def main(argv):
## Formating the argv from reseource manager
for i in range(len(argv)):
argv[i] = str(argv[i]).replace('[', '').replace(']',
'').replace(',', '')
## Formatiing the workers IP to build
## a tensorflow cluster in executor class
temp_workers = []
for i in range(len(argv) - 3):
temp_workers.append(argv[i + 3])
temp_workers = ','.join(temp_workers)
## 1) Define stacked layers funtion activation, number of layers and their neurons
layer_1 = [Linear(1300, 4)]
## 2) Select the neural network architecture and pass the hyper-parameters
cnn_model = CustomGraph(input_size_1=1300,
input_size_2=1,
output_size=4,
loss=CrossEntropy,
optimizer=Adam(lr=0.0001),
layers=layer_1)
## 3) Dataset configurations for splitting, batching and target selection
data_config = Batching(dataset_name="ECG",
valid_size=0.1,
test_size=0.1,
batch_size=50,
devices_number=4)
## 4) Select the computational platform and pass the DNN and Dataset configurations
if ON_ASTRO:
testbed_path = "/home/mpiuser/cloud/0/diagnosenet/samples/1_customgraph/testbed"
else:
testbed_path = "/home/mpiuser/cloud/diagnosenet/samples/1_customgraph/testbed"
platform = Distibuted_GRPC(model=cnn_model,
datamanager=data_config,
monitor=enerGyPU(testbed_path=testbed_path,
machine_type="arm",
file_path=workspace_path),
max_epochs=20,
early_stopping=3,
ip_ps=argv[2],
ip_workers=temp_workers)
## 5) Uses the platform modes for training in an efficient way
if ON_ASTRO:
dataset_path = "/home/mpiuser/cloud/0/diagnosenet/samples/1_customgraph//dataset/"
else:
dataset_path = "/home/mpiuser/cloud/diagnosenet/samples/1_customgraph/dataset"
platform.asynchronous_training(dataset_name="ECG",
dataset_path=dataset_path,
inputs_name="xdata.npy",
targets_name="undim-ydata.npy",
job_name=argv[0],
task_index=int(argv[1]))
optimizer=Adam(lr=0.001),
dropout=0.8)
## 3) Dataset configurations for splitting, batching and target selection
data_config = MultiTask(dataset_name="MCP-PMSI",
valid_size=0.05,
test_size=0.15,
batch_size=250,
target_name='Y11',
target_start=0,
target_end=14)
## 4) Select the computational platform and pass the DNN and Dataset configurations
platform = DesktopExecution(model=mlp_model_1,
datamanager=data_config,
monitor=enerGyPU(machine_type="arm",
file_path=workspace_path),
max_epochs=10,
early_stopping=5)
### Read the PMSI-Dataset using Pickle from diagnosenet.io_functions
X = IO_Functions()._read_file("dataset/patients_features.txt")
y = IO_Functions()._read_file("dataset/medical_targets_Y14.txt")
## 5) Uses the platform modes for training in an efficient way
platform.training_memory(X, y)
print("Execution Time: {}".format((time.time() - execution_start)))
####################################################
## Path for Octopus Machine: Full Representation
#path = "/data_B/datasets/drg-PACA/healthData/sandbox-FULL-W1_x1_x2_x3_x4_x5_x7_x8_Y1/1_Mining-Stage/binary_representation/"
layers=layers,
loss=CrossEntropy,
optimizer=Adam(lr=0.001),
dropout=0.8)
## 3) Dataset configurations for splitting, batching and target selection
data_config_1 = Batching(dataset_name="MCP-PMSI",
valid_size=0.05,
test_size=0.10,
devices_number=7,
batch_size=40)
## 4) Select the computational platform and pass the DNN and Dataset configurations
platform = Distibuted_MPI(model=mlp_model,
datamanager=data_config_1,
monitor=enerGyPU(machine_type="arm",
file_path=file_path),
max_epochs=20,
early_stopping=3)
## 5) Uses the platform modes for training in an efficient way
platform.asynchronous_training(dataset_name="MCP-PMSI",
dataset_path="dataset/",
inputs_name="patients_features.txt",
targets_name="medical_targets_Y14.txt",
weighting=1)
# platform.synchronous_training(dataset_name="MCP-PMSI",
# dataset_path="dataset/",
# inputs_name="patients_features.txt",
# targets_name="medical_targets_Y14.txt")
def main(argv):
## Formating the argv from reseource manager
for i in range(len(argv)):
argv[i] = str(argv[i]).replace('[', '').replace(']',
'').replace(',', '')
## Formatiing the workers IP to build
## a tensorflow cluster in executor class
temp_workers = []
for i in range(len(argv) - 3):
temp_workers.append(argv[i + 3])
temp_workers = ','.join(temp_workers)
## PMSI-ICU Dataset shapes
X_shape = 10833 #14637
y_shape = 381
Y1_shape = 14
Y2_shape = 239
Y3_shape = 5
## 1) Define stacked layers funtion activation, number of layers and their neurons
stacked_layers_1 = [
Relu(X_shape, 1024),
Relu(1024, 1024),
Linear(1024, Y1_shape)
]
stacked_layers_2 = [
Relu(X_shape, 512),
Relu(512, 512),
Relu(512, 512),
Relu(512, 512),
Linear(512, Y1_shape)
]
stacked_layers_3 = [
Relu(X_shape, 256),
Relu(256, 256),
Relu(256, 256),
Relu(256, 256),
Relu(256, 256),
Relu(256, 256),
Relu(256, 256),
Relu(256, 256),
Linear(256, Y1_shape)
]
## 2) Select the neural network architecture and pass the hyper-parameters
mlp_model = SequentialGraph(input_size=X_shape,
output_size=Y1_shape,
layers=stacked_layers_3,
loss=CrossEntropy,
optimizer=Adam(lr=0.001),
dropout=0.8)
## 3) Dataset configurations for splitting, batching and target selection
data_config_1 = Batching(dataset_name="MCP-PMSI",
valid_size=0.05,
test_size=0.10,
devices_number=4,
batch_size=10)
## 4) Select the computational platform and pass the DNN and Dataset configurations
if ON_ASTRO:
testbed_path = "/home/mpiuser/cloud/0/diagnosenet/samples/0_sequentialgraph/testbed"
else:
testbed_path = "/home/mpiuser/cloud/diagnosenet/samples/0_sequentialgraph/testbed"
platform = Distibuted_GRPC(model=mlp_model,
datamanager=data_config_1,
monitor=enerGyPU(testbed_path=testbed_path,
machine_type="arm",
file_path=workspace_path),
max_epochs=20,
early_stopping=3,
ip_ps=argv[2],
ip_workers=temp_workers)
## 5) Uses the platform modes for training in an efficient way
if ON_ASTRO:
dataset_path = "/home/mpiuser/cloud/0/diagnosenet/samples/0_sequentialgraph/dataset/"
else:
dataset_path = "/home/mpiuser/cloud/PMSI-Dataset/"
platform.asynchronous_training(dataset_name="MCP-PMSI",
dataset_path=dataset_path,
inputs_name="patients_features-full.txt",
targets_name="medical_targets-full.txt",
job_name=argv[0],
task_index=int(argv[1]))