def test_selectedrows_gradient1(self):
places = [fluid.CPUPlace()]
if core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
for place in places:
for dtype in ["float32", "float64"]:
for sort_sum_gradient in [True, False]:
paddle.disable_static(place)
fluid.set_flags(
{'FLAGS_sort_sum_gradient': sort_sum_gradient})
# grad_clip = fluid.clip.GradientClipByGlobalNorm(5.0)
input_word = np.array([[1, 2], [2, 1]]).astype('int64')
input = paddle.to_tensor(input_word)
simplenet = SimpleNet(20, 32, dtype)
adam = SGDOptimizer(learning_rate=0.001,
parameter_list=simplenet.parameters()
) # grad_clip=grad_clip
input_emb, emb = simplenet(input)
self.assertTrue(emb.weight.gradient() is None)
self.assertTrue(input_emb.gradient() is None)
input_emb.backward()
adam.minimize(input_emb)
self.assertTrue(emb.weight.gradient() is not None)
emb.clear_gradients()
self.assertTrue(emb.weight.gradient() is None)
input_emb.clear_gradient()
self.assertTrue(input_emb.gradient() is not None)
paddle.enable_static()
def test_NoDetachMulti_DetachMulti(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
array_no_detach_multi = self.no_detach_multi()
array_detach_multi = self.detach_multi()
assert not np.array_equal(array_no_detach_multi, array_detach_multi)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def test_selectedrows_gradient2(self):
places = [fluid.CPUPlace()]
if core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
for place in places:
for sort_sum_gradient in [True, False]:
with fluid.dygraph.guard(place):
fluid.set_flags({
'FLAGS_sort_sum_gradient': sort_sum_gradient
})
grad_clip = fluid.clip.GradientClipByGlobalNorm(5.0)
input_word = np.array([[1, 2], [2, 1]]).astype('int64')
input = to_variable(input_word)
simplenet = SimpleNet(20, 32, "float32")
adam = SGDOptimizer(
learning_rate=0.001,
parameter_list=simplenet.parameters(),
grad_clip=grad_clip)
input_emb, emb = simplenet(input)
self.assertTrue(emb.weight.gradient() is None)
self.assertTrue(input_emb.gradient() is None)
input_emb.backward()
adam.minimize(input_emb)
self.assertTrue(emb.weight.gradient() is not None)
emb.clear_gradients()
self.assertTrue(emb.weight.gradient() is None)
input_emb.clear_gradient()
self.assertTrue(input_emb.gradient() is not None)
def test_coo_values_grad(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
with _test_eager_guard():
indices = [[0, 0, 1, 2, 2], [1, 3, 2, 0, 1]]
values = [1.0, 2.0, 3.0, 4.0, 5.0]
sparse_x = paddle.incubate.sparse.sparse_coo_tensor(
paddle.to_tensor(indices),
paddle.to_tensor(values),
shape=[3, 4],
stop_gradient=False)
values_tensor = sparse_x.values()
out_grad = [2.0, 3.0, 5.0, 8.0, 9.0]
# test coo_values_grad
values_tensor.backward(paddle.to_tensor(out_grad))
assert np.array_equal(out_grad, sparse_x.grad.values().numpy())
indices = [[0, 0, 1, 2, 2], [1, 3, 2, 0, 1]]
values = [[1.0, 1.0], [2.0, 2.0], [3.0, 3.0], [4.0, 4.0],
[5.0, 5.0]]
sparse_x = paddle.incubate.sparse.sparse_coo_tensor(
paddle.to_tensor(indices),
paddle.to_tensor(values),
shape=[3, 4, 2],
stop_gradient=False)
values_tensor = sparse_x.values()
