-
Notifications
You must be signed in to change notification settings - Fork 0
/
figure2.py
189 lines (158 loc) · 8.84 KB
/
figure2.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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
# Script for generating Figure 1, showing the cost - accuracy trade off
import sklearn, sklearn.svm, sklearn.kernel_ridge, sklearn.datasets
import hierarchyrl.powerset, hierarchyrl.utils, hierarchyrl.policy2
import numpy as np, os, functools, itertools, argparse, collections, tqdm, dill
import matplotlib.pyplot as plt, multiprocessing
from experiments import dataReading, tools
def dice(k1, k2):
intersection = len([k for k in k1 if k in k2])
if len(k1) == 0. and len(k2) == 0.:
return 1
return 2 * intersection/(len(k1) + len(k2))
classifier = lambda: tools.addPredictProba(hierarchyrl.utils.PredictorStandardise(
sklearn.svm.SVC(probability = False, C = 1., class_weight = 'balanced')
)
)
regressor = lambda: hierarchyrl.utils.PredictorStandardise(
sklearn.model_selection.GridSearchCV(
sklearn.kernel_ridge.KernelRidge(alpha = 1, kernel = 'rbf'),
param_grid = {'alpha' : np.logspace(-2, 2, num = 5)},
cv = sklearn.model_selection.KFold(n_splits=5, shuffle= True, random_state = 0)
),
scaleY = True)
metrics_fcts = {}
metrics_fcts['acc'] = lambda y, y1: sklearn.metrics.accuracy_score(y, np.argmax(y1, axis = 1))
def doComputations(t, nReps, costs, stateCost, X_p_2, X_p_test):
dataResultSameTestset = tools.DataResult(metrics_fcts, X_p_2.DATA_MASK, stateCost)
for i in range(nReps):
X_p_train, _, _, _ = X_p_2.split(.90, random_rng= np.random.RandomState(i))
X_p_train_1, _, X_p_val, _ = X_p_train.split(.8)
policy = hierarchyrl.policy2.PolicyPowerset(X_p_train,
modelQAction = regressor,
modelClassification = classifier,
loss='accuracy',
acquisitionCost = {k:v *t for k,v in costs.items()}
)
policy.train(debug = False, nIts = 1, offPolicyEpsilon = 0.5)
yPred_dqn, ks, _ = policy.simulateEvaluateInPolicy( X_powerset= X_p_test)
dataResultSameTestset.addResult('Reinforcement learning', X_p_test.y, yPred_dqn,ks)
model = classifier()
best_val_acc = -10000
nVars = len(X_p_train.variablesNames)
for v in itertools.chain(*[itertools.combinations(range(nVars), i) for i in range(0, nVars + 1)]):
k = functools.reduce(lambda x, y: x | y, [1 << vv for vv in v], 0)
if k != 0:
model.fit(X_p_train_1.getData(k),X_p_train_1.y)
val_acc = np.mean(policy.loss(X_p_val.y, model.predict_proba(X_p_val.getData(k)))) - policy.stateCost[k]
else:
yPred = np.repeat([1 - np.mean(X_p_train_1.y), np.mean(X_p_train_1.y)] ,X_p_val.nSamples).reshape((2, -1)).T
val_acc = np.mean(policy.loss(X_p_val.y, yPred))
if val_acc > best_val_acc:
best_val_acc = val_acc
best_k = k
if best_k != 0:
model.fit(X_p_train.getData(best_k),X_p_train.y)
yPred = model.predict_proba(X_p_test.getData(best_k))
else:
yPred = np.repeat([1 - np.mean(X_p_train.y), np.mean(X_p_train.y)] ,X_p_test.nSamples).reshape((2, -1)).T
dataResultSameTestset.addResult('Populationwise feature selection', X_p_test.y, yPred, [best_k])
meanDice = collections.defaultdict(list)
for i,j in itertools.combinations( range(nReps), 2):
if i == j:
continue
for algorithm in dataResultSameTestset.ks:
nSamples = len(dataResultSameTestset.ks[algorithm][i])
mean = np.mean([dice(X_p_test.getVariablesFromEncoding(dataResultSameTestset.ks[algorithm][i][k]),
X_p_test.getVariablesFromEncoding(dataResultSameTestset.ks[algorithm][j][k]))
for k in range(nSamples)])
meanDice[algorithm].append(mean)
return meanDice
if __name__ == '__main__':
# Parameters
parser = argparse.ArgumentParser(description= 'Figure 2 of the paper, experiment in which the reproducibility of the modalities chosen by the policy is evaluated.')
