-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluate.py
380 lines (293 loc) · 13.7 KB
/
evaluate.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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
import numpy as np
import os
from random import randrange
from pandas import DataFrame, Series
import sklearn.metrics
from sklearn.metrics import plot_roc_curve, precision_recall_curve, f1_score, roc_auc_score, auc, confusion_matrix
from sklearn.model_selection import train_test_split, StratifiedKFold # Import train_test_split function
from sklearn.linear_model import LogisticRegressionCV, LogisticRegression
# from sklearn import svm, datasets
import matplotlib
matplotlib.use('Agg') # use a non-interactive backend such as Agg (for PNGs), PDF, SVG or PS.
from matplotlib import pyplot as plt
from plot_utils import saveFig, plot_roc
def eval_performance(X, y, model=None, cv=5, random_state=53, **kargs):
"""
Memo
----
1. https://scikit-learn.org/stable/modules/classes.html#module-sklearn.metrics
"""
# from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
import time
#### parameters ####
verbose = kargs.get('verbose', 0)
algo_name = kargs.get('method', 'LG')
output_dir = kargs.get('output_dir', 'plot')
plot_prefix = kargs.get('prefix', '')
save_plot = kargs.get('save', True)
show_plot = kargs.get('show', False)
# plot_ext = kargs.get('ext', 'tif')
####################
# clf = LogisticRegressionCV(cv=5, random_state=random_state, scoring=).fit(X, y)
if not model:
model = LogisticRegression(class_weight='balanced', penalty='l1', solver='saga')
elif isinstance(model, str):
if model.startswith(('log', 'defaut')):
model = LogisticRegression(class_weight='balanced', penalty='l1', solver='saga')
else:
raise NotImplementedError
kf = StratifiedKFold(n_splits=cv, shuffle=True, random_state=random_state)
if len(y.shape) == 1:
y = y.reshape((y.shape[0], 1))
cv_scores =[]
cv_data = []
predictions = {i: {} for i in range(cv)}
fold_number_act = np.random.randint(0, cv)
for i, (train, test) in enumerate(kf.split(X,y)):
fold_number = i
if verbose: print('[eval] {} of KFold {}'.format(fold_number, kf.n_splits))
X_train, X_test = X[train], X[test]
y_train, y_test = y[train], y[test]
#model
model.fit(X_train, y_train)
y_pred = model.predict_proba(X_test)[:,1]
#evaluation
cv_data.append((y_test, y_pred))
score, p_th = fmax_score_threshold(y_test, y_pred)
score_auc = roc_auc_score(y_test, y_pred)
# other metrics
y_pred_label = model.predict(X_test) # to_label_prediction(y_pred, p_th=0.5) #
score_f1 = f1_score(y_test, y_pred_label)
predictions[i]['scores'] = calculate_metrics(y_test, y_pred,
plot=save_plot, show=show_plot and i==fold_number_act,
method=algo_name, fold=fold_number, phase='test', prefix=plot_prefix)
# ... add a prefix to plot name to distinguish between different training data sets
if verbose:
print('> Fmax: {} p_th: {} | F1: {}, AUC: {}'.format(score, p_th, score_f1, score_auc))
cv_scores.append(score)