out_grad = [[2.0, 2.0], [3.0, 3.0], [5.0, 5.0], [8.0, 8.0],
[9.0, 9.0]]
# test coo_values_grad
values_tensor.backward(paddle.to_tensor(out_grad))
assert np.array_equal(out_grad, sparse_x.grad.values().numpy())
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def run_program(enable_addto):
np.random.seed(10)
paddle.seed(10)
paddle.framework.random._manual_program_seed(10)
if fluid.core.is_compiled_with_cuda():
fluid.set_flags({"FLAGS_cudnn_deterministic": True})
fluid.set_flags({"FLAGS_max_inplace_grad_add": 2})
loss, main, startup, w = create_program(data_format=data_format)
place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda(
) else fluid.CPUPlace()
exe = fluid.Executor(place)
strategy = fluid.BuildStrategy()
strategy.enable_addto = enable_addto
compiled = fluid.CompiledProgram(main).with_data_parallel(
loss_name=loss.name, build_strategy=strategy)
exe.run(startup)
img = np.random.uniform(-128, 128,
[8, 3, 224, 224]).astype(np.float32)
for i in range(10):
res = exe.run(compiled,
feed={'img': img},
fetch_list=[loss.name, w.name])
return res
def func_sum_op(self):
x = np.ones([2, 2], np.float32)
with fluid.dygraph.guard():
inputs = []
for _ in range(10):
tmp = paddle.to_tensor(x)
tmp.stop_gradient = False
inputs.append(tmp)
ret = paddle.add_n(inputs)
loss = fluid.layers.reduce_sum(ret)
loss.backward()
with fluid.dygraph.guard():
inputs2 = []
for _ in range(10):
tmp = paddle.to_tensor(x)
tmp.stop_gradient = False
inputs2.append(tmp)
ret2 = paddle.add_n(inputs2)
loss2 = fluid.layers.reduce_sum(ret2)
fluid.set_flags({'FLAGS_sort_sum_gradient': True})
loss2.backward()
self.assertTrue(np.allclose(ret.numpy(), x * 10))
self.assertTrue(np.allclose(inputs[0].gradient(), x))
self.assertTrue(np.allclose(ret2.numpy(), x * 10))
a = inputs2[0].gradient()
self.assertTrue(np.allclose(inputs2[0].gradient(), x))
def test_coo_to_dense(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
with _test_eager_guard():
indices = [[0, 0, 1, 2, 2], [1, 3, 2, 0, 1]]
values = [1.0, 2.0, 3.0, 4.0, 5.0]
sparse_x = paddle.incubate.sparse.sparse_coo_tensor(
paddle.to_tensor(indices),
paddle.to_tensor(values),
shape=[3, 4],
stop_gradient=False)
dense_tensor = sparse_x.to_dense()
#test to_dense_grad backward
out_grad = [[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0],
[9.0, 10.0, 11.0, 12.0]]
dense_tensor.backward(paddle.to_tensor(out_grad))
#mask the out_grad by sparse_x.indices()
correct_x_grad = [2.0, 4.0, 7.0, 9.0, 10.0]
assert np.array_equal(correct_x_grad,
sparse_x.grad.values().numpy())
paddle.device.set_device("cpu")
sparse_x_cpu = paddle.incubate.sparse.sparse_coo_tensor(
paddle.to_tensor(indices),
paddle.to_tensor(values),
shape=[3, 4],
stop_gradient=False)
dense_tensor_cpu = sparse_x_cpu.to_dense()
dense_tensor_cpu.backward(paddle.to_tensor(out_grad))
assert np.array_equal(correct_x_grad,
sparse_x_cpu.grad.values().numpy())
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def test_assign_LoDTensorArray(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
main_program = Program()
startup_program = Program()
with program_guard(main_program):
