def test_distributed_basic(self):
checker = acp._get_checker()
fs = HDFSClient(checker.hdfs_home, None)
fs.delete(checker.hdfs_checkpoint_path)
self._reset_generator()
logger.info("begin test_distributed_basic")
fs = LocalFS()
save_dir = "./run_save_0"
fs.delete(save_dir)
#basic
exe, main_prog, startup_prog = self._generate()
compiled, data_loader, optimizer, loss, image, label = \
self._init_env(exe, main_prog, startup_prog, minimize=False)
#fleet
os.environ["TRAINING_ROLE"] = "TRAINER"
os.environ["PADDLE_TRAINER_ID"] = "0"
os.environ["PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:6070"
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
with fluid.program_guard(main_prog, startup_prog):
dist_optimizer = fleet.distributed_optimizer(optimizer)
dist_optimizer.minimize(loss)
exe.run(startup_prog)
o = None
i = 0
name = None
for i in acp.train_epoch_range(3, 0):
o = acp._get_train_epoch_range()
name = o.name
logger.info("_run_save_0 name:{} epoch_no:{}".format(o.name, i))
for data in data_loader():
fetch = exe.run(fleet.main_program,
feed=data,
fetch_list=[loss])
self.assertEqual(len(o._exe_status), 1)
o = acp._get_train_epoch_range()
assert o == None, "now train epoch must not exits now"
self.assertEqual(i, 2)
fs.delete(save_dir)
logger.info("end test_distributed_basic")
def _not_use_train(self):
logger.info("begin _not_use_train")
exe, main_prog, startup_prog = self._generate()
compiled, data_loader, optimizer, loss, image, label = \
self._init_env(exe, main_prog, startup_prog)
epochs = []
for i in acp.train_epoch_range(3, 0):
epochs.append(i)
for data in data_loader():
fetch = exe.run(compiled, feed=data, fetch_list=[loss])
self.assertEqual(epochs, [0, 1, 2])
logger.info("end _not_use_train")
def test_multiple(self):
checker = acp._get_checker()
fs = HDFSClient(checker.hdfs_home, None)
fs.delete(checker.hdfs_checkpoint_path)
self._reset_generator()
logger.info("begin test_multiple")
fs = LocalFS()
save_dir = "./run_save_0"
fs.delete(save_dir)
exe, main_prog1, startup_prog1 = self._generate()
_, main_prog2, startup_prog2 = self._generate()
compiled1, data_loader1, optimizer1, loss1, image1, label1 = \
self._init_env(exe, main_prog1, startup_prog1)
compiled2, data_loader2, optimizer2, loss2, image2, label2 = \
self._init_env(exe, main_prog2, startup_prog2)
o = None
epochs = []
for i in acp.train_epoch_range(3, 0):
for data in data_loader1():
fetch = exe.run(compiled1, feed=data, fetch_list=[loss1])
for data in data_loader2():
fetch = exe.run(compiled2, feed=data, fetch_list=[loss2])
o = acp._get_train_epoch_range()
self.assertEqual(len(o._exe_status), 2)
print(o._exe_status)
epochs.append(i)
o = acp._get_train_epoch_range()
self.assertTrue(o == None, "now train epoch must not exits now")
self.assertEqual(i, 2)
self.assertEqual(epochs, [0, 1, 2])
fs.delete(save_dir)
logger.info("end test_multiple")
def _run_load_0(self, break_epoch_no=None):
logger.info("begin _run_load_0")
exe, main_prog, startup_prog = self._generate()
fs = LocalFS()
save_dir = "./run_load_0"
fs.delete(save_dir)
compiled, data_loader, optimizer, loss, image, label = self._init_env(
exe, main_prog, startup_prog)
o = None
i = 0
check = False
epochs = []
for i in acp.train_epoch_range(3, 0):
epochs.append(i)
for data in data_loader():
fetch = exe.run(compiled, feed=data, fetch_list=[loss])
o = acp._get_train_epoch_range()
self.assertTrue(o == None, "now train epoch must not exits now")
self.assertEqual(i, 2)
if break_epoch_no is not None:
if break_epoch_no == 0:
self.assertEqual(epochs, [0, 1, 2])
elif break_epoch_no == 1:
self.assertEqual(epochs, [1, 2])
elif break_epoch_no == 2:
self.assertEqual(epochs, [2])
else:
self.assertEqual(epochs, [2])
fs.delete(save_dir)
logger.info("begin _run_load_0")
def _run_save_0(self, break_epoch_no=None):
logger.info("begin _run_save_0")
fs = LocalFS()
save_dir = "./run_save_0"
fs.delete(save_dir)
exe, main_prog, startup_prog = self._generate()
compiled, data_loader, optimizer, loss, image, label = \
self._init_env(exe, main_prog, startup_prog)
o = None
i = 0
name = None
for i in acp.train_epoch_range(3, 0):
o = acp._get_train_epoch_range()
name = o.name
for data in data_loader():
fetch = exe.run(compiled, feed=data, fetch_list=[loss])
self.assertEqual(len(o._exe_status), 1)
if break_epoch_no is not None:
if i == break_epoch_no:
break
o = acp._get_train_epoch_range()
assert o == None, "now train epoch must not exits now"
if break_epoch_no is None:
self.assertEqual(i, 2)
else:
self.assertEqual(i, break_epoch_no)
fs.delete(save_dir)
logger.info("end _run_save_0")