def feature_precisions(A, y, feature_sels, test_size=0.2, ITER_TIMES=50, classifier='xgboost', params=None):
precisions = defaultdict(list)
xgb_params = {'silent': 1, 'objective': 'multi:softmax', 'num_class': 10}
if params is not None and classifier == 'xgboost':
xgb_params.update(params)
elif params is None:
params = {}
num_round = 50
for r in range(ITER_TIMES):
A_train, A_test, y_train, y_test = train_test_split(A, y, test_size=test_size)
ms = []
for key, features_selection in feature_sels.items():
_A_train = select_features(A_train, features_selection)
_A_test = select_features(A_test, features_selection)
if classifier == 'xgboost':
_A_train = xgb.DMatrix(_A_train, label=y_train)
_A_test = xgb.DMatrix(_A_test, label=y_test)
clf = xgb.train(xgb_params, _A_train, num_round)
elif classifier == 'svm':
clf = OneVsOneClassifier(SVC(**params))
clf.fit(_A_train, y_train)
h = np.array(clf.predict(_A_test)).astype(int)
p = accuracy_score(h, y_test)
precisions[key].append(p)
ms.append('{:>7s} precision:{:7.2%}'.format(key, p))
sys.stdout.flush()
sys.stdout.write('Round {:3d}/{:3d}:{}\r'.format(r+1, ITER_TIMES, '|'.join(ms)))
precision_info(precisions, ITER_TIMES)
return precisions
def get_coords(routes:list, output='dict'):
# output is either 'df' or 'dict'
url = make_request_url(routes, verbose=True)
json = utils.grab(url)
trains = parse_train_data(json)
features = ['lat', 'lon']
trains = utils.select_features(trains, features)
if output == 'df':
trains = pd.DataFrame.from_dict(trains, orient='index')
return trains
def run(data_fn,
prop_missing=0.,
max_num_feature=-1,
feature_selection='random',
k=10,
data_dir='_data',
out_dir='_out'):
"""Run RIDDLE classification interpretation pipeline.
Arguments:
data_fn: string
data file filename
prop_missing: float
proportion of feature observations which should be randomly masked;
values in [0, 1)
max_num_feature: int
maximum number of features to use
feature_selection: string
feature selection method; values = {'random', 'frequency', 'chi2'}
k: int
number of partitions for k-fold cross-validation
interpret_model: bool
whether to interpret the trained model for first k-fold partition
which_half: str
which half of experiments to do; values = {'first', 'last', 'both'}
data_dir: string
directory where data files are located
cache_dir: string
directory where cached files (e.g., saved parameters) are located
out_dir: string
outer directory where outputs (e.g., results) should be saved
"""
from keras.models import load_model
from riddle import emr, feature_importance
from riddle.models import MLP
start = time.time()
base_out_dir = get_base_out_dir(out_dir, 'riddle', data_fn, prop_missing,
max_num_feature, feature_selection)
recursive_mkdir(base_out_dir)
# get common data
x_unvec, y, idx_feat_dict, idx_class_dict, icd9_descript_dict, perm_indices = (
get_preprocessed_data(data_dir, data_fn, prop_missing=prop_missing))
num_feature = len(idx_feat_dict)
num_class = len(idx_class_dict)
list_sums_D, list_sums_D2, list_sums_contribs = [], [], []
for k_idx in range(k):
full_out_dir = '{}/k_idx={}'.format(base_out_dir, k_idx)
print('\nPartition k = {}'.format(k_idx))
x_train_unvec, y_train, _, _, x_test_unvec, y_test = emr.get_k_fold_partition(
x_unvec, y, k_idx=k_idx, k=k, perm_indices=perm_indices)
if max_num_feature > 0: # select features and re-encode