x = fluid.data(name='x', shape=[100, 10], dtype='float32')
x.stop_gradient = False
y = fluid.layers.fill_constant(
shape=[100, 10], dtype='float32', value=1)
z = fluid.layers.elementwise_add(x=x, y=y)
i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=0)
init_array = fluid.layers.array_write(x=z, i=i)
array = fluid.layers.assign(init_array)
sums = fluid.layers.array_read(array=init_array, i=i)
mean = fluid.layers.mean(sums)
append_backward(mean)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
) else fluid.CPUPlace()
exe = fluid.Executor(place)
feed_x = np.random.random(size=(100, 10)).astype('float32')
ones = np.ones((100, 10)).astype('float32')
feed_add = feed_x + ones
res = exe.run(main_program,
feed={'x': feed_x},
fetch_list=[sums.name, x.grad_name])
self.assertTrue(np.allclose(res[0], feed_add))
self.assertTrue(np.allclose(res[1], ones / 1000.0))
def test_in_static_mode_mkldnn(self):
fluid.set_flags({'FLAGS_use_mkldnn': True})
try:
if paddle.fluid.core.is_compiled_with_mkldnn():
self.resnet_helper.train(to_static=True)
finally:
fluid.set_flags({'FLAGS_use_mkldnn': False})
def test_clone(self):
paddle.disable_static()
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
self.python_api = paddle.clone
x = paddle.ones([2])
x.stop_gradient = False
clone_x = paddle.clone(x)
y = clone_x**3
y.backward()
self.assertTrue(np.array_equal(x, [1, 1]), True)
self.assertTrue(np.array_equal(clone_x.grad.numpy(), [3, 3]), True)
self.assertTrue(np.array_equal(x.grad.numpy(), [3, 3]), True)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
paddle.enable_static()
with program_guard(Program(), Program()):
x_np = np.random.randn(2, 3).astype('float32')
x = paddle.static.data("X", shape=[2, 3])
clone_x = paddle.clone(x)
exe = paddle.static.Executor()
y_np = exe.run(paddle.static.default_main_program(),
feed={'X': x_np},
fetch_list=[clone_x])[0]
self.assertTrue(np.array_equal(y_np, x_np), True)
def test_dim2_offset1(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
expected_np = np.array([[2, 1, 2], [2, 2, 1], [2, 2, 2],
[2, 2, 2]]).astype('float32')
expected_grad = np.array([[1, 0, 1], [1, 1, 0], [1, 1, 1],
[1, 1, 1]]).astype('float32')
for idx, p in enumerate(self.places):
if idx == 0:
paddle.set_device('cpu')
else:
paddle.set_device('gpu')
for dtype in self.typelist:
v = paddle.ones((2, ), dtype=dtype)
var = (np.random.random() + 1)
x = paddle.ones((4, 3), dtype=dtype)
x.stop_gradient = False
y = x * 2
ny = y.fill_diagonal_tensor(v, offset=1, dim1=0, dim2=1)
loss = ny.sum()
loss.backward()
self.assertEqual(
(ny.numpy().astype('float32') == expected_np).all(), True)
self.assertEqual(
(y.grad.numpy().astype('float32') == expected_grad).all(),
True)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def test_environ(self):
self.input_np = np.random.random([2, 3, 5, 5]).astype("float32")
for place in [paddle.CPUPlace(), paddle.CUDAPlace(0)]:
fluid.set_flags({'FLAGS_conv2d_disable_cudnn': False})
self.run_all(place)
fluid.set_flags({'FLAGS_conv2d_disable_cudnn': True})
self.run_all(place)
def func(self, place):
