def compute_opt_tree_depth(algorithm, n_test_tree=2000000):
opt_tree_depth = None
if algorithm in ["buffer_kd_tree", "kd_tree"]:
# the different tree depths that shall
# be tested for this data set
if algorithm == "buffer_kd_tree":
tree_depths = range(4, 12)
elif algorithm == "kd_tree":
tree_depths = range(8, 16)
# search for optimal tree depth
model = NearestNeighbors(n_neighbors=n_neighbors,
algorithm=algorithm,
n_jobs=n_jobs,
plat_dev_ids=plat_dev_ids,
verbose=verbose)
opt_tree_depth = compute_optimal_tree_depth(model,
Xtrain,
Xtest[:n_test_tree],
target="test",
tree_depths=tree_depths)
print("Optimal tree depth found: %i " % opt_tree_depth)
return opt_tree_depth
def run_algorithm(n_test, tree_depth=None, algorithm="buffer_kd_tree"):
print("----------------------------------------------------------------------")
print("\n\nRunning %s for n_test=%i ...\n" % (algorithm, n_test))
print("----------------------------------------------------------------------")
Xtest_local = Xtest[:n_test, :]
# instantiate model
nbrs = NearestNeighbors(n_neighbors=n_neighbors, \
algorithm=algorithm, \
n_jobs=n_jobs, \
tree_depth=opt_tree_depth, \
plat_dev_ids=plat_dev_ids, \
verbose=verbose)
# train model
start_time = time.time()
nbrs.fit(Xtrain)
end_time = time.time()
train_time = (end_time - start_time)
print("Fitting time: %f" % train_time)
# apply model (testing phase)
start_time = time.time()
_, _ = nbrs.kneighbors(Xtest_local)
end_time = time.time()
test_time = (end_time - start_time)
print("Testing time: %f" % test_time)
return train_time, test_time
def prepare_input(self, x, y):
tree = NearestNeighbors(
algorithm="buffer_kd_tree",
plat_dev_ids=plat_dev_ids,
tree_depth=9,
)
tree.fit(x)
return tree, y
def run_algorithm(n_test, tree_depth=None, algorithm="buffer_kd_tree"):
print(
"----------------------------------------------------------------------"
)
print("\n\nRunning %s for n_test=%i ...\n" % (algorithm, n_test))
print(
"----------------------------------------------------------------------"
)
Xtest_local = Xtest[:n_test, :]
# instantiate model
nbrs = NearestNeighbors(n_neighbors=n_neighbors, \
algorithm=algorithm, \
n_jobs=n_jobs, \
tree_depth=opt_tree_depth, \
plat_dev_ids=plat_dev_ids, \
verbose=verbose)
# train model
start_time = time.time()
nbrs.fit(Xtrain)
end_time = time.time()
train_time = (end_time - start_time)
print("Fitting time: %f" % train_time)
# apply model (testing phase)
start_time = time.time()
dists, inds = nbrs.kneighbors(Xtest_local)
end_time = time.time()
test_time = (end_time - start_time)
print("Testing time: %f" % test_time)
# store results
if algorithm not in results.keys():
results[algorithm] = {}
results[algorithm][n_test] = {
'train': train_time,
'test': test_time,
'opt_tree_depth': opt_tree_depth
}
def run_algorithm(n_test_local, leaf_size=30, algorithm="kd_tree"):
print("----------------------------------------------------------------------")
print(("\n\nRunning %s for n_test=%i ...\n" % (algorithm, n_test_local)))
print("----------------------------------------------------------------------")
Xtest_local = Xtest[:n_test_local, :]
# instantiate model
if algorithm == "kd_tree":
nbrs = NearestNeighbors(n_neighbors=n_neighbors,
algorithm=algorithm,
n_jobs=n_jobs,
leaf_size=leaf_size,
verbose=verbose)
else:
nbrs = NearestNeighborsSKLEARN(n_neighbors=n_neighbors,
algorithm="kd_tree",
n_jobs=n_jobs,
leaf_size=leaf_size,
)
# train model
start_time = time.time()
nbrs.fit(Xtrain)
end_time = time.time()
train_time = (end_time - start_time)
print(("Fitting time: %f" % train_time))
# apply model (testing phase)
start_time = time.time()
_, _ = nbrs.kneighbors(Xtest_local)
end_time = time.time()
test_time = (end_time - start_time)
print(("Testing time: %f" % test_time))
return train_time, test_time
def buildGraph(ip):
"""Builds the knn grap with intial params.
