def run(self):
lock = self.lock
remain_trees = self.remain_trees
cpu_classifier = self.cpu_classifier
max_features = self.max_features
bootstrap = self.bootstrap
n_jobs = self.n_jobs
result_queue = self.result_queue
forests = list()
X = self.X
Y = self.Y
if max_features == None:
max_features = "auto"
self.Y = self.Y.astype(np.uint16)
classifier_name = cpu_classifier.__name__
n_trees_before = 0
while True:
lock.acquire()
#The trees trained by GPU in last round
n_trees_by_gpu = n_trees_before - remain_trees.value
#Stop trainning if the remaining tree is smaller than the sum of n_jobs and n_trees trained by gpu
if remain_trees.value < n_jobs or (
n_trees_before != 0
and remain_trees.value - n_jobs <= n_trees_by_gpu):
lock.release()
break
remain_trees.value -= n_jobs
n_trees_before = remain_trees.value
lock.release()
util.log_info("%s got %s jobs.", classifier_name, n_jobs)
f = cpu_classifier(n_estimators=n_jobs,
n_jobs=n_jobs,
bootstrap=bootstrap,
max_features=max_features)
f.fit(X, Y)
forests.append(f)
result_queue.put(forests)
util.log_info("CPU DONE")
def run(self):
reload(util)
driver.init()
ctx = driver.Device(self.gpu_id).make_context()
X = self.X
Y = self.Y
bootstrap = self.bootstrap
max_features = self.max_features
bfs_threshold = self.bfs_threshold
remain_trees = self.remain_trees
lock = self.lock
trees = list()
forest = cdRF(n_estimators=1,
bootstrap=bootstrap,
max_features=max_features)
forest.fit_init(X, Y)
while True:
lock.acquire()
if remain_trees.value == 0:
lock.release()
break
remain_trees.value -= 1
lock.release()
util.log_info("CUDA gets 1 job")
tree = RandomClassifierTree(forest)
si, n_samples = forest._get_sorted_indices(forest.sorted_indices)
tree.fit(forest.samples, forest.target, si, n_samples)
trees.append(tree)
forest.fit_release()
self.result_queue.put(trees)
ctx.detach()
del ctx
util.log_info("GPU DONE")
def run(self):
lock = self.lock
remain_trees = self.remain_trees
cpu_classifier = self.cpu_classifier
max_features = self.max_features
bootstrap = self.bootstrap
n_jobs = self.n_jobs
result_queue = self.result_queue
forests = list()
X = self.X
Y = self.Y
if max_features == None:
max_features = "auto"
self.Y = self.Y.astype(np.uint16)
classifier_name = cpu_classifier.__name__
n_trees_before = 0
while True:
lock.acquire()
#The trees trained by GPU in last round
n_trees_by_gpu = n_trees_before - remain_trees.value
#Stop trainning if the remaining tree is smaller than the sum of n_jobs and n_trees trained by gpu
if remain_trees.value < n_jobs or (n_trees_before != 0 and remain_trees.value - n_jobs <= n_trees_by_gpu):
lock.release()
break
remain_trees.value -= n_jobs
n_trees_before = remain_trees.value
lock.release()
util.log_info("%s got %s jobs.", classifier_name, n_jobs)
f = cpu_classifier(n_estimators = n_jobs, n_jobs = n_jobs,
bootstrap = bootstrap, max_features = max_features)
f.fit(X, Y)
forests.append(f)
result_queue.put(forests)
util.log_info("CPU DONE")
def run(self):
reload(util)
driver.init()
ctx = driver.Device(self.gpu_id).make_context()
X = self.X
Y = self.Y
bootstrap = self.bootstrap
max_features = self.max_features
bfs_threshold = self.bfs_threshold
remain_trees = self.remain_trees
lock = self.lock
trees = list()
forest = cdRF(n_estimators = 1,
bootstrap = bootstrap,
max_features = max_features)
forest.fit_init(X, Y)
while True:
lock.acquire()
if remain_trees.value == 0:
lock.release()
break
remain_trees.value -= 1
lock.release()
util.log_info("CUDA gets 1 job")
tree = RandomClassifierTree(forest)
si, n_samples = forest._get_sorted_indices(forest.sorted_indices)
tree.fit(forest.samples, forest.target, si, n_samples)
trees.append(tree)
forest.fit_release()
self.result_queue.put(trees)
ctx.detach()
del ctx
util.log_info("GPU DONE")
def cpu_build(cpu_classifier,
X,
Y,
n_estimators,
bootstrap,
max_features,
n_jobs,
remain_trees,
result_queue, lock):
"""
Build some trees on cpu, the cpu classifier should be cpu
implementation of random forest classifier.
"""
forests = list()
if max_features == None:
max_features = "auto"
Y = Y.astype(np.uint16)
classifier_name = cpu_classifier.__name__
while True:
lock.acquire()
if remain_trees.value < 2 * n_jobs:
lock.release()
break
remain_trees.value -= n_jobs
lock.release()
util.log_info("%s got %s jobs.", classifier_name, n_jobs)
f = cpu_classifier(n_estimators = n_jobs, n_jobs = n_jobs,
bootstrap = bootstrap, max_features = max_features)
f.fit(X, Y)
forests.append(f)
result_queue.put(forests)
util.log_info("%s's job done", classifier_name)