feat_encoding_dict, idx_feat_dict = select_features(
x_train_unvec,
y_train,
idx_feat_dict,
method=feature_selection,
num_feature=num_feature,
max_num_feature=max_num_feature)
x_test_unvec = subset_reencode_features(x_test_unvec,
feat_encoding_dict)
num_feature = max_num_feature
# interpret
start = time.time()
temp_mlp = MLP(num_feature=num_feature, num_class=num_class)
hdf5_path = full_out_dir + '/model.h5'
sums_D, sums_D2, sums_contribs, pairs = \
feature_importance.get_diff_sums(
hdf5_path,
x_test_unvec,
process_x_func=temp_mlp.process_x,
num_feature=num_feature,
num_class=num_class)
with open(full_out_dir + '/sums_D.pkl', 'wb') as f:
pickle.dump(sums_D, f)
with open(full_out_dir + '/sums_D2.pkl', 'wb') as f:
pickle.dump(sums_D2, f)
with open(full_out_dir + '/sums_contribs.pkl', 'wb') as f:
pickle.dump(sums_contribs, f)
list_sums_D.append(sums_D)
list_sums_D2.append(sums_D2)
list_sums_contribs.append(sums_contribs)
def compute_total_sums(list_sums):
total_sums = list_sums[0]
for i in range(1, len(list_sums)):
for j in range(len(total_sums)):
total_sums[j] = np.add(total_sums[j], list_sums[i][j])
return total_sums
total_sums_D = compute_total_sums(list_sums_D)
total_sums_D2 = compute_total_sums(list_sums_D2)
total_sums_contribs = compute_total_sums(list_sums_contribs)
num_sample = len(x_unvec)
run_interpretation_summary(x_unvec,
y,
total_sums_D,
total_sums_D2,
total_sums_contribs,
idx_feat_dict=idx_feat_dict,
idx_class_dict=idx_class_dict,
icd9_descript_dict=icd9_descript_dict,
pairs=pairs,
num_sample=num_sample,
full_out_dir=base_out_dir)
print('Computed DeepLIFT scores and analysis in {:.4f} seconds'.format(
time.time() - start))
print('-' * 72)
print()
def run(method, x_unvec, y, idx_feat_dict, num_feature, max_num_feature,
num_class, max_num_sample, feature_selection, k_idx, k, num_search,
perm_indices):
"""Run a parameter search for a single k-fold partitions
Arguments:
method: string
name of classification method; values = {'logit', 'random_forest',
'linear_svm', 'poly_svm', 'rbf_svm', 'gbdt', 'riddle'}
x_unvec: [[int]]
feature indices that have not been vectorized; each inner list
collects the indices of features that are present (binary on)
for a sample
y: [int]
list of class labels as integer indices
idx_feat_dict: {int: string}
dictionary mapping feature indices to features
num_feature: int
number of features present in the dataset
max_num_feature: int
maximum number of features to use
num_class: int
number of classes present
feature_selection: string
feature selection method; values = {'random', 'frequency', 'chi2'}
k_idx: int
index of the k-fold partition to use
k: int
number of partitions for k-fold cross-validation
num_search: int
number of searches (parameter configurations) to try
perm_indices: np.ndarray, int
array of indices representing a permutation of the samples with
shape (num_sample, )
Returns:
best_param: {string: ?}
dictionary mapping parameter names to the best values found
"""
print('-' * 72)
print('Partition k = {}'.format(k_idx))
x_train_unvec, y_train, x_val_unvec, y_val, _, _ = (
emr.get_k_fold_partition(x_unvec,
y,
k_idx=k_idx,
k=k,
perm_indices=perm_indices))
if max_num_feature > 0: # select features and re-encode
feat_encoding_dict, _ = select_features(
x_train_unvec,
y_train,
idx_feat_dict,
method=feature_selection,
