def run(training, cross_v, perc_test, file_output, n_trees=100, nmodels=1, max_features = 20, cpus=-1 ):
base_clf=RandomForestClassifier(n_estimators= n_trees , class_weight="balanced", oob_score=True );
samples, groups, dimensions, values = read_kmer_matrix(training)
classnames, Y = np.unique(groups, return_inverse=True)
n_dim=values.shape[1]
if n_dim < max_features :
max_features = n_dim
groupcount = np.bincount(Y)
print("Starting the creation of {} models.".format(nmodels))
print("Original data have {} dimensions with {} samples, divided in {} groups:".format(n_dim, values.shape[0],len(classnames)))
for c in range(0, len(classnames)) :
print("\t{} - {} \t {} samples".format(c, classnames[c] ,groupcount[c]))
ofs=open(file_output, "w")
ofs.write("round\tfeatures\taccuracy\ttrain_accuracy\toob_score\n")
for i in range(0, nmodels):
print("\n\nRound {}".format(i), flush=True)
fsel = SelectFromModel(DecisionTreeClassifier(), max_features=max_features, threshold=-np.inf)
fsel.fit(values, Y)
for j in range(2, max_features+1):
fsel.max_features=j
clf=clone(base_clf)
tmp_acc = cross_validate(clf, fsel.transform(values), Y, scoring="balanced_accuracy", cv=StratifiedShuffleSplit(n_splits=cross_v, test_size=perc_test), n_jobs=cpus, return_train_score=True)
acc = np.mean( tmp_acc["test_score"])
acc_train = np.mean(tmp_acc["train_score"] )
clf.fit(fsel.transform(values), Y)
toprint="{}\t{}\t{}\t{}\t{}".format(i, j, acc, acc_train, clf.oob_score_ )
ofs.write(toprint+"\n")
print(toprint, flush=True)
return;
def run(training,
test,
file_output,
n_trees=100,
nmodels=1,
max_features=20,
cpus=-1):
base_clf = RandomForestClassifier(n_estimators=n_trees,
class_weight="balanced",
oob_score=True)
samples, groups, dimensions, values = read_kmer_matrix(training)
test_samples, test_groups, test_dimensions, test_values = read_kmer_matrix(
test)
support = []
for d in dimensions:
try:
support.append(test_dimensions.index(d))
except:
print("ERROR! Feature {} is not in the input matrix.".format(d))
exit(1)
test_values = test_values[:, support]
classnames, Y = np.unique(groups, return_inverse=True)
Y_test = np.zeros(test_groups.shape, dtype=int)
for i in range(0, len(classnames)):
Y_test[test_groups == classnames[i]] = i
n_dim = values.shape[1]
groupcount = np.bincount(Y)
print("Starting the creation of {} models.".format(nmodels))
print(
"Original data have {} dimensions with {} samples, divided in {} groups:"
.format(n_dim, values.shape[0], len(classnames)))
for c in range(0, len(classnames)):
print("\t{} - {} \t {} samples".format(c, classnames[c],
groupcount[c]))
ofs = open(file_output, "w")
for i in range(0, nmodels):
print("\n\nRound {}".format(i), flush=True)
fsel = SelectFromModel(DecisionTreeClassifier(),
max_features=max_features,
threshold=-np.inf)
fsel.fit(values, Y)
for j in range(2, max_features + 1):
fsel.max_features = j
clf = clone(base_clf)
clf.fit(fsel.transform(values), Y)
acc = balanced_accuracy_score(
Y_test, clf.predict(fsel.transform(test_values)))
toprint = "{}\t{}\t{}\t{}".format(i, j, acc, clf.oob_score_)
ofs.write(toprint + "\n")
print(toprint, flush=True)
return