params:
------
ip: initial params
return:
------
graph: graph object of Graph
"""
# find the nearest neighbors on the gpu
start = time()
nbrs = NearestNeighbors(n_neighbors=ip.k+1, algorithm="buffer_kd_tree", tree_depth=9, plat_dev_ids={0:[0]})
nbrs.fit(ip.signal)
dists, inds = nbrs.kneighbors(ip.signal)
dists_gpu = dists
dists_gpu = dists_gpu[0:,1:]
dists_gpu = unroll(dists_gpu)
dists_gpu = dists_gpu.astype('float32')
ngbrs_gpu = inds
ngbrs_gpu = ngbrs_gpu[0:,1:]
ngbrs_gpu = unroll(ngbrs_gpu)
ngbrs_gpu = ngbrs_gpu.astype('int32')
k = ip.k
scale = ip.sigma
n, chnl = ip.signal.shape
# now build the graph using those nns using gpu
platform = cl.get_platforms()[0]
print(platform)
device = platform.get_devices()[0]
print(device)
context = cl.Context([device])
print(context)
program = cl.Program(context, open(mywf).read()).build()
print(program)
queue = cl.CommandQueue(context)
print(queue)
# create the buffers on the device, intensity, nbgrs, weights
mem_flags = cl.mem_flags
dists_buf = cl.Buffer(context, mem_flags.READ_ONLY | mem_flags.COPY_HOST_PTR,hostbuf=dists_gpu)
weight_vec = np.ndarray(shape=(n*k,), dtype=np.float32)
weight_buf = cl.Buffer(context, mem_flags.WRITE_ONLY, weight_vec.nbytes)
# run the kernel to compute the weights
program.compute_weights(queue, (n*k,), None, dists_buf, weight_buf, np.int32(k), np.float32(scale))
queue.finish()
# copy the weihts to the host memory
cl.enqueue_copy(queue, weight_vec, weight_buf)
queue.finish()
end = time() - start
print('total time taken by the gpu python:', end)
# save the graph
graph = Graph(weight_vec,ngbrs_gpu,k)
return graph
ap.add_argument("-s",
"--s",
type=float,
required=True,
help="scale value in the graph")
args = vars(ap.parse_args())
position = numpy.load(args["position"])
position = position['position']
position = position.astype('float32')
print(position.shape)
# find the nearest neighbors on the gpu
start = time()
nbrs = NearestNeighbors(n_neighbors=args["k"] + 1,
algorithm="buffer_kd_tree",
tree_depth=9,
plat_dev_ids={0: [0]}) # use the arg parser here
nbrs.fit(position)
dists, inds = nbrs.kneighbors(position)
# now build the graph using those nns using gpu
platform = cl.get_platforms()[0]
print(platform)
device = platform.get_devices()[0]
print(device)
context = cl.Context([device])
print(context)
program = cl.Program(context, open("kernels.cl").read()).build()
def _conduct_tree_depths_comparison(params,
Xtrain,
Xtest,
target="test",
tree_depths=None,
verbose=1):
runtimes = {}
model = NearestNeighbors(**params)
if target == "test":
for tree_depth in tree_depths:
#model = copy.deepcopy(model)
model.tree_depth = tree_depth
model.fit(Xtrain)
start = time.time()
model.kneighbors(Xtest)
end = time.time()
if model.verbose:
print("tree_depth %i -> %f" % (tree_depth, end - start))
runtimes[tree_depth] = end - start
elif target == "train":
for tree_depth in tree_depths:
#model = copy.deepcopy(model)
model.tree_depth = tree_depth
start = time.time()
model.fit(Xtrain)
end = time.time()
if model.verbose:
print("tree_depth %i -> %f" % (tree_depth, end - start))
runtimes[tree_depth] = end - start
elif target == "both":
for tree_depth in tree_depths:
#model = copy.deepcopy(model)
model.tree_depth = tree_depth
start = time.time()
model.fit(Xtrain)
model.kneighbors(Xtest)
end = time.time()
if verbose > 0:
print("tree_depth %i -> %f" % (tree_depth, end - start))
runtimes[tree_depth] = end - start
else:
raise Exception("Unknown target: " + unicode(target))
return min(runtimes, key=runtimes.get)
def buildGraph(ip, dev=0):
"""Builds the knn grap with intial params.
params:
------
ip: initial params
return:
------
graph: graph object of Graph
"""
start = time()
nbrs = NearestNeighbors(n_neighbors = ip.k + 1, algorithm="buffer_kd_tree", tree_depth=9, plat_dev_ids={0:[0]})
nbrs.fit(ip.position)
dists, inds = nbrs.kneighbors(ip.position)
print("success") if bool_1 else print()
# now build the graph using those nns using gpu
platform = cl.get_platforms()[0]
print(platform)
device = platform.get_devices()[dev]
print(device)
context = cl.Context([device])
print(context)
program = cl.Program(context, open(mywf).read()).build()
print(program)
queue = cl.CommandQueue(context)
print(queue)
# define the input here which is the ndbrs gpu
ngbrs_gpu = inds
ngbrs_gpu = ngbrs_gpu[0:,1:]
ngbrs_gpu = unroll(ngbrs_gpu)
ngbrs_gpu = ngbrs_gpu.astype('int32')
# define the second input here which is the signal levels
signal = ip.signal
n, chnl = signal.shape
signal = np.reshape(signal,(n*chnl,),order='F')
signal = signal.astype('float32')
print("signal",signal.shape) if bool_1 else print()
k = ip.k
print("n is :", n) if bool_1 else print()
scale = ip.sigma
# create the buffers on the device, intensity, nbgrs, weights
mem_flags = cl.mem_flags
ngbrs_buf = cl.Buffer(context, mem_flags.READ_ONLY | mem_flags.COPY_HOST_PTR,hostbuf=ngbrs_gpu)
signal_buf = cl.Buffer(context, mem_flags.READ_ONLY | mem_flags.COPY_HOST_PTR, hostbuf=signal)
weight_vec = np.ndarray(shape=(n*k,), dtype=np.float32)
weight_buf = cl.Buffer(context, mem_flags.WRITE_ONLY, weight_vec.nbytes)
# run the kernel to compute the weights
program.compute_weights(queue, (n,), None, signal_buf, ngbrs_buf, weight_buf, np.int32(k), np.float32(scale), np.int32(chnl))
queue.finish() #OT
# copy the weihts to the host memory
cl.enqueue_copy(queue, weight_vec, weight_buf)
end = time() - start
print('total time taken by the gpu python:', end) if bool_1 else print()
# save the graph
graph = Graph(weight_vec,ngbrs_gpu,k)
return graph