def train(self, place, iters, feed, use_cinn=False, seed=1234):
np.random.seed(seed)
paddle.seed(seed)
if paddle.is_compiled_with_cuda():
paddle.set_flags({'FLAGS_cudnn_deterministic': 1})
set_cinn_flag(use_cinn)
startup_program = paddle.static.Program()
main_program = paddle.static.Program()
loss = self.build_program(main_program, startup_program)
exe = paddle.static.Executor(place)
compiled_prog = paddle.static.CompiledProgram(
main_program).with_data_parallel(loss_name=loss.name)
loss_vals = []
scope = paddle.static.Scope()
with paddle.static.scope_guard(scope):
exe.run(startup_program)
for step in range(iters):
loss_v = exe.run(compiled_prog,
feed=feed[step],
fetch_list=[loss],
return_numpy=True)
loss_vals.append(loss_v[0][0])
return loss_vals
def train(dot_save_dir, prefix, seed=1234):
np.random.seed(seed)
paddle.seed(seed)
if paddle.is_compiled_with_cuda():
paddle.set_flags({'FLAGS_cudnn_deterministic': 1})
startup_program = paddle.static.Program()
main_program = paddle.static.Program()
img, label, loss = build_program(main_program, startup_program)
place = paddle.CUDAPlace(
0) if paddle.is_compiled_with_cuda() else paddle.CPUPlace()
exe = paddle.static.Executor(place)
exe.run(startup_program)
build_strategy = paddle.static.BuildStrategy()
build_strategy.debug_graphviz_path = os.path.join(dot_save_dir, prefix)
compiled_program = paddle.static.CompiledProgram(
main_program, build_strategy).with_data_parallel(loss_name=loss.name)
iters = 100
feed = rand_data(img.name, label.name, iters)
loss_values = []
for step in range(iters):
loss_v = exe.run(compiled_program,
feed=feed[step],
fetch_list=[loss],
return_merged=False)
loss_values.append(loss_v[0][0][0])
return loss_values
def set_cinn_flag(val):
cinn_compiled = False
try:
paddle.set_flags({'FLAGS_use_cinn': val})
cinn_compiled = True
except ValueError:
logger.warning("The used paddle is not compiled with CINN.")
return cinn_compiled
def setUpClass(cls):
if not paddle.is_compiled_with_cuda():
return
paddle.enable_static()
paddle.set_flags({'FLAGS_cudnn_deterministic': True})
_clip_by_global_norm_using_mp_type(True)
fleet.init(role_maker=get_role_maker())
def setUp(self):
if can_use_cuda_graph():
paddle.set_flags({
'FLAGS_allocator_strategy': 'auto_growth',
'FLAGS_sync_nccl_allreduce': False,
'FLAGS_cudnn_deterministic': True,
'FLAGS_use_stream_safe_cuda_allocator': False,
})
def init(self):
if paddle.is_compiled_with_cuda():
paddle.set_flags({'FLAGS_cudnn_deterministic': 1})
self.rtol = 1e-6
self.atol = 1e-8
rank = paddle.distributed.get_rank()
paddle.seed(rank + 2021)
random.seed(rank + 2021)
np.random.seed(rank + 2021)
def setUp(self):
paddle.enable_static()
if paddle.is_compiled_with_cuda():
paddle.set_flags({'FLAGS_cudnn_deterministic': 1})
seed = int(os.environ.get('SEED', -1))
if seed <= 0:
seed = np.random.randint(low=1, high=1000000, size=[1])[0]
os.environ['SEED'] = str(seed)
self.seed = seed
paddle.seed(self.seed)
self.rtol = 1e-5
self.atol = 1e-8
self.equal_nan = False
self.init()
def test_compute_type_fp16_nan(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
with fluid.dygraph.guard(place):
paddle.set_flags(
{'FLAGS_gemm_use_half_precision_compute_type': True})
input_x = np.random.random([2, 8, 16]).astype("float16")
input_y = np.random.random([2, 16, 8]).astype("float16")
for i in range(0, 16, 2):
input_x[:, :, i] += 60000
input_x[:, :, i + 1] -= 60000
input_y[:, :, :] = 1.5
x = paddle.to_tensor(input_x)
y = paddle.to_tensor(input_y)
result = paddle.matmul(x, y)
result_np = np.matmul(input_x, input_y)
self.assertFalse(
paddle.isfinite(result)[0, 0, 0]) # contains nan/inf
self.assertTrue(np.isfinite(result_np)[0, 0, 0])
paddle.set_flags(
{'FLAGS_gemm_use_half_precision_compute_type': False})
def use_fast_math(enabled):
paddle.set_flags({'FLAGS_use_fast_math': enabled})
def set_flags(self, enable_autotune):
if paddle.is_compiled_with_cuda():
if enable_autotune:
paddle.set_flags({'FLAGS_conv_workspace_size_limit': -1})
else:
paddle.set_flags({'FLAGS_conv_workspace_size_limit': 512})
def setUp(self):
paddle.enable_static()
if paddle.is_compiled_with_cuda():
paddle.set_flags({'FLAGS_cudnn_deterministic': 1})
def setUp(self):
paddle.set_flags({'FLAGS_eager_delete_tensor_gb': 0.0})
