def main():
ray.init(object_store_memory=4000000000)
n_data = 80
test_data = np.random.uniform(-1, 1, (n_data, 224, 224, 3))
model = dKeras(ResNet50,
weights='imagenet',
init_ray=False,
wait_for_workers=True,
n_workers=4)
start_time = time.time()
preds = model.predict(test_data)
elapsed = time.time() - start_time
model.close()
time.sleep(2)
#
# model = ResNet50(weights='imagenet')
# preds2 = model.predict(test_data)
#
# # for i in range(n_data):
# # for j,n in enumerate(preds[i]):
# # print(n, preds2[i][j])
# # print('-'*80)
print("Time elapsed: {}\nFPS: {}".format(elapsed, n_data / elapsed))
def main():
"""
Test out using multiple models at the same time, this case is running
ResNet50 and MobileNet
"""
ray.init()
n_data = 20
model1 = dKeras(ResNet50,
weights='imagenet',
wait_for_workers=True,
n_workers=3)
model2 = dKeras(MobileNet,
weights='imagenet',
wait_for_workers=True,
n_workers=3)
test_data = np.random.uniform(-1, 1, (n_data, 224, 224, 3))
model1.predict(test_data)
model2.predict(test_data)
model1.close()
model2.close()
def main():
n_data = 80
test_data = np.random.uniform(-1, 1, (n_data, 224, 224, 3))
model = dKeras(ResNet50,
weights='imagenet',
wait_for_workers=True,
n_workers=4)
start_time = time.time()
preds = model.predict(test_data,
use_multiprocessing=True,
multiprocessing_workers=4)
elapsed = time.time() - start_time
model.close()
def main():
ray.init()
n_data = 20
model = dKeras(ResNet50,
weights='imagenet',
wait_for_workers=True,
n_workers=3)
for i in range(5):
test_data = np.random.uniform(-1, 1, (n_data, 224, 224, 3))
start_time = time.time()
preds_1 = model.predict(test_data)
elapsed = time.time() - start_time
print(np.asarray(preds_1).shape)
print("Time elapsed: {}\nFPS: {}".format(elapsed, n_data / elapsed))
model.close()
from tensorflow.keras.applications import ResNet50
from dkeras import dKeras
from dkeras.utils.qsub_functions import init_pbs_ray
import numpy as np
import time
import ray
init_pbs_ray()
print(ray.nodes())
data = np.random.uniform(-1, 1, (10000, 224, 224, 3))
start_time = time.time()
model = dKeras(ResNet50, init_ray=False, wait_for_workers=True, n_workers=500)
elapsed = time.time() - start_time
print("Workers initialized after {}".format(elapsed))
start_time = time.time()
preds = model.predict(data)
elapsed = time.time() - start_time
print("Preds after {}".format(elapsed))
model1.predict(test_data)
model2.predict(test_data)
model1.close()
model2.close()
if __name__ == "__main__":
main()
ray.init()
n_data = 20
model1 = dKeras(ResNet50,
weights='imagenet',
wait_for_workers=True,
n_workers=3)
model2 = dKeras(MobileNet,
weights='imagenet',
wait_for_workers=True,
n_workers=3)
test_data = np.random.uniform(-1, 1, (n_data, 224, 224, 3))
model1.predict(test_data)
model2.predict(test_data)
model1.close()
model2.close()
def main():
model_names = {
'densenet121': DenseNet121,
'densenet169': DenseNet169,
'densenet201': DenseNet201,
'inception_v3': InceptionV3,
'inception_resnet_v2': InceptionResNetV2,
'mobilenet': MobileNet,
'mobilenet_v2': MobileNetV2,
'nasnet_large': NASNetLarge,
'nasnet_mobile': NASNetMobile,
'resnet50': ResNet50,
'vgg16': VGG16,
'vgg19': VGG19,
'xception': Xception
}
# os.environ["KERAS_BACKEND"] = "plaidml.keras.backend"
parser = argparse.ArgumentParser()
parser.add_argument("--n_data",
help="Number of fake datapoints",
default=1000,
type=int)
parser.add_argument("--n_workers",
help="Number of Ray workers",
default=5,
type=int)
parser.add_argument("--test",
help="0: Local, 1: dKeras",
default=1,
type=int)
parser.add_argument("--search",
help="True or False, Find best n_workers",
default=False,
type=bool)
parser.add_argument("--search-pool",
help="n workers to search from, separated by commas",
default='2,3,5,10,20,25,50,60,100',
type=str)
parser.add_argument("--model",
help="Model options: {}".format(model_names.keys()),
default='resnet50',
type=str)
args = parser.parse_args()
n_workers = args.n_workers
test_type = args.test
n_data = args.n_data
model_name = args.model
use_search = args.search
search_pool = args.search_pool
try:
search_pool = list(map(int, search_pool.split(',')))
except TypeError:
raise UserWarning("Search pool arg must be int separated by commas")
if not (model_name in model_names.keys()):
raise UserWarning("Model name not found: {}, options: {}".format(
model_name, model_names))
test_data = np.float16(np.random.uniform(-1, 1, (n_data, 224, 224, 3)))
if test_type == 0:
model = model_names[model_name]()
start_time = time.time()
preds = model.predict(test_data)
elapsed = time.time() - start_time
print("Time elapsed: {}\nFPS: {}".format(elapsed, n_data / elapsed))
elif test_type == 1:
if use_search:
results = {}
best_time = np.inf
best_n_workers = -1
for n in search_pool:
model = dKeras(model_names[model_name],
wait_for_workers=True,
n_workers=n)
print("Workers are ready")
start_time = time.time()
preds = model.predict(test_data)
elapsed = time.time() - start_time
time.sleep(3)
if elapsed < best_time:
best_time = elapsed
best_n_workers = n
results[str(n)] = elapsed
model.close()
print('{}\nN\tElapsed Time'.format('=' * 80))
for k in results.keys():
print("{}\t{}".format(k, results[k]))
print(
"{}\nTests completed:\n\tBest N workers: {}\t FPS: {}".format(
'=' * 80, best_n_workers, n_data / best_time))
else:
model = dKeras(model_names[model_name],
wait_for_workers=True,
n_workers=n_workers)
start_time = time.time()
preds = model.predict(test_data)
elapsed = time.time() - start_time
model.close()
time.sleep(3)
print("Time elapsed: {}\nFPS: {}".format(elapsed,
n_data / elapsed))
