
sum_similarity += similarity_matrix[ii, jj]
similarity[i, j] = sum_similarity / cnt
np.save(matrixfile, similarity)
print(similarity)
# draw figure
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
fig = plt.figure()
ax = Axes3D(fig)
x = np.arange(similarity.shape[0])
y = np.arange(similarity.shape[1])
x, y = np.meshgrid(x, y)
ax.plot_surface(x, y, similarity, cmap=plt.cm.winter)
plt.show()