parser.add_argument('-output', default = './FiguresMICCAI')
parser.add_argument('-nIts', default = 20, type =int)
parser.add_argument('-nReps', default = 3, type =int)
parser.add_argument('-numTsamples', default = 10, type = int)
parser.add_argument('--parallel', action = 'store_true', default = False)
parser.add_argument('--ignore-cleveland', action = 'store_true', default = False)
parser.add_argument('--readOnly', action = 'store_true', default = False)
args = parser.parse_args()
num_t_samples = args.numTsamples
nIts = args.nIts
nReps = args.nReps
resultPath = args.output
# Data reading
all_costs = {}
all_data = {}
if not args.ignore_cleveland:
all_data['Cleveland'], all_costs['Cleveland'], _ = dataReading.readDataCleveland()
all_data['Hypertension'], all_costs['Hypertension'], _ = dataReading.readDataHypertensionCensored()
#
if args.readOnly:
results = {}
allDiceByDataset = {}
for dataset in all_costs:
costs = all_costs[dataset]
X_p = all_data[dataset]
allDice= collections.defaultdict(lambda: collections.defaultdict(list))
allDiceByDataset[dataset] = allDice
ts = np.logspace(-3, -1, num = num_t_samples)
policyTest = hierarchyrl.policy2.PolicyPowerset(X_p,
modelQAction = regressor,
modelClassification = classifier,
loss='accuracy',
acquisitionCost = {k:v for k,v in costs.items()}
)
results[dataset] = tools.DataResultList(ts, metrics_fcts, X_p.DATA_MASK, policyTest.stateCost)
for ii in tqdm.tqdm(range(nIts)):
X_p_2, _, X_p_test, _ = X_p.split(.90, random_rng= np.random.RandomState(ii))
if not args.parallel:
for t in ts:
meanDice = doComputations(t, nReps, costs, policyTest.stateCost, X_p_2, X_p_test)
for algorithm in meanDice:
allDice[algorithm][t].append(meanDice[algorithm])
else:
with multiprocessing.Pool() as pool:
rs = pool.starmap(doComputations, zip(ts,
itertools.repeat(nReps), itertools.repeat(costs), itertools.repeat(policyTest.stateCost),
itertools.repeat(X_p_2), itertools.repeat(X_p_test)))
for t, meanDice in zip(ts, rs):
for algorithm in meanDice:
allDice[algorithm][t].append(meanDice[algorithm])
else:
with open(os.path.join(resultPath, 'exp2.pkl'), 'rb' ) as f:
allDiceByDataset = dill.load(f)
# Save data and generate figures
score = 'acc'
_, fs = plt.subplots(ncols = max(len(allDiceByDataset), 2), figsize = (12, 2.5))
for i, (k, allDice) in enumerate(allDiceByDataset.items()):
plt.sca(fs[i])
maxCosts = []
minCosts = []
for algo in allDice:
if algo == 'NIPS - 2015':
continue
meanCost = np.array([ np.mean(allCosts[k][algo][t]) for t in ts])
maxCosts.append(np.max(meanCost))
minCosts.append(np.min(meanCost))
mean = np.array([ np.mean(allDice[algo][t]) for t in ts])
std = np.array([ np.std(allDice[algo][t]) for t in ts])
line = plt.plot(meanCost, mean, '-o', label = algo)
plt.fill_between(meanCost, mean - std, mean + std, alpha = .15, color = line[0].get_color())
plt.fill_between(meanCost, mean - 1.96/np.sqrt(nIts) *std, mean + 1.96/np.sqrt(nIts) *std, alpha = .45, color = line[0].get_color())
plt.ylabel('Dice Score', fontsize = 12)
plt.xlabel('Average cost [Arbitrary unit]', fontsize = 12)
plt.xlim(max(minCosts), min(maxCosts))
plt.legend(loc = 4, fontsize = 12)
if k == 'Hypertension':
k = 'Hypertense'
elif k == 'Cleveland':
k = 'Heart Disease'
plt.title(k, y = 1, pad = -14, fontsize = 14, bbox = dict(boxstyle='round', facecolor='lightgray', alpha=0.85))
plt.ylim(0, 1)
plt.tight_layout()
plt.savefig('FiguresMICCAI/exp2.pdf')
# Save data
with open(os.path.join(resultPath, 'exp2.pkl'), 'wb' ) as f:
dill.dump(allDiceByDataset, f)