def train(fpath, max_depth, max_features, n_estimators):
"""
:param params: hyperparameters. Its structure is consistent with how search space is defined. See below.
:param fpath: Path or URL for the training data used with the model.
:param max_depth: RF max_depth parameter
:param max_features: RF max_features parameter
:param n_estimators: RF n_estimators parameter
:return: trained model
"""
X_train, X_test, y_train, y_test = load_data(fpath)
mod = RandomForestClassifier(max_depth=max_depth,
max_features=max_features,
n_estimators=n_estimators)
mod.fit(X_train, y_train)
preds = mod.predict(X_test)
acc = accuracy_score(y_test, preds)
mlparams = {
"max_depth": str(max_depth),
"max_features": str(max_features),
"n_estimators": str(n_estimators),
}
mlflow.log_params(mlparams)
mlflow.log_metric("accuracy", acc)
mlflow.sklearn.log_model(mod, "saved_models")
return mod
def GridSearch_random_forest(X_train, y_train):
# Encode as float32
X_train = X_train.to_numpy().astype('float32')
y_train = y_train.to_numpy().astype('float32')
# Init Kfolds
folds = KFold(n_splits=5)
# Init hyperparam vals
n_estimators_lst = [128, 256, 512, 1024]
max_features_lst = ['sqrt', 'log2']
fin_arr = []
# Run GridSearch for all hyperparam combos
for n_estimators in n_estimators_lst:
for max_features in max_features_lst:
# Init clf
clf = RandomForestClassifier(n_estimators=n_estimators,
max_features=max_features)
predicted_y = []
true_y = []
# Run CV and calc metrics
for train, holdout in folds.split(X_train):
clf.fit(X_train[train], y_train[train])
predicted_y.append(clf.predict(X_train[holdout]))
true_y.append(y_train[holdout])
predicted_y = np.concatenate(predicted_y)
true_y = np.concatenate(true_y)
accuracy_train = accuracy_score(true_y, predicted_y)
f1_train = f1_score(true_y, predicted_y)
roc_auc_train = roc_auc_score(true_y, predicted_y)
fin_arr.append([
n_estimators, max_features, accuracy_train, f1_train,
roc_auc_train
])
# Create final dataframe from GridSearch results
fin_arr = np.array(fin_arr).reshape(
(len(n_estimators_lst) * len(max_features_lst)), 5)
columns = [
'n_estimators', 'max_features', 'mean_accuracy', 'mean_f1', 'mean_auc'
]
results = pd.DataFrame(data=fin_arr, columns=columns)
results.n_estimators = results.n_estimators.astype(int)
return results
def fast_rf_classifier(
X,
y,
*,
num_classes=2,
split_algo=1,
split_criterion=0,
min_rows_per_node=2,
min_impurity_decrease=0.0,
bootstrap_features=False,
rows_sample=1.0,
max_leaves=-1,
n_estimators=100,
max_depth=16,
max_features='auto',
bootstrap=True,
n_bins=8,
n_cols=None,
dtype=None,
accuracy_metric=None,
quantile_per_tree=False,
n_streams=8,
random_state: int = 1,
n_jobs: Optional[int] = None,
framework: Literal['auto', 'cuml', 'sklearn'] = 'auto',
**kwargs,
):
kw = dict(locals())
kwargs = kw.pop('kwargs')
X = kw.pop('X')
y = kw.pop('y')
kw.update(kwargs)
framework = kw.pop('framework')
### import
is_cuml = False
if framework == 'sklearn':
RFC = RandomForestClassifier
else:
try:
from cuml.ensemble import RandomForestClassifier as RFC
is_cuml = True
except ImportError as e:
RFC = RandomForestClassifier
### fine-tune keywords
if is_cuml:
kw['output_type'] = 'numpy'
kw['seed'] = kw.pop('random_state')
else:
kw = dict()
### training
tree = RFC()
for k, v in tree.__dict__.items():
print(k, v)
exit()
tree.fit(X, y)
return tree
def train_and_eval(X_param, y_param, max_depth=16, n_estimators=100):
X_train, X_valid, y_train, y_valid = train_test_split(X_param,
y_param,
random_state=77)
classifier = RandomForestClassifier(max_depth=max_depth,
n_estimators=n_estimators)
classifier.fit(X_train, y_train)
y_pred = classifier.predict(X_valid)
score = accuracy_score(y_valid, y_pred)
return score
def fit(X, y):
global clf
clf = RandomForestClassifier(split_criterion=params.criterion,
split_algo=params.split_algorithm,
n_estimators=params.num_trees,
max_depth=params.max_depth,
max_features=params.max_features,
min_samples_split=params.min_samples_split,
max_leaves=params.max_leaf_nodes,
min_impurity_decrease=params.min_impurity_decrease,
bootstrap=params.bootstrap)
return clf.fit(X, y)
def _train(params, fpath, hyperopt=False):
"""
:param params: hyperparameters. Its structure is consistent with how search space is defined. See below.