num_feature=num_feature,
max_num_feature=max_num_feature)
x_val_unvec = subset_reencode_features(x_val_unvec, feat_encoding_dict)
num_feature = max_num_feature
# cap number of validation samples
if max_num_sample != None and len(x_val_unvec) > max_num_sample:
x_val_unvec = x_val_unvec[0:max_num_sample]
y_val = y_val[0:max_num_sample]
start = time.time()
if method == 'riddle':
model_class = MLP
init_args = {'num_feature': num_feature, 'num_class': num_class}
param_dist = {
'num_hidden_layer': 2, # [1, 2]
'num_hidden_node': 512, # [128, 256, 512]
'activation': ['prelu', 'relu'],
'dropout': tuning.Uniform(lo=0.2, hi=0.8),
'learning_rate': tuning.UniformLogSpace(10, lo=-6, hi=-1),
}
best_param = tuning.random_search(model_class,
init_args,
param_dist,
x_val_unvec,
y_val,
num_class=num_class,
k=TUNING_K,
num_search=num_search)
else: # scikit-learn methods
x_val = vectorize_features(x_val_unvec, num_feature)
if method == 'logit': # logistic regression
from sklearn.linear_model import LogisticRegression
estimator = LogisticRegression(multi_class='multinomial',
solver='lbfgs')
param_dist = {'C': tuning.UniformLogSpace(base=10, lo=-3, hi=3)}
elif method == 'random_forest':
from sklearn.ensemble import RandomForestClassifier
estimator = RandomForestClassifier()
param_dist = {
'max_features': ['sqrt', 'log2', None],
'max_depth': tuning.UniformIntegerLogSpace(base=2, lo=0, hi=7),
'n_estimators': tuning.UniformIntegerLogSpace(base=2,
lo=4,
hi=8)
}
elif method == 'linear_svm':
from sklearn.svm import SVC
# remark: due to a bug in scikit-learn / libsvm, the sparse 'linear'
# kernel is much slower than the sparse 'poly' kernel, so we use
# the 'poly' kernel with degree=1 over the 'linear' kernel
estimator = SVC(kernel='poly',
degree=1,
coef0=0.,
gamma=1.,
probability=True,
cache_size=1000)
param_dist = {'C': tuning.UniformLogSpace(base=10, lo=-2, hi=1)}
elif method == 'poly_svm':
from sklearn.svm import SVC
estimator = SVC(kernel='poly', probability=True, cache_size=1000)
param_dist = {
'C': tuning.UniformLogSpace(base=10, lo=-2, hi=1),
'degree': [2, 3, 4],
'gamma': tuning.UniformLogSpace(base=10, lo=-5, hi=1)
}
elif method == 'rbf_svm':
from sklearn.svm import SVC
estimator = SVC(kernel='rbf', probability=True, cache_size=1000)
param_dist = {
'C': tuning.UniformLogSpace(base=10, lo=-2, hi=1),
'gamma': tuning.UniformLogSpace(base=10, lo=-5, hi=1)
}
elif method == 'gbdt':
from xgboost import XGBClassifier
estimator = XGBClassifier(objective='multi:softprob')
param_dist = {
'max_depth': tuning.UniformIntegerLogSpace(base=2, lo=0, hi=5),
'n_estimators': tuning.UniformIntegerLogSpace(base=2,
lo=4,
hi=8),
'learning_rate': tuning.UniformLogSpace(base=10, lo=-3, hi=0)
}
else:
raise ValueError('unknown method: {}'.format(method))
param_search = RandomizedSearchCV(estimator,
param_dist,
refit=False,
n_iter=num_search,
scoring=loss_scorer)
param_search.fit(x_val, y_val)
best_param = param_search.best_params_
print('Best parameters for {} for k_idx={}: {} found in {:.3f} s'.format(
method, k_idx, best_param,
time.time() - start))
return best_param
def run(ModelClass, x_unvec, y, idx_feat_dict, num_feature, max_num_feature,
num_class, feature_selection, k_idx, k, params, perm_indices,
init_args, full_out_dir):
"""Run a classification pipeline for a single k-fold partition.