# the shape of input variable should be clearly specified, not inlcude -1.
shape = [2, 3, 4, 5]
eps = 0.005
dtype = np.float64
x = layers.data('x', shape, False, dtype)
y = layers.data('y', shape, False, dtype)
x.persistable = True
y.persistable = True
out = layers.elementwise_mul(x, y)
x_arr = np.random.uniform(-1, 1, shape).astype(dtype)
y_arr = np.random.uniform(-1, 1, shape).astype(dtype)
gradient_checker.triple_grad_check([x, y],
out,
x_init=[x_arr, y_arr],
place=place,
eps=eps)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
gradient_checker.triple_grad_check_for_dygraph(self.multiply_wrapper,
[x, y],
out,
x_init=[x_arr, y_arr],
place=place)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def test_all_cases(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
self.func_exception()
self.func_example_with_gradient_and_create_graph()
with _test_eager_guard():
self.func_exception()
self.func_example_with_gradient_and_create_graph()
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def init_dtype_type(self):
if fluid.is_compiled_with_cuda():
fluid.set_flags({'FLAGS_cudnn_deterministic': True})
self.x_type = np.float32
self.index_type = np.int32
self.dim = -2
self.x_shape = (10, 10, 4, 10)
self.index_size = 10
def load_and_train_dygraph(self):
place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
) else fluid.CPUPlace()
with fluid.dygraph.guard(place):
fluid.default_startup_program().random_seed = self.seed
fluid.default_main_program().random_seed = self.seed
fluid.set_flags({'FLAGS_sort_sum_gradient': True})
mnist = fluid.dygraph.static_runner.StaticModelRunner(
model_dir=self.save_dirname,
model_filename=self.model_filename,
params_filename=self.params_filename)
suffix_varname_dict = mnist._program_holder_dict[
'forward']._suffix_varname_dict
dict_old_new = {v: k for k, v in suffix_varname_dict.items()}
dy_param_init_value = {}
for param in mnist.parameters():
dy_param_init_value[param.name] = param.numpy()
sgd = fluid.optimizer.SGD(learning_rate=0.001,
parameter_list=mnist.parameters())
train_reader = paddle.batch(
self.reader_decorator(paddle.dataset.mnist.train()),
batch_size=self.batch_size,
drop_last=True)
train_loader = fluid.io.DataLoader.from_generator(capacity=10)
train_loader.set_sample_list_generator(train_reader, places=place)
mnist.train()
for epoch in range(self.epoch_num):
for batch_id, data in enumerate(train_loader()):
img = data[0]
label = data[1]
label.stop_gradient = True
cost = mnist(img)
loss = fluid.layers.cross_entropy(cost, label)
avg_loss = fluid.layers.mean(loss)
avg_loss.backward()
sgd.minimize(avg_loss)
mnist.clear_gradients()
if batch_id >= self.batch_num:
break
dy_x_data = img.numpy()
dy_out = avg_loss.numpy()
dy_param_value = {}
for param in mnist.parameters():
dy_param_value[param.name] = param.numpy()
return dy_x_data, dy_out, dy_param_init_value, dy_param_value, dict_old_new
def test_dygraph_static_same_loss(self):
if fluid.is_compiled_with_cuda():
fluid.set_flags({"FLAGS_cudnn_deterministic": True})
args = parse_args()
fake_data_reader = FakeDataReader("train", parse_config(args.config))
dygraph_loss = train(args, fake_data_reader, to_static=False)
static_loss = train(args, fake_data_reader, to_static=True)
self.assertTrue(np.allclose(dygraph_loss, static_loss),
msg="dygraph_loss: {} \nstatic_loss: {}".format(
dygraph_loss, static_loss))
def test_dygraph_static_same_loss(self):
if fluid.is_compiled_with_cuda():
fluid.set_flags({"FLAGS_cudnn_deterministic": True})
conf_dict = create_conf_dict()
dygraph_loss = train(conf_dict, to_static=False)
static_loss = train(conf_dict, to_static=True)
self.assertEqual(len(dygraph_loss), len(static_loss))
for i in range(len(dygraph_loss)):
self.assertAlmostEqual(dygraph_loss[i], static_loss[i])
def test_errors(self):
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
with program_guard(Program(), Program()):
# The type of input must be Variable or numpy.ndarray.
x1 = fluid.create_lod_tensor(
np.array([[-1]]), [[1]], fluid.CPUPlace())
self.assertRaises(TypeError, paddle.assign, x1)
