def test_dedup_input(self):
os.environ['CGO'] = 'true'
lp, models = self._build_logical_with_mnist(3)
opt = DedupInputOptimizer()
opt.convert(lp)
with open('dedup_logical_graph.json', 'r') as fp:
correct_dump = fp.readlines()
lp_dump = lp.logical_graph._dump()
self.assertTrue(correct_dump[0] == json.dumps(lp_dump))
advisor = RetiariiAdvisor()
cgo = CGOExecutionEngine()
phy_models = cgo._assemble(lp)
self.assertTrue(len(phy_models) == 1)
# logging.info(phy_models[0][0]._dump())
# script=model_to_pytorch_script(phy_models[0][0], placement = phy_models[0][1])
# logging.info(script)
# with open('generated/debug_dedup_input.py', 'w') as fp:
# fp.write(script)
# sys.path.insert(0, 'generated')
# multi_model = import_('debug_dedup_input.logical_0')
# trainer = PyTorchMultiModelTrainer(
# multi_model(), phy_models[0][0].training_config.kwargs
# )
# trainer.fit()
advisor.stopping = True
advisor.default_worker.join()
advisor.assessor_worker.join()
def test_submit_models(self):
_reset()
nni.retiarii.debug_configs.framework = 'pytorch'
os.makedirs('generated', exist_ok=True)
import nni.runtime.platform.test as tt
protocol._set_out_file(
open('generated/debug_protocol_out_file.py', 'wb'))
protocol._set_in_file(
open('generated/debug_protocol_out_file.py', 'rb'))
models = _load_mnist(2)
advisor = RetiariiAdvisor('ws://_unittest_placeholder_')
advisor._channel = protocol.LegacyCommandChannel()
advisor.default_worker.start()
advisor.assessor_worker.start()
remote = RemoteConfig(machine_list=[])
remote.machine_list.append(
RemoteMachineConfig(host='test', gpu_indices=[0, 1, 2, 3]))
cgo_engine = CGOExecutionEngine(training_service=remote,
batch_waiting_time=0)
set_execution_engine(cgo_engine)
submit_models(*models)
time.sleep(3)
if torch.cuda.is_available() and torch.cuda.device_count() >= 2:
cmd, data = protocol.receive()
params = nni.load(data)
tt.init_params(params)
trial_thread = threading.Thread(
target=CGOExecutionEngine.trial_execute_graph)
trial_thread.start()
last_metric = None
while True:
time.sleep(1)
if tt._last_metric:
metric = tt.get_last_metric()
if metric == last_metric:
continue
if 'value' in metric:
metric['value'] = json.dumps(metric['value'])
advisor.handle_report_metric_data(metric)
last_metric = metric
if not trial_thread.is_alive():
trial_thread.join()
break
trial_thread.join()
advisor.stopping = True
advisor.default_worker.join()
advisor.assessor_worker.join()
cgo_engine.join()
def test_submit_models(self):
_reset()
nni.retiarii.debug_configs.framework = 'pytorch'
os.makedirs('generated', exist_ok=True)
from nni.runtime import protocol
import nni.runtime.platform.test as tt
protocol._set_out_file(
open('generated/debug_protocol_out_file.py', 'wb'))
protocol._set_in_file(
open('generated/debug_protocol_out_file.py', 'rb'))
models = _load_mnist(2)
advisor = RetiariiAdvisor()
cgo_engine = CGOExecutionEngine(devices=[
GPUDevice("test", 0),
GPUDevice("test", 1),
GPUDevice("test", 2),
GPUDevice("test", 3)
],
batch_waiting_time=0)
set_execution_engine(cgo_engine)
submit_models(*models)
time.sleep(3)
if torch.cuda.is_available() and torch.cuda.device_count() >= 2:
cmd, data = protocol.receive()
params = nni.load(data)
tt.init_params(params)
trial_thread = threading.Thread(
target=CGOExecutionEngine.trial_execute_graph)
trial_thread.start()
last_metric = None
while True:
time.sleep(1)
if tt._last_metric:
metric = tt.get_last_metric()
if metric == last_metric:
continue
if 'value' in metric:
metric['value'] = json.dumps(metric['value'])
advisor.handle_report_metric_data(metric)
last_metric = metric
if not trial_thread.is_alive():
trial_thread.join()
break
trial_thread.join()
advisor.stopping = True
advisor.default_worker.join()
advisor.assessor_worker.join()
cgo_engine.join()
def test_dedup_input_two_devices(self):
_reset()
lp, models = self._build_logical_with_mnist(3)
opt = DedupInputOptimizer()
opt.convert(lp)
advisor = RetiariiAdvisor()
available_devices = [GPUDevice("test", 0), GPUDevice("test", 1)]
cgo = CGOExecutionEngine(devices=available_devices,
batch_waiting_time=0)
phy_models = cgo._assemble(lp)
self.assertTrue(len(phy_models) == 2)
advisor.stopping = True
advisor.default_worker.join()
advisor.assessor_worker.join()
cgo.join()
def test_dedup_input_two_devices(self):
_reset()
lp, models = self._build_logical_with_mnist(3)
opt = DedupInputOptimizer()
opt.convert(lp)
advisor = RetiariiAdvisor('ws://_unittest_placeholder_')
advisor._channel = protocol.LegacyCommandChannel()
advisor.default_worker.start()
advisor.assessor_worker.start()
remote = RemoteConfig(machine_list=[])
remote.machine_list.append(
RemoteMachineConfig(host='test', gpu_indices=[0, 1]))
cgo = CGOExecutionEngine(training_service=remote, batch_waiting_time=0)
phy_models = cgo._assemble(lp)
self.assertTrue(len(phy_models) == 2)
advisor.stopping = True
advisor.default_worker.join()
advisor.assessor_worker.join()
cgo.join()
def test_submit_models(self):
os.environ['CGO'] = 'true'
os.makedirs('generated', exist_ok=True)
from nni.runtime import protocol, platform
import nni.runtime.platform.test as tt
protocol._out_file = open('generated/debug_protocol_out_file.py', 'wb')
protocol._in_file = open('generated/debug_protocol_out_file.py', 'rb')
models = _load_mnist(2)
advisor = RetiariiAdvisor()
submit_models(*models)
if torch.cuda.is_available() and torch.cuda.device_count() >= 2:
cmd, data = protocol.receive()
params = json.loads(data)
params['parameters']['training_kwargs']['max_steps'] = 100
tt.init_params(params)
trial_thread = threading.Thread(
target=CGOExecutionEngine.trial_execute_graph())
trial_thread.start()
last_metric = None
while True:
time.sleep(1)
if tt._last_metric:
metric = tt.get_last_metric()
if metric == last_metric:
continue
advisor.handle_report_metric_data(metric)
last_metric = metric
if not trial_thread.is_alive():
break
trial_thread.join()
advisor.stopping = True
advisor.default_worker.join()
advisor.assessor_worker.join()