Arguments:
ModelClass: Python class
classification model
x_unvec: [[int]]
feature indices that have not been vectorized; each inner list
collects the indices of features that are present (binary on)
for a sample
y: [int]
list of class labels as integer indices
idx_feat_dict: {int: string}
dictionary mapping feature indices to features
num_feature: int
number of features present in the dataset
max_num_feature: int
maximum number of features to use
num_class: int
number of classes present
feature_selection: string
feature selection method; values = {'random', 'frequency', 'chi2'}
k_idx: int
index of the k-fold partition to use
k: int
number of partitions for k-fold cross-validation
params: [{string: ?}]
list of dictionary mapping parameter names to values for each
k-fold partition
perm_indices: np.ndarray, int
array of indices representing a permutation of the samples with
shape (num_sample, )
init_args: {string: ?}
dictionary mapping initialization argument names to values
out_dir: string
directory where outputs (e.g., results) should be saved
"""
print('-' * 72)
print('Partition k = {}'.format(k_idx))
print(params[k_idx])
x_train_unvec, y_train, _, _, x_test_unvec, y_test = (
emr.get_k_fold_partition(x_unvec,
y,
k_idx=k_idx,
k=k,
perm_indices=perm_indices))
if max_num_feature > 0: # select features and re-encode
feat_encoding_dict, _ = select_features(
x_train_unvec,
y_train,
idx_feat_dict,
method=feature_selection,
num_feature=num_feature,
max_num_feature=max_num_feature)
x_train_unvec = subset_reencode_features(x_train_unvec,
feat_encoding_dict)
x_test_unvec = subset_reencode_features(x_test_unvec,
feat_encoding_dict)
num_feature = max_num_feature
x_train = vectorize_features(x_train_unvec, num_feature)
x_test = vectorize_features(x_test_unvec, num_feature)
args = dict(init_args) # copy dictionary
args.update(params[k_idx])
start = time.time()
model = ModelClass(**args)
model.fit(x_train, y_train)
y_test_probas = model.predict_proba(x_test)
runtime = time.time() - start
evaluate(y_test,
y_test_probas,
runtime,
num_class=num_class,
out_dir=full_out_dir)
def run(x_unvec, y, idx_feat_dict, idx_class_dict, icd9_descript_dict,
num_feature, max_num_feature, num_class, feature_selection, k_idx, k,
params, perm_indices, full_out_dir):
"""Run a RIDDLE classification pipeline for a single k-fold partition.
Arguments:
x_unvec: [[int]]
feature indices that have not been vectorized; each inner list
collects the indices of features that are present (binary on)
for a sample
y: [int]
list of class labels as integer indices
idx_feat_dict: {int: string}
dictionary mapping feature indices to features
idx_class_dict: {int: string}
dictionary mapping class indices to classes
icd9_descript_dict: {string: string}
dictionary mapping ICD9 codes to description text
num_feature: int
number of features present in the dataset
max_num_feature: int
maximum number of features to use
num_class: int
number of classes
feature_selection: string
feature selection method; values = {'random', 'frequency', 'chi2'}
k_idx: int
index of the k-fold partition to use
k: int
number of partitions for k-fold cross-validation
params: [{string: ?}]
list of dictionary mapping parameter names to values for each
k-fold partition
perm_indices: np.ndarray, int
array of indices representing a permutation of the samples with
shape (num_sample, )
full_out_dir: string
directory where outputs (e.g., results) should be saved
"""
from keras import backend as K
from riddle import emr, feature_importance
from riddle.models import MLP
print('Partition k = {}'.format(k_idx))
print()
x_train_unvec, y_train, x_val_unvec, y_val, x_test_unvec, y_test = (
emr.get_k_fold_partition(x_unvec, y, k_idx=k_idx, k=k,
perm_indices=perm_indices))
if max_num_feature > 0: # select features and re-encode
feat_encoding_dict, idx_feat_dict = select_features(
x_train_unvec, y_train, idx_feat_dict,
method=feature_selection, num_feature=num_feature,
max_num_feature=max_num_feature)
x_train_unvec = subset_reencode_features(x_train_unvec,
feat_encoding_dict)
x_val_unvec = subset_reencode_features(x_val_unvec,
feat_encoding_dict)
x_test_unvec = subset_reencode_features(x_test_unvec,
feat_encoding_dict)
num_feature = max_num_feature
# set up
max_num_epoch = -1