:param fpath: Path or URL for the training data used with the model.
:param hyperopt: Use hyperopt for hyperparameter search during training.
:return: dict with fields 'loss' (scalar loss) and 'status' (success/failure status of run)
"""
max_depth, max_features, n_estimators = params
max_depth, max_features, n_estimators = (int(max_depth),
float(max_features),
int(n_estimators))
# Log all of our training parameters for this run.
pyver = sys.version_info
mlparams = {
'cudf_version': str(cudf.__version__),
'cuml_version': str(cuml.__version__),
'max_depth': str(max_depth),
'max_features': str(max_features),
'n_estimators': str(n_estimators),
'python_version': f"{pyver[0]}.{pyver[1]}.{pyver[2]}.{pyver[3]}",
}
mlflow.log_params(mlparams)
X_train, X_test, y_train, y_test = load_data(fpath)
mod = RandomForestClassifier(max_depth=max_depth,
max_features=max_features,
n_estimators=n_estimators)
mod.fit(X_train, y_train)
preds = mod.predict(X_test)
acc = accuracy_score(y_test, preds)
mlflow.log_metric("accuracy", acc)
mlflow.sklearn.log_model(mod, "saved_models")
if not hyperopt:
return mod
return {"loss": acc, "status": STATUS_OK}
def _train(params, fpath, hyperopt=False):
"""
:param params: hyperparameters. Its structure is consistent with how search space is defined. See below.
:param fpath: Path or URL for the training data used with the model.
:param hyperopt: Use hyperopt for hyperparameter search during training.
:return: dict with fields 'loss' (scalar loss) and 'status' (success/failure status of run)
"""
max_depth, max_features, n_estimators = params
max_depth, max_features, n_estimators = (int(max_depth),
float(max_features),
int(n_estimators))
X_train, X_test, y_train, y_test = load_data(fpath)
mod = RandomForestClassifier(max_depth=max_depth,
max_features=max_features,
n_estimators=n_estimators)
mod.fit(X_train, y_train)
preds = mod.predict(X_test)
acc = accuracy_score(y_test, preds)
mlparams = {
"max_depth": str(max_depth),
"max_features": str(max_features),
"n_estimators": str(n_estimators)
}
mlflow.log_params(mlparams)
mlflow.log_metric("accuracy", acc)
mlflow.sklearn.log_model(mod, "saved_models")
if (not hyperopt):
return mod
return {'loss': acc, 'status': STATUS_OK}
print('Copying data to GPU done in {:.2f} seconds'.format(time() - t0))
# ### Learning
#
# Random forest classifiers are quick to train, quite robust to
# hyperparameter values, and often work relatively well.
print()
print('Learning begins')
t0 = time()
n_estimators = 100
max_depth = 16
clf_rf = RandomForestClassifier(n_estimators=n_estimators, max_depth=max_depth)
print(clf_rf)
clf_rf.fit(cu_X_train, cu_y_train)
print('Learning done in {:.2f} seconds'.format(time() - t0))
# ### Inference
#
# We will use GPU-based inference to predict the classes for the test
# data.
print()
print('Inference begins')
t0 = time()
pred_rf = clf_rf.predict(X_test, predict_model='GPU')
pred_rf = [chr(x) for x in pred_rf + ord('A')]
pred_rf = np.array(pred_rf)
def run_random_forest(scaled_df):
raw_train_arr = []
raw_test_arr = []
# Over five trials
for i in range(5):
# Split data into train and test
X_train, X_test, y_train, y_test = train_test_split(
scaled_df.iloc[:, :-1], scaled_df.y, train_size=5000)
# Run GridSearch
search_results = GridSearch_random_forest(X_train, y_train)
results = search_results
# Get optimal clfs using gridsearch results
opt_acc_inf = results.sort_values(by='mean_accuracy',
ascending=False).iloc[0]
opt_f1_inf = results.sort_values(by='mean_f1', ascending=False).iloc[0]
opt_auc_inf = results.sort_values(by='mean_auc',
ascending=False).iloc[0]
# Init optimal clfs
opt_acc_clf = RandomForestClassifier(
n_estimators=opt_acc_inf.n_estimators,
max_features=opt_acc_inf.max_features)
opt_f1_clf = RandomForestClassifier(
n_estimators=opt_f1_inf.n_estimators,
max_features=opt_f1_inf.max_features)
opt_auc_clf = RandomForestClassifier(
n_estimators=opt_auc_inf.n_estimators,
max_features=opt_auc_inf.max_features)
# Encode as float32 for cuML
X_train_np = X_train.to_numpy().astype('float32')
