def parameter_test_worker(i, p, num_params, alpha, group_size, data_x, data_y):
"""
Get parameter sets from a queue and find the random forest performance. Send them back through another queue.
:param qu_gl: queue for global objects, such as the dataset
:param qu_in: queue for the parameter sets
:param qu_out: queue for the output
"""
out_str = "# Parameter set %d of %d" % (
i + 1, num_params) + "\n" + json.dumps(p) + "\n\n"
rf_split_counts = []
rf_performance = []
rf_num_nodes = []
rf_train_time = []
fg = {
a: {
"split_counts": [],
"performance": [],
"num_nodes": [],
"train_time": []
}
for a in alpha
}
kf = sklearn.cross_validation.KFold(data_x.shape[0], n_folds=10)
for kf_i, (train, test) in enumerate(kf):
print "## kf %d of %d" % (kf_i + 1, len(kf))
train_x = data_x[train]
train_y = data_y[train]
test_x = data_x[test]
test_y = data_y[test]
# Train the rf and get the performance.
rf = randomforest.RandomForestClassifier(n_rand_dims="auto",
n_jobs=1,
**p)
start = time.time()
rf.fit(train_x, train_y)
end = time.time()
pred, split_counts = rf.predict(test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum(1 for a, b in zip(test_y, pred) if a == b)
performance = count / float(len(pred))
rf_split_counts.append(split_counts)
rf_performance.append(performance)
rf_num_nodes.append(rf.num_nodes())
rf_train_time.append(end - start)
# Train the forest garrote and get the performance.
for a_i, a in enumerate(alpha):
print "## forest garrote %d of %d" % (a_i + 1, len(alpha))
start = time.time()
if rf.num_trees() <= group_size:
refined_rf = forest_garrote(rf,
train_x,
train_y,
group_size=None,
alpha=a)
else:
refined_rf = forest_garrote(rf,
train_x,
train_y,
group_size=group_size,
alpha=a)
end = time.time()
pred, split_counts = refined_rf.predict(test_x,
return_split_counts=True)
split_counts /= float(len(pred))
count = sum(1 for a, b in zip(test_y, pred) if a == b)
performance = count / float(len(pred))
fg[a]["split_counts"].append(split_counts)
fg[a]["performance"].append(performance)
fg[a]["num_nodes"].append(refined_rf.num_nodes())
fg[a]["train_time"].append(end - start)
# Create the output string.
out_str += "# performance\n" + str(numpy.mean(rf_performance)) + " " + str(
numpy.std(rf_performance)) + "\n"
out_str += "# train_time\n" + str(numpy.mean(rf_train_time)) + " " + str(
numpy.std(rf_train_time)) + "\n"
out_str += "# split_counts\n" + str(
numpy.mean(rf_split_counts)) + " " + str(
numpy.std(rf_split_counts)) + "\n"
out_str += "# num_nodes\n" + str(numpy.mean(rf_num_nodes)) + " " + str(
numpy.std(rf_num_nodes)) + "\n\n"
for a in alpha:
out_str += "fg " + str(a) + "\n\n"
out_str += "# performance\n" + str(
numpy.mean(fg[a]["performance"])) + " " + str(
numpy.std(fg[a]["performance"])) + "\n"
out_str += "# train_time\n" + str(
numpy.mean(fg[a]["train_time"])) + " " + str(
numpy.std(fg[a]["train_time"])) + "\n"
out_str += "# split_counts\n" + str(numpy.mean(
fg[a]["split_counts"])) + " " + str(
numpy.std(fg[a]["split_counts"])) + "\n"
out_str += "# num_nodes\n" + str(
numpy.mean(fg[a]["num_nodes"])) + " " + str(
numpy.std(fg[a]["num_nodes"])) + "\n\n"
print out_str[:-1]
with open("LOGFILE.txt", "a") as f:
f.write(out_str)
def parameter_test_worker(i, p, num_params, alpha, group_size, data_x, data_y):
"""
Get parameter sets from a queue and find the random forest performance. Send them back through another queue.
