def __init__(self,
total_process,
mi,
batch_template,
cb_remote_initialize=None,
cb_remote_batch_process=None,
args=None):
''' Initialize `SharedData` class with a few hooks
Args:
total_process: number of processes
mi: ModelInterface
batch_template:
cb_remote_initialize: Callbacks for remote Initialization
cb_remote_batch_process: Callbacks for remote process
args: additional arguments
'''
self.server = ParameterServer(total_process)
self.cb_remote_initialize = cb_remote_initialize
self.cb_remote_batch_process = cb_remote_batch_process
self.args = args
# def get_gpu_id(i): return i + 1
def get_gpu_id(i):
return 0
# Share only training batches.
shared_batches = []
cvs_send = []
cvs_recv = []
qs = []
for i in range(total_process - 1):
# gpu_id = get_gpu_id(i)
# shared_batches.append(
# cpu2gpu(all_batches[train_idx][0], gpu=gpu_id))
shared_batches.append(utils_elf.pin_clone(batch_template))
qs.append(mp.Queue(1))
qs[-1].put(shared_batches[i])
cvs_send.append(Cond())
cvs_recv.append(Cond())
self.cvs_send = cvs_send
self.cvs_recv = cvs_recv
self.shared_batches = shared_batches
self.qs = qs
self.b = mp.Barrier(total_process)
self.optimizers = [
mp.process(target=self.process_main, args=(i, get_gpu_id(i)))
for i in range(total_process - 1)
]
for optimizer in self.optimizers:
optimizer.start()
# Wait until all models have received the shared memory.
self.b.wait()
self.server.server_send_model(mi)
def __init__(self, total_process, mi, batch_template,
cb_remote_initialize=None,
cb_remote_batch_process=None,
args=None):
''' Initialize `SharedData` class with a few hooks
Args:
total_process: number of processes
mi: ModelInterface
batch_template:
cb_remote_initialize: Callbacks for remote Initialization
cb_remote_batch_process: Callbacks for remote process
args: additional arguments
'''
self.server = ParameterServer(total_process)
self.cb_remote_initialize = cb_remote_initialize
self.cb_remote_batch_process = cb_remote_batch_process
self.args = args
# def get_gpu_id(i): return i + 1
def get_gpu_id(i): return 0
# Share only training batches.
shared_batches = []
cvs_send = []
cvs_recv = []
qs = []
for i in range(total_process - 1):
# gpu_id = get_gpu_id(i)
# shared_batches.append(
# cpu2gpu(all_batches[train_idx][0], gpu=gpu_id))
shared_batches.append(utils_elf.pin_clone(batch_template))
qs.append(mp.Queue(1))
qs[-1].put(shared_batches[i])
cvs_send.append(Cond())
cvs_recv.append(Cond())
self.cvs_send = cvs_send
self.cvs_recv = cvs_recv
self.shared_batches = shared_batches
self.qs = qs
self.b = mp.Barrier(total_process)
self.optimizers = [
mp.process(
target=self.process_main, args=(
i, get_gpu_id(i))) for i in range(
total_process - 1)]
for optimizer in self.optimizers:
optimizer.start()
# Wait until all models have received the shared memory.
self.b.wait()
self.server.server_send_model(mi)
def __init__(self,
total_process,
mi,
batch_template,
cb_remote_initialize=None,
cb_remote_batch_process=None,
args=None):
self.server = ParameterServer(total_process)
self.cb_remote_initialize = cb_remote_initialize
self.cb_remote_batch_process = cb_remote_batch_process
self.args = args
#def get_gpu_id(i): return i + 1
def get_gpu_id(i):
return 0
# Share only training batches.
shared_batches = []
cvs_send = []
cvs_recv = []
qs = []
for i in range(total_process - 1):
# gpu_id = get_gpu_id(i)
# shared_batches.append(cpu2gpu(all_batches[train_idx][0], gpu=gpu_id))
shared_batches.append(utils_elf.pin_clone(batch_template))
qs.append(mp.Queue(1))
qs[-1].put(shared_batches[i])
cvs_send.append(Cond())
cvs_recv.append(Cond())
self.cvs_send = cvs_send
self.cvs_recv = cvs_recv
self.shared_batches = shared_batches
self.qs = qs
self.b = mp.Barrier(total_process)
self.optimizers = [
mp.Process(target=self.process_main, args=(i, get_gpu_id(i)))
for i in range(total_process - 1)
]
for optimizer in self.optimizers:
optimizer.start()
# Wait until all models have received the shared memory.
self.b.wait()
self.server.server_send_model(mi)
def __init__(self, total_process, mi, batch_template,
cb_remote_initialize=None,
cb_remote_batch_process=None,
args=None):
self.server = ParameterServer(total_process)
self.cb_remote_initialize = cb_remote_initialize
self.cb_remote_batch_process = cb_remote_batch_process
self.args = args
#def get_gpu_id(i): return i + 1
def get_gpu_id(i): return 0
# Share only training batches.
shared_batches = []
cvs_send = []
cvs_recv = []
qs = []
for i in range(total_process - 1):
# gpu_id = get_gpu_id(i)
# shared_batches.append(cpu2gpu(all_batches[train_idx][0], gpu=gpu_id))
shared_batches.append(utils_elf.pin_clone(batch_template))
qs.append(mp.Queue(1))
qs[-1].put(shared_batches[i])
cvs_send.append(Cond())
cvs_recv.append(Cond())
self.cvs_send = cvs_send
self.cvs_recv = cvs_recv
self.shared_batches = shared_batches
self.qs = qs
self.b = mp.Barrier(total_process)
self.optimizers = [mp.Process(target=self.process_main, args=(i, get_gpu_id(i))) for i in range(total_process - 1)]
for optimizer in self.optimizers: optimizer.start()
# Wait until all models have received the shared memory.
self.b.wait()
self.server.server_send_model(mi)