# [note]
# The ROC curve within each CV fold can be combined while still maintaining visual clarity
# However, it's not the case for the PR curve, and therefore, I've decided to
# plot the PR curve seperately for each CV fold.
if save_plot:
plot_roc(cv_data, output_dir=output_dir, method=algo_name, prefix=plot_prefix, show=show_plot)
return cv_scores, predictions
def to_label_prediction(y_score, p_th=0.5):
# turn probabilities into label prediction given threshold at 'p_th'
yhat = np.zeros(len(y_score))
for i, yp in enumerate(y_score):
if yp >= p_th:
yhat[i] = 1
return yhat
def calculate_accuracy(y_true, y_hat):
"""
Calculate accuracy percentage
"""
correct = 0
for i in range(len(y_true)):
if y_true[i] == y_hat[i]:
correct += 1
return correct / float(len(y_true))
def eval_AUPRC(y_true, y_score, method='LG', plot=True, **kargs):
precision, recall, _ = precision_recall_curve(y_true, y_score)
auprc = auc(recall, precision)
show_plot = kargs.get('show', False)
phase = kargs.get("phase", 'test')
prefix_id = kargs.get('prefix', '')
if plot: # plot the precision-recall curves
plt.clf()
fold_number = kargs.get('fold', 0)
y_true = np.array(y_true)
no_skill = len(y_true[y_true==1]) / len(y_true)
plt.plot([0, 1], [no_skill, no_skill], linestyle='--', label='No Skill')
plt.plot(recall, precision, marker='.', label=method)
# axis labels
plt.xlabel('Recall')
plt.ylabel('Precision')
if prefix_id:
plt.title('PRC for {}: method={}, nfold={})'.format(prefix_id, method, fold_number))
else:
plt.title('PRC for {}, nfold={})'.format(method, fold_number))
# show the legend
plt.legend()
if prefix_id:
prefix = f'{prefix_id}-precision_recall_curve-train' % prefix_id if phase.startswith('tr') else f'{prefix_id}-precision_recall_curve'
else:
prefix = 'precision_recall_curve-train' % prefix_id if phase.startswith('tr') else 'precision_recall_curve'
filename = kargs.get('filename', '{}-{}'.format(prefix, fold_number) if fold_number > 0 else prefix)
output_dir = kargs.get("output_dir", "plot")
output_path = os.path.join(output_dir, filename) # example path: System.analysisPath
saveFig(plt, output_path, ext='tif', dpi=300, message='[output] precision recall curve', verbose=True)
# show the plot
if show_plot:
plt.show()
return auprc
def eval_AUROC(y_true, y_score, method='LG', plot=False, **kargs):
auc = roc_auc_score(y_true, y_score)
if plot:
pass # use utils_plot.plot_roc() instead
# Note: ROC curve in each CV fold can be easily combined while still maintaining visual clarity,
# and therefore, the plotting is deferred until all CV data are collected (see evaluate_algorithm())
return auc
def calculate_metrics(y_true, y_score, p_th=0.5, **kargs):
"""
Calculate performance metrics and plot related performance curves (e.g. precision-recall curve).
Params
------
y_true: true labels
y_score: class conditional probabilities P(y=1|x)
method: the name of the algorithm (for display only)
phase: 'train' for the training phase or 'test' for the test phase
use this to make distinction between the plot associaetd with the training or test
Most use cases only generate performance plot on the test data and therefore
the file name for the plot does not have this keyword.
If, however, we wish to diagnose overfitting by comparing the performance gap
between the training phase and the test phase, the performance plot can be
generated accordingly but with the keyword 'train' added to the plots' file names.
"""
# optional plot params
index = kargs.get("fold", 0)
plot = kargs.get("plot", True)
method = kargs.get("method", 'LR')
phase = kargs.get("phase", 'test') #
metrics = {}
metrics['AUROC'] = eval_AUROC(y_true, y_score, fold=index)
metrics['AUPRC'] = eval_AUPRC(y_true, y_score, fold=index, plot=plot, method=method, phase=phase, prefix=kargs.get('prefix', ''))
y_hat = to_label_prediction(y_score, p_th=p_th)
# ret['f1'] = f1_score(y_true, y_hat)
nTP = nTN = nFP = nFN = 0
for i, y in enumerate(y_true):
if y == 1:
if y_hat[i] == 1:
nTP += 1
else:
nFN += 1
else: # y == 0
if y_hat[i] == 0:
nTN += 1
else:
nFP += 1
metrics['precision'] = nTP/(nTP+nFP+0.0)
metrics['recall'] = nTP/(nTP+nFN+0.0)
metrics['accuracy'] = calculate_accuracy(y_true, y_hat)
return metrics
def perturb(X, cols_x=[], cols_y=[], lower_bound=0, alpha=100.):
def add_noise():
min_nonnegative = np.min(X[np.where(X>lower_bound)])
Eps = np.random.uniform(min_nonnegative/(alpha*10), min_nonnegative/alpha, X.shape)
return X + Eps
# from pandas import DataFrame
if isinstance(X, DataFrame):
from data_processor import toXY
X, y, fset, lset = toXY(X, cols_x=cols_x, cols_y=cols_y, scaler=None, perturb=False)
X = add_noise(X)
dfX = DataFrame(X, columns=fset)
dfY = DataFrame(y, columns=lset)
return pd.concat([dfX, dfY], axis=1)
X = add_noise()
return X
def fmax_score(labels, predictions, beta = 1.0, pos_label = 1):
"""
Reference
---------
Radivojac, P. et al. (2013). A Large-Scale Evaluation of Computational Protein Function Prediction. Nature Methods, 10(3), 221-227.
Manning, C. D. et al. (2008). Evaluation in Information Retrieval. In Introduction to Information Retrieval. Cambridge University Press.