:param qu_gl: queue for global objects, such as the dataset
:param qu_in: queue for the parameter sets
:param qu_out: queue for the output
"""
out_str = "# Parameter set %d of %d" % (i+1, num_params) + "\n" + json.dumps(p) + "\n\n"
rf_split_counts = []
rf_performance = []
rf_num_nodes = []
rf_train_time = []
fg = {a: {"split_counts": [],
"performance": [],
"num_nodes": [],
"train_time": []}
for a in alpha}
kf = sklearn.cross_validation.KFold(data_x.shape[0], n_folds=10)
for kf_i, (train, test) in enumerate(kf):
print "## kf %d of %d" % (kf_i+1, len(kf))
train_x = data_x[train]
train_y = data_y[train]
test_x = data_x[test]
test_y = data_y[test]
# Train the rf and get the performance.
rf = randomforest.RandomForestClassifier(n_rand_dims="auto", n_jobs=1, **p)
start = time.time()
rf.fit(train_x, train_y)
end = time.time()
pred, split_counts = rf.predict(test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum(1 for a, b in zip(test_y, pred) if a == b)
performance = count/float(len(pred))
rf_split_counts.append(split_counts)
rf_performance.append(performance)
rf_num_nodes.append(rf.num_nodes())
rf_train_time.append(end-start)
# Train the forest garrote and get the performance.
for a_i, a in enumerate(alpha):
print "## forest garrote %d of %d" % (a_i+1, len(alpha))
start = time.time()
if rf.num_trees() <= group_size:
refined_rf = forest_garrote(rf, train_x, train_y, group_size=None, alpha=a)
else:
refined_rf = forest_garrote(rf, train_x, train_y, group_size=group_size, alpha=a)
end = time.time()
pred, split_counts = refined_rf.predict(test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum(1 for a, b in zip(test_y, pred) if a == b)
performance = count/float(len(pred))
fg[a]["split_counts"].append(split_counts)
fg[a]["performance"].append(performance)
fg[a]["num_nodes"].append(refined_rf.num_nodes())
fg[a]["train_time"].append(end-start)
# Create the output string.
out_str += "# performance\n" + str(numpy.mean(rf_performance)) + " " + str(numpy.std(rf_performance)) + "\n"
out_str += "# train_time\n" + str(numpy.mean(rf_train_time)) + " " + str(numpy.std(rf_train_time)) + "\n"
out_str += "# split_counts\n" + str(numpy.mean(rf_split_counts)) + " " + str(numpy.std(rf_split_counts)) + "\n"
out_str += "# num_nodes\n" + str(numpy.mean(rf_num_nodes)) + " " + str(numpy.std(rf_num_nodes)) + "\n\n"
for a in alpha:
out_str += "fg " + str(a) + "\n\n"
out_str += "# performance\n" + str(numpy.mean(fg[a]["performance"])) + " " + str(numpy.std(fg[a]["performance"])) + "\n"
out_str += "# train_time\n" + str(numpy.mean(fg[a]["train_time"])) + " " + str(numpy.std(fg[a]["train_time"])) + "\n"
out_str += "# split_counts\n" + str(numpy.mean(fg[a]["split_counts"])) + " " + str(numpy.std(fg[a]["split_counts"])) + "\n"
out_str += "# num_nodes\n" + str(numpy.mean(fg[a]["num_nodes"])) + " " + str(numpy.std(fg[a]["num_nodes"])) + "\n\n"
print out_str[:-1]
with open("LOGFILE.txt", "a") as f:
f.write(out_str)
def train_rf(n_trees,
n_jobs,
predict=True,
save=False,
load=False,
filename=None,
refine=False,
group_size=None):
"""
Train a random forest and compute the accuracy on a test set.
:param n_trees: number of trees
:param n_jobs: number of jobs
:param predict: use the random forest to predict on a test set
:param save: save the random forest to a file
:param load: load the random forest from a file
:param filename: file name
"""
# train_x, train_y, test_x, test_y = load_data([3, 8])
train_x, train_y, test_x, test_y = load_neuro_data()
if load:
assert os.path.isfile(filename)
print "Loading random forest from file %s." % filename
with open(filename, "r") as f:
rf_str = f.read()
rf = randomforest.RandomForestClassifier.from_string(rf_str)
if n_jobs is not None:
rf._n_jobs = n_jobs
else:
print "Training random forest with %d trees." % n_trees
rf = randomforest.RandomForestClassifier(
n_estimators=n_trees,
n_rand_dims="auto",
n_jobs=n_jobs,
# bootstrap_sampling=True, use_sample_label_count=True, resample_count=None,
# bootstrap_sampling=False, use_sample_label_count=False, resample_count=None,
bootstrap_sampling=True,
use_sample_label_count=False,
resample_count=None,
# bootstrap_sampling=False, use_sample_label_count=True, resample_count=None, # does not make sense
# resample_count=20,
# loggamma_tau=1e-6,
split_selection="gini")
with Timer("Training took %.03f seconds"):
rf.fit(train_x, train_y)
print "The random forest has %d nodes." % rf.num_nodes()
if save and not load:
print "Saving random forest to file %s." % filename
with open(filename, "w") as f:
f.write(rf.to_string())
if predict:
print "Predicting on a test set with the random forest."