# When the type of input is numpy.ndarray, the dtype of input must be float32, int32.
x2 = np.array([[2.5, 2.5]], dtype='uint8')
self.assertRaises(TypeError, paddle.assign, x2)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def test_mnist_declarative_cpu_vs_mkldnn(self):
dygraph_loss_cpu = self.train_dygraph()
fluid.set_flags({'FLAGS_use_mkldnn': True})
try:
dygraph_loss_mkldnn = self.train_dygraph()
finally:
fluid.set_flags({'FLAGS_use_mkldnn': False})
self.assertTrue(
np.allclose(dygraph_loss_cpu, dygraph_loss_mkldnn),
msg='cpu dygraph is {}\n mkldnn dygraph is \n{}'.format(
dygraph_loss_cpu, dygraph_loss_mkldnn))
def run_dygraph():
fluid.set_flags({'FLAGS_sort_sum_gradient': True})
paddle.seed(seed)
paddle.framework.random._manual_program_seed(seed)
ocr_attention = OCRAttention()
if Config.learning_rate_decay == "piecewise_decay":
learning_rate = fluid.layers.piecewise_decay(
[50000], [Config.LR, Config.LR * 0.01])
else:
learning_rate = Config.LR
optimizer = fluid.optimizer.SGD(
learning_rate=0.001, parameter_list=ocr_attention.parameters())
dy_param_init_value = {}
for param in ocr_attention.parameters():
dy_param_init_value[param.name] = param.numpy()
for epoch in range(epoch_num):
for batch_id in range(batch_num):
label_in = to_variable(label_in_np)
label_out = to_variable(label_out_np)
label_out.stop_gradient = True
img = to_variable(image_np)
dy_prediction = ocr_attention(img, label_in)
label_out = fluid.layers.reshape(label_out, [-1, 1],
inplace=False)
dy_prediction = fluid.layers.reshape(
dy_prediction, [label_out.shape[0], -1], inplace=False)
loss = fluid.layers.cross_entropy(input=dy_prediction,
label=label_out)
avg_loss = fluid.layers.reduce_sum(loss)
dy_out = avg_loss.numpy()
if epoch == 0 and batch_id == 0:
for param in ocr_attention.parameters():
if param.name not in dy_param_init_value:
dy_param_init_value[param.name] = param.numpy()
avg_loss.backward()
dy_grad_value = {}
for param in ocr_attention.parameters():
if param.trainable:
np_array = np.array(
param._grad_ivar().value().get_tensor())
dy_grad_value[param.name +
core.grad_var_suffix()] = np_array
optimizer.minimize(avg_loss)
ocr_attention.clear_gradients()
dy_param_value = {}
for param in ocr_attention.parameters():
dy_param_value[param.name] = param.numpy()
return dy_out, dy_param_init_value, dy_param_value
def test_rnn(self):
np_inp = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0],
[10.0, 11.0, 12.0]])
np_inp = np_inp.reshape((1, 4, 3))
np_inp = np_inp.astype(np.float32)
with fluid.dygraph.guard():
var_inp = fluid.dygraph.base.to_variable(np_inp)
var_inp = fluid.layers.reshape(var_inp, shape=[1, 4, 3])
simple_rnn = SimpleRNN()
outs, pre_hiddens = simple_rnn.forward(var_inp)
dy_out = outs[3].numpy()
outs[3].backward()
dy_grad_h2o = simple_rnn._cell._h2o_w.gradient()
dy_grad_h2h = simple_rnn._cell._h2h_w.gradient()
dy_grad_i2h = simple_rnn._cell._i2h_w.gradient()
with fluid.dygraph.guard():
var_inp2 = fluid.dygraph.base.to_variable(np_inp)
var_inp2 = fluid.layers.reshape(var_inp2, shape=[1, 4, 3])