"""
# import sklearn.metrics
precision, recall, threshold = sklearn.metrics.precision_recall_curve(labels, predictions, pos_label)
# the general formula for positive beta
f1 = (1 + beta**2) * (precision * recall) / ((beta**2 * precision) + recall)
# if beta == 1, then this is just f1 score, harmonic mean between precision and recall
# i = np.nanargmax(f1)
# return (f1[i], threshold[i])
return nanmax(f1)
def fmax_score_threshold(labels, predictions, beta = 1.0, pos_label = 1):
"""
Return the fmax score and the probability threhold where the max of f1 (fmax) is reached
"""
# import sklearn.metrics
precision, recall, threshold = sklearn.metrics.precision_recall_curve(labels, predictions, pos_label)
# the general formula for positive beta
# ... if beta == 1, then this is just f1 score, harmonic mean between precision and recall
f1 = (1 + beta**2) * (precision * recall) / ((beta**2 * precision) + recall)
i = np.nanargmax(f1) # the position for which f1 is the max
th = threshold[i] if i < len(threshold) else 1.0 # len(threshold) == len(precision) -1
# assert f1[i] == nanmax(f1)
return (f1[i], th)
def get_sample_sizes(y, col='ICD10'):
import collections
if isinstance(y, DataFrame):
sizes = collections.Counter( y[col].values )
else:
# df is a numpy array or list
if len(y.shape) == 2:
y = y.reshape( (y.shape[0],) )
sizes = collections.Counter(y)
return sizes # label/col -> sample size
def one_vs_all_encoding(df, target_label, codebook={'pos': 1, 'neg': 0}, col='ICD10', col_target='target'):
# inplace operation
if isinstance(df, DataFrame):
assert col in df.columns
cond_pos = df[col] == target_label # target loinc
cond_neg = df[col] != target_label
print("> target: {} (dtype: {}) | n(pos): {}, n(neg): {}".format(target, type(target), np.sum(cond_pos), np.sum(cond_neg)))
df[col_target] = df[col]
df.loc[cond_pos, col_target] = codebook['pos']
df.loc[cond_neg, col_target] = codebook['neg']
else:
df = np.where(df == target_label, codebook['pos'], codebook['neg'])
return df
def encode_labels(df, pos_label, neg_label=None, col_label='ICD10', codebook={}, verbose=1):
if not codebook: codebook = {'pos': 1, 'neg': 0, '+': 1, '-': 0}
y = df[col_label] if isinstance(df, DataFrame) else df
sizes = get_sample_sizes(y)
n0 = sizes[pos_label]
# df.loc[df[col_target] == pos_label]
# col_target='target'
y = one_vs_all_encoding(y, target_label=pos_label, codebook=codebook)
# ... if df is a DataFrame, then df has an additional attribute specified by col_target/'target'
sizes = get_sample_sizes(y)
assert sizes[codebook['pos']] == n0
print("(encode_labels) sample size: {}".format(sizes))
return y
def demo_evaluate_performance():
from sklearn.datasets import load_iris
from utils import summarize_dict
n_folds = 5
X, y = load_iris(return_X_y=True)
# print("> before y:\n{}\n".format(y))
y = encode_labels(y, pos_label=1)
# print("> after y:\n{}\n".format(y))
scores, predictions = eval_performance(X, y, model=None, cv=n_folds, random_state=53, prefix='iris')
print("[demo] Fmax average: {}, std: {}".format(np.mean(scores), np.std(scores)))
for i in range(n_folds):
print(f"... Fold #[{i}]: ")
summarize_dict(predictions[i], topn=15, sort_=True, prefix=' ' * 5)
return scores
def demo_evaluate_health_status_prediction(target_codes=[], model=None, show_plot=False, n_folds=5):
from data_pipeline import load_XY
from utils import summarize_dict
from icd_utils import encode, decode
if not target_codes:
target_codes = ['I10', 'N39.0', 'E78.2', 'F41.9', 'K21.9', 'J20.9', 'M54.5', 'L70.0', 'J06.9', 'E11.9']
# W = (60, 60) # select a window
for target_code in target_codes:
try:
X, y = load_XY(suffix=target_code)
except:
print(f"[demo] Trouble loading data set for diagnosis code: {target_code}! Skipping ...")
continue
plot_id = encode(target_code)
scores, predictions = eval_performance(X, y, model=model, cv=n_folds,
random_state=53, prefix=plot_id, show=show_plot)
print("[demo] Fmax average: {}, std: {}".format(np.mean(scores), np.std(scores)))
for i in range(n_folds):
print(f"[demo] Fold #[{i}]: ")
summarize_dict(predictions[i], topn=15, sort_=True, prefix=' ' * 7)
def test():
# test evaluating performance
# demo_evaluate_performance()
demo_evaluate_health_status_prediction()
return
if __name__ == "__main__":
test()