with Timer("Random forest prediction took %.03f seconds."):
pred, split_counts = rf.predict(test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum([1 if a == b else 0 for a, b in zip(test_y, pred)])
print "%d of %d correct (%.03f%%), used %.02f splits per instance" % (
count, len(pred), (100.0 * count) / len(pred), split_counts)
if refine:
print "Refining the random forest using forest garrote."
with Timer("Refining took %.03f seconds."):
refined_rf = forest_garrote(rf,
train_x,
train_y,
group_size=group_size)
# refined_rf = global_refinement(rf, train_x, train_y)
print "The refined forest has %d nodes." % refined_rf.num_nodes()
if save:
f0, f1 = os.path.split(filename)
refined_filename = os.path.join(f0, "refined_" + f1)
print "Saving refined random forest to file %s." % refined_filename
with open(refined_filename, "w") as f:
f.write(refined_rf.to_string())
if predict:
print "Predicting on a test set with the forest garrote."
with Timer("Forest garrote prediction took %.03f seconds."):
pred, split_counts = refined_rf.predict(
test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum([1 if a == b else 0 for a, b in zip(test_y, pred)])
print "%d of %d correct (%.03f%%), used %.02f splits per instance" % (
count, len(pred), (100.0 * count) / len(pred), split_counts)
def train_rf(n_trees, n_jobs, predict=True, save=False, load=False, filename=None, refine=False, group_size=None):
"""
Train a random forest and compute the accuracy on a test set.
:param n_trees: number of trees
:param n_jobs: number of jobs
:param predict: use the random forest to predict on a test set
:param save: save the random forest to a file
:param load: load the random forest from a file
:param filename: file name
"""
# train_x, train_y, test_x, test_y = load_data([3, 8])
train_x, train_y, test_x, test_y = load_neuro_data()
if load:
assert os.path.isfile(filename)
print "Loading random forest from file %s." % filename
with open(filename, "r") as f:
rf_str = f.read()
rf = randomforest.RandomForestClassifier.from_string(rf_str)
if n_jobs is not None:
rf._n_jobs = n_jobs
else:
print "Training random forest with %d trees." % n_trees
rf = randomforest.RandomForestClassifier(n_estimators=n_trees, n_rand_dims="auto", n_jobs=n_jobs,
# bootstrap_sampling=True, use_sample_label_count=True, resample_count=None,
# bootstrap_sampling=False, use_sample_label_count=False, resample_count=None,
bootstrap_sampling=True, use_sample_label_count=False, resample_count=None,
# bootstrap_sampling=False, use_sample_label_count=True, resample_count=None, # does not make sense
# resample_count=20,
# loggamma_tau=1e-6,
split_selection="gini"
)
with Timer("Training took %.03f seconds"):
rf.fit(train_x, train_y)
print "The random forest has %d nodes." % rf.num_nodes()
if save and not load:
print "Saving random forest to file %s." % filename
with open(filename, "w") as f:
f.write(rf.to_string())
if predict:
print "Predicting on a test set with the random forest."
with Timer("Random forest prediction took %.03f seconds."):
pred, split_counts = rf.predict(test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum([1 if a == b else 0 for a, b in zip(test_y, pred)])
print "%d of %d correct (%.03f%%), used %.02f splits per instance" % (count, len(pred), (100.0*count)/len(pred), split_counts)
if refine:
print "Refining the random forest using forest garrote."
with Timer("Refining took %.03f seconds."):
refined_rf = forest_garrote(rf, train_x, train_y, group_size=group_size)
# refined_rf = global_refinement(rf, train_x, train_y)
print "The refined forest has %d nodes." % refined_rf.num_nodes()
if save:
f0, f1 = os.path.split(filename)
refined_filename = os.path.join(f0, "refined_" + f1)
print "Saving refined random forest to file %s." % refined_filename
with open(refined_filename, "w") as f:
f.write(refined_rf.to_string())
if predict:
print "Predicting on a test set with the forest garrote."
with Timer("Forest garrote prediction took %.03f seconds."):
pred, split_counts = refined_rf.predict(test_x, return_split_counts=True)
split_counts /= float(len(pred))
count = sum([1 if a == b else 0 for a, b in zip(test_y, pred)])
print "%d of %d correct (%.03f%%), used %.02f splits per instance" % (count, len(pred), (100.0*count)/len(pred), split_counts)