simple_rnn2 = SimpleRNN()
outs2, pre_hiddens2 = simple_rnn2.forward(var_inp2)
dy_out2 = outs2[3].numpy()
fluid.set_flags({'FLAGS_sort_sum_gradient': True})
outs2[3].backward()
dy_grad_h2o2 = simple_rnn2._cell._h2o_w.gradient()
dy_grad_h2h2 = simple_rnn2._cell._h2h_w.gradient()
dy_grad_i2h2 = simple_rnn2._cell._i2h_w.gradient()
with new_program_scope():
inp = fluid.layers.data(name="inp",
shape=[1, 4, 3],
append_batch_size=False)
simple_rnn = SimpleRNN()
outs, pre_hiddens = simple_rnn(inp)
param_grads = fluid.backward.append_backward(outs[3])
exe = fluid.Executor(fluid.CPUPlace())
exe.run(fluid.default_startup_program())
static_out, static_grad_h2o, static_grad_h2h, static_grad_i2h = exe.run(
feed={inp.name: np_inp},
fetch_list=[
outs[3].name, param_grads[0][1].name,
param_grads[1][1].name, param_grads[2][1].name
])
self.assertTrue(np.allclose(dy_out, static_out))
self.assertTrue(np.allclose(dy_grad_h2o, static_grad_h2o))
self.assertTrue(np.allclose(dy_grad_h2h, static_grad_h2h))
self.assertTrue(np.allclose(dy_grad_i2h, static_grad_i2h))
self.assertTrue(np.allclose(dy_out2, static_out))
self.assertTrue(np.allclose(dy_grad_h2o2, static_grad_h2o))
self.assertTrue(np.allclose(dy_grad_h2h2, static_grad_h2h))
self.assertTrue(np.allclose(dy_grad_i2h2, static_grad_i2h))
def func_append_activation_in_dygraph_global_use_mkldnn(self):
a_np = np.random.uniform(-2, 2, (10, 20, 30)).astype(np.float32)
helper = LayerHelper(fluid.unique_name.generate("test"), act="relu")
func = helper.append_activation
with fluid.dygraph.guard(fluid.core.CPUPlace()):
a = paddle.to_tensor(a_np)
fluid.set_flags({'FLAGS_use_mkldnn': True})
try:
res1 = func(a)
finally:
fluid.set_flags({'FLAGS_use_mkldnn': False})
res2 = fluid.layers.relu(a)
self.assertTrue(np.array_equal(res1.numpy(), res2.numpy()))
def __init__(self, cfg):
paddle.seed(1)
paddle.framework.random._manual_program_seed(1)
self.generator = Generator(cfg)
self.discriminator = Discriminator(cfg)
self.g_optimizer = build_optimizer(self.generator, cfg)
self.d_optimizer = build_optimizer(self.discriminator, cfg)
self.cfg = cfg
fluid.set_flags({'FLAGS_sort_sum_gradient': cfg.sort_sum_gradient})
def test_mlp(self):
np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
with fluid.dygraph.guard():
var_inp = fluid.dygraph.base.to_variable(np_inp)
mlp = MLP(input_size=2)
out = mlp(var_inp)
dy_out = out.numpy()
out.backward()
dy_grad = mlp._linear1.weight.gradient()
with fluid.dygraph.guard():
var_inp2 = fluid.dygraph.base.to_variable(np_inp)
mlp2 = MLP(input_size=2)
out2 = mlp2(var_inp2)
dy_out2 = out2.numpy()
fluid.set_flags({'FLAGS_sort_sum_gradient': True})
out2.backward()
dy_grad2 = mlp2._linear1.weight.gradient()
with new_program_scope():
inp = fluid.layers.data(name="inp",
shape=[2, 2],
append_batch_size=False)
mlp = MLP(input_size=2)
out = mlp(inp)
param_grads = fluid.backward.append_backward(
out, parameter_list=[mlp._linear1.weight.name])[0]
exe = fluid.Executor(fluid.CPUPlace(
) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
