/
cross_validation.py
180 lines (149 loc) · 7.19 KB
/
cross_validation.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
import numpy as np
import pandas as pd
from joblib import Parallel, delayed
from utils import flatten_list
from evaluate import evaluate
##------------------------------------------------------------------##
## Cross-Validation Utils
##------------------------------------------------------------------##
# For finding the folds of each trial
def fold_trials(num_trials, num_folds=10):
fold_ids = np.array(list(range(num_folds)))
n_reps = np.ceil(float(num_trials) / num_folds)
n_excessive = int((num_folds * n_reps) - num_trials)
smaller_folds = np.random.choice(fold_ids, n_excessive, replace=False)
reps_per_fold = [n_reps - 1 if fid in smaller_folds else n_reps for fid in fold_ids]
grouping_factor = np.repeat(fold_ids, reps_per_fold, axis=0)
np.random.shuffle(grouping_factor)
return grouping_factor
def cross_validate_time_point(X, y, trial_folds, train_predict_fn, use_features=None):
folds = np.unique(trial_folds)
target_collection = []
predicted_probs_collection = []
predicted_class_collection = []
trial_id_collection = []
fold_collection = []
for fold in folds:
train_indices = np.where(trial_folds != fold)[0]
test_indices = np.where(trial_folds == fold)[0]
if isinstance(X, pd.DataFrame):
if use_features is not None:
X = X.loc[:, use_features]
X_train = np.asarray(X.iloc[train_indices])
X_test = np.asarray(X.iloc[test_indices])
else:
X_train = X[train_indices]
X_test = X[test_indices]
y_train = y[train_indices]
y_test = y[test_indices]
# Fit model and predict test set
predicted_probs, predicted_class = train_predict_fn(X_train=X_train, X_test=X_test, y_train=y_train)
# Append to collections
trial_id_collection.append(test_indices)
target_collection.append(y_test)
predicted_probs_collection.append(predicted_probs)
predicted_class_collection.append(predicted_class)
fold_collection.append([fold] * len(test_indices))
return pd.DataFrame({"Fold": flatten_list(fold_collection),
"Trial": flatten_list(trial_id_collection),
"Target": flatten_list(target_collection),
"Predicted Probability": flatten_list(predicted_probs_collection),
"Predicted Class": flatten_list(predicted_class_collection)})
def cross_validate_all_time_points(time_point_dfs, y, trial_folds, train_predict_fn,
use_features, parallel=False, cores=7):
"""
Note: time_point_dfs should be list of tuples with time point and df, e.g. (0, df_0)
"""
# Repeated CV
repeats = len(trial_folds)
def cv_single(time_point_df, time_point, rep):
# Run cross-validation
# returns predictions as data frame
predictions = cross_validate_time_point(X=time_point_df, y=y, trial_folds=trial_folds[rep],
train_predict_fn=train_predict_fn,
use_features=use_features)
predictions["Time Point"] = time_point
predictions["Repetition"] = rep
# Evaluate predictions
eval = evaluate(list(predictions["Target"]),
list(predictions["Predicted Class"]))
eval["Time Point"] = time_point
eval["Repetition"] = rep
print(eval)
return predictions, eval
if parallel:
preds, evals = zip(*Parallel(n_jobs=cores)(delayed(cv_single)(df, tp, rep) \
for tp, df in time_point_dfs for rep in range(repeats)))
else:
preds, evals = zip(*[cv_single(df, tp, rep) for tp, df in time_point_dfs for rep in range(repeats)])
all_predictions = pd.concat(preds)
all_evaluations = pd.concat(evals)
return all_predictions, all_evaluations
## Multigroup
def leave_one_group_out_cv_single_time_point(X, y, group_names, train_predict_fn, use_features=None):
if not isinstance(X, pd.DataFrame) or not isinstance(y, pd.DataFrame):
raise KeyError("leave_one_group_out_cv expects X and y to be data frames.")
groups = np.unique(group_names)
target_collection = []
predicted_probs_collection = []
predicted_class_collection = []
trial_id_collection = []
group_id_collection = []
for group in groups:
# Subset X and y
train_set = X[X['group'] != group]
test_set = X[X['group'] == group]
train_labels = y[y['group'] != group]
test_labels = y[y['group'] == group]
# Add trial info to collection
trial_id_collection.append(list(test_set["Trial"]))
# Extract sensors
if use_features is not None:
train_set = train_set.loc[:, use_features]
test_set = test_set.loc[:, use_features]
# Convert to numpy arrays
X_train = np.asarray(train_set)
X_test = np.asarray(test_set)
y_train = np.asarray(train_labels["label"])
y_test = np.asarray(test_labels["label"])
# Fit model and predict test set
predicted_probs, predicted_class = train_predict_fn(X_train=X_train, X_test=X_test, y_train=y_train)
# Append to collections
target_collection.append(y_test)
predicted_probs_collection.append(predicted_probs)
predicted_class_collection.append(predicted_class)
group_id_collection.append([group] * len(predicted_class))
return pd.DataFrame({"Group": flatten_list(group_id_collection),
"Trial": flatten_list(trial_id_collection),
"Target": flatten_list(target_collection),
"Predicted Probability": flatten_list(predicted_probs_collection),
"Predicted Class": flatten_list(predicted_class_collection)})
def cross_validate_all_time_points_by_group(time_point_dfs, y, group_names,
train_predict_fn,
use_features, parallel=False, cores=7):
"""
Note: time_point_dfs should be list of tuples with group and a list of paths
(group_name, [(tp_1, path_1),(tp_2, path_2),...])
"""
def cv_single(time_point_df, time_point):
# Run cross-validation
# returns predictions as data frame
predictions = leave_one_group_out_cv_single_time_point(
X=time_point_df, y=y, group_names=group_names,
train_predict_fn=train_predict_fn,
use_features=use_features)
predictions["Time Point"] = time_point
# Evaluate predictions
eval = evaluate(list(predictions["Target"]),
list(predictions["Predicted Class"]))
eval["Time Point"] = time_point
print(eval)
return predictions, eval
if parallel:
preds, evals = zip(*Parallel(n_jobs=cores)(delayed(cv_single)(df, tp) \
for tp, df in time_point_dfs))
else:
preds, evals = zip(*[cv_single(df, tp) for tp, df in time_point_dfs])
all_predictions = pd.concat(preds)
all_evaluations = pd.concat(evals)
return all_predictions, all_evaluations