y_train_np = y_train.to_numpy().astype('float32')
X_test_np = X_test.to_numpy().astype('float32')
y_test_np = y_test.to_numpy().astype('float32')
# Fit clfs
opt_acc_clf.fit(X_train_np, y_train_np)
opt_f1_clf.fit(X_train_np, y_train_np)
opt_auc_clf.fit(X_train_np, y_train_np)
# Get train and test metrics
train_score_acc = opt_acc_clf.score(X_train_np, y_train_np)
train_score_f1 = f1_score(y_train_np, opt_f1_clf.predict(X_train_np))
train_score_auc = roc_auc_score(y_train_np,
opt_auc_clf.predict(X_train_np))
test_score_acc = opt_acc_clf.score(X_test_np, y_test_np)
test_score_f1 = f1_score(y_test_np, opt_f1_clf.predict(X_test_np))
test_score_auc = roc_auc_score(y_test_np,
opt_auc_clf.predict(X_test_np))
raw_train_arr.append(
[train_score_acc, train_score_f1, train_score_auc])
raw_test_arr.append([test_score_acc, test_score_f1, test_score_auc])
raw_train_arr = np.array(raw_train_arr).reshape(5, 3)
raw_test_arr = np.array(raw_test_arr).reshape(5, 3)
raw_train_df = pd.DataFrame(data=raw_train_arr,
columns=['accuracy', 'f1', 'auc'])
raw_test_df = pd.DataFrame(data=raw_test_arr,
columns=['accuracy', 'f1', 'auc'])
return raw_train_df, raw_test_df
class CUMLTrainable(tune.Trainable):
def _setup(self, config):
# [X_train, X_test, y_train, y_test] = get_pinned_object(data_id)
self._gpu_id = os.environ.get("CUDA_VISIBLE_DEVICES",
0) #ray.get_gpu_ids()[0]
print("Starting new trainable on {}.".format(self._gpu_id))
# self._wait_for_gpus()
with FileLock(os.path.expanduser("~/.tune.gpulock")):
X_cudf_train = cudf.DataFrame.from_pandas(X_train)
self.train_mat = X_cudf_train.as_gpu_matrix(order="F")
del X_cudf_train
self.X_cudf_test = cudf.DataFrame.from_pandas(X_test)
self.y_cudf_train = cudf.Series(y_train.values)
self.y_test = y_test
config = {k: int(v) for k, v in config.items()}
self.cuml_model = GPURandomForestClassifier(**config)
def _train(self):
self.cuml_model.fit(self.train_mat, self.y_cudf_train)
fil_preds_orig = self.cuml_model.predict(self.X_cudf_test)
accuracy = accuracy_score(self.y_test, fil_preds_orig)
return {"mean_accuracy": accuracy}
def _stop(self):
import time
import GPUtil
gpu_object = GPUtil.getGPUs()[self._gpu_id]
print("Deleting the model. Mem: {:0.3f}".format(gpu_object.memoryUsed))
del self.cuml_model
print("Deleting the test set. Mem: {:0.3f}".format(
gpu_object.memoryUsed))
del self.X_cudf_test
print("Deleting the test labels. Mem: {:0.3f}".format(
gpu_object.memoryUsed))
del self.y_test
print("Deleting the training labels. Mem: {:0.3f}".format(
gpu_object.memoryUsed))
del self.y_cudf_train
print("Deleting the training matrix. Mem: {:0.3f}".format(
gpu_object.memoryUsed))
del self.train_mat
# self._wait_for_gpus(retry=1)
def _wait_for_gpus(self, retry=10):
import GPUtil
import time
gpu_object = GPUtil.getGPUs()[self._gpu_id]
for i in range(int(retry)):
if gpu_object.memoryUsed > 0.1:
print("Waiting for GPU memory to free. Mem: {:0.3f}".format(
gpu_object.memoryUsed))
time.sleep(5)
time.sleep(5)
def reset_config(self, config):
del self.cuml_model
config = {k: int(v) for k, v in config.items()}
self.cuml_model = GPURandomForestClassifier(**config)
return True
import pickle
from datasets import prepare_dataset
from cuml.ensemble import RandomForestClassifier as GPURandomForestClassifier
data = prepare_dataset("/data", "airline", None)
X_train, X_test, y_train, y_test = data.X_train, data.X_test, data.y_train, data.y_test
y_train = y_train.astype(np.int32)
y_test = y_test.astype(np.int32)
QUARTER = len(X_train) // 2
X_train = X_train[QUARTER:]
y_train = y_train[QUARTER:]
X_cudf_train = cudf.DataFrame.from_pandas(X_train)
X_cudf_test = cudf.DataFrame.from_pandas(X_test)
train_mat = X_cudf_train.as_gpu_matrix(order="F")
del X_cudf_train
y_cudf_train = cudf.Series(y_train.values)
cuml_model = GPURandomForestClassifier(n_estimators=467,
max_depth=19,
max_features=1.0)
cuml_model.fit(train_mat, y_cudf_train)
fil_preds_orig = cuml_model.predict(X_cudf_test)
fil_acc_orig = accuracy_score(y_test, fil_preds_orig)