if 'debug' in full_out_dir:
max_num_epoch = 3
model = MLP(num_feature=num_feature, num_class=num_class,
max_num_epoch=max_num_epoch, **params[k_idx])
# train and test
start = time.time()
model.train(x_train_unvec, y_train, x_val_unvec, y_val)
y_test_probas = model.predict_proba(x_test_unvec)
runtime = time.time() - start
print('Completed training and testing in {:.4f} seconds'.format(runtime))
print('-' * 72)
print()
# evaluate model performance
evaluate(y_test, y_test_probas, runtime, num_class=num_class,
out_dir=full_out_dir)
model.save_model(path=full_out_dir + '/model.h5')
K.clear_session()
print('Finished with partition k = {}'.format(k_idx))
print('=' * 72)
print()
def main():
"""
Module to execute the entire package from data retrieval to model
performance metrics
@:param: None
:return: Post process results
"""
# Importing inhibitor notation data
# The SMILES and InChI logs of the same material have identical indices
# Creating and joining the SMILES and InChI dataframes along the same index
utils.check_files()
df_compounds_smiles = utils.create_dataframe(
'data/chemical_notation_'
'data/compounds_smiles.txt', 'smiles')
df_compounds_smiles.rename(columns={'ID': 'CID'}, inplace=True)
df_compounds_smiles.sort_values(by='CID', inplace=True)
# Importing inhibitor activity data
activity = pd.read_csv('data/activity_data/AID_743255_datatable.csv')
activity = utils.clean_activity_dataframe(activity)
# Merging activity data and compound notation data
df = activity.merge(df_compounds_smiles)
df.sort_values(by='CID', inplace=True)
df.reset_index(drop=True, inplace=True)
# Drop non-descriptor columns before feature space reduction
df_target = df.drop(['SMILES', 'CID', 'Phenotype'], axis=1)
# Extracting molecular descriptors for all compounds
# print("Sending data for descriptor calculation")
# utils.extract_all_descriptors(df, 'SMILES')
# Importing feature sets
df_charge = pd.DataFrame.from_csv('data/df_charge.csv')
df_basak = pd.DataFrame.from_csv('data/df_basak.csv')
df_con = pd.DataFrame.from_csv('data/df_con.csv')
df_estate = pd.DataFrame.from_csv('data/df_estate.csv')
df_constitution = pd.DataFrame.from_csv('data/df_constitution.csv')
df_property = pd.DataFrame.from_csv('data/df_property.csv')
df_kappa = pd.DataFrame.from_csv('data/df_kappa.csv')
df_moe = pd.DataFrame.from_csv('data/df_moe.csv')
print("Joining dataframes")
df_descriptor = df_kappa.join(df_moe).join(df_constitution).\
join(df_property).join(df_charge).join(df_estate).join(df_con).join(
df_basak)
print("Joining dataframes done")
print("Checking dataframe for NaN, infinite or too large values")
df_descriptor = utils.remove_nan_infinite(df_descriptor)
# Transform all column values to mean 0 and unit variance
print("Transforming dataframe using mean and variance")
df_descriptor = utils.transform_dataframe(df_descriptor)
print("Transforming dataframe using mean and variance done")
# Feature selection and space reduction
print("Selecting best features in dataframe")
df_features = utils.select_features(df_descriptor, df_target)
print("Selecting best features in dataframe done")
df = df_features.join(df_target)
# Data to training task
# Type check inputs for sanity
if df is None:
raise ValueError('df is None')
if not isinstance(df, pd.DataFrame):
raise TypeError('df is not a dataframe')
if TARGET_COLUMN is None:
raise ValueError('target_column is None')
if not isinstance(TARGET_COLUMN, basestring):
raise TypeError('target_column is not a string')
if TARGET_COLUMN not in df.columns:
raise ValueError('target_column (%s) is not a valid column name' %
TARGET_COLUMN)
# Train, validation and test split
df_train, df_test = sklearn.cross_validation.train_test_split(
df, test_size=0.25)
# Remove the classification column from the dataframe
x_train = df_train.drop(TARGET_COLUMN, 1)
x_test = df_test.drop(TARGET_COLUMN, 1)
y_train = pd.DataFrame(df_train[TARGET_COLUMN])
y_test = pd.DataFrame(df_test[TARGET_COLUMN])
with open(XY_PICKLE, 'wb') as results:
pickle.dump(x_train, results, pickle.HIGHEST_PROTOCOL)
pickle.dump(x_test, results, pickle.HIGHEST_PROTOCOL)
pickle.dump(y_train, results, pickle.HIGHEST_PROTOCOL)
pickle.dump(y_test, results, pickle.HIGHEST_PROTOCOL)
models.run_models(x_train, y_train, x_test, y_test)
post_process.results()