exe.run(fluid.default_startup_program())
static_out, static_grad = exe.run(
feed={inp.name: np_inp},
fetch_list=[out.name, param_grads[1].name])
self.assertTrue(np.allclose(dy_out, static_out))
self.assertTrue(np.allclose(dy_grad, static_grad))
self.assertTrue(np.allclose(dy_out2, static_out))
self.assertTrue(np.allclose(dy_grad2, static_grad))
params = mlp.parameters(True)
self.assertEqual("linear_0.w_0", params[0].name)
self.assertEqual("linear_0.b_0", params[1].name)
self.assertEqual("linear_1.w_0", params[2].name)
self.assertEqual("linear_1.b_0", params[3].name)
self.assertEqual(len(params), 4)
sublayers = mlp.sublayers(True)
self.assertEqual(mlp._linear1, sublayers[0])
self.assertEqual(mlp._linear2, sublayers[1])
self.assertEqual(len(sublayers), 2)
def setUp(self):
paddle.enable_static()
if paddle.is_compiled_with_cuda():
fluid.set_flags({
'FLAGS_cudnn_deterministic': 1,
'FLAGS_max_inplace_grad_add': 6,
})
self.place = paddle.CUDAPlace(0)
else:
self.place = paddle.CPUPlace()
self.use_cuda = isinstance(self.place, paddle.CUDAPlace)
self.executor = paddle.static.Executor(self.place)
self.num_classes = 1000
self.seed = 1
def test_single_api(sort_sum_gradient):
fluid.set_flags({'FLAGS_sort_sum_gradient': sort_sum_gradient})
x = paddle.to_tensor(5., stop_gradient=False)
for i in range(10):
y = paddle.pow(x, 4.0)
y.backward()
self.assertEqual(x.grad.numpy(), (i + 1) * 500)
x.clear_gradient()
self.assertEqual(x.grad.numpy(), 0.)
for i in range(10):
y = paddle.pow(x, 4.0)
y.backward()
self.assertEqual(x.grad.numpy(), (i + 1) * 500)
x.clear_grad()
self.assertEqual(x.grad.numpy(), 0.)
def func(self, place):
shape = [2, 3, 7, 9]
eps = 0.0005
dtype = np.float64
x = layers.data('x', shape, False, dtype=dtype)
x.persistable = True
y = layers.tanh(x)
x_arr = np.random.random(shape).astype(dtype)
x_arr[np.abs(x_arr) < 0.005] = 0.002
gradient_checker.triple_grad_check(
[x], y, x_init=x_arr, place=place, eps=eps)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
gradient_checker.triple_grad_check_for_dygraph(
self.tanh_wrapper, [x], y, x_init=x_arr, place=place)
fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": False})
def grad(self,
outputs,
inputs,
grad_outputs=None,
no_grad_vars=None,
retain_graph=None,
create_graph=False,
allow_unused=False):
fluid.set_flags({'FLAGS_sort_sum_gradient': self.sort_sum_gradient})
return fluid.dygraph.grad(outputs=outputs,
inputs=inputs,
grad_outputs=grad_outputs,
no_grad_vars=no_grad_vars,
retain_graph=retain_graph,
create_graph=create_graph,
allow_unused=allow_unused)
def test_api(self):
flags = {
'FLAGS_eager_delete_tensor_gb': 1.0,
'FLAGS_check_nan_inf': True
}
fluid.set_flags(flags)
flags_list = ['FLAGS_eager_delete_tensor_gb', 'FLAGS_check_nan_inf']
flag = 'FLAGS_eager_delete_tensor_gb'
res_list = fluid.get_flags(flags_list)
res = fluid.get_flags(flag)
self.assertTrue(res_list['FLAGS_eager_delete_tensor_gb'], 1.0)
self.assertTrue(res_list['FLAGS_check_nan_inf'], True)
self.assertTrue(res['FLAGS_eager_delete_tensor_gb'], 1.0)