m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix the error of init variable in StaticRNN when stop_gradient=ON (#21118)

Browse Source
revert-21172-masked_select_api
guofei 6 years ago committed by Huihuang Zheng
parent 3c98ec90ce
commit 56b5d14704
2 changed files with 89 additions and 15 deletions
Show Stats Download Patch File Download Diff File

8
paddle/fluid/operators/recurrent_op.cc
Unescape View File

@ -635,11 +635,9 @@ class RecurrentGradOpShapeInference : public framework::InferShapeBase {
RecurrentBase::kOutputs);
// In some case the kInitialStates is empty.
if (ctx->HasInputs(RecurrentBase::kInitialStates)) {
PADDLE_ENFORCE_EQ(ctx->HasOutputs(framework::GradVarName(
RecurrentBase::kInitialStates)),
true, "The output of(%s) should not be empty.",
framework::GradVarName(RecurrentBase::kInitialStates));
if (ctx->HasInputs(RecurrentBase::kInitialStates) &&
ctx->HasOutputs(
framework::GradVarName(RecurrentBase::kInitialStates))) {
ctx->SetOutputsDim(framework::GradVarName(RecurrentBase::kInitialStates),
ctx->GetInputsDim(RecurrentBase::kInitialStates));
}

96
python/paddle/fluid/tests/unittests/test_recurrent_op.py
Unescape View File

@ -123,7 +123,8 @@ class RecurrentOpTest1(unittest.TestCase):
def setUp(self):
self.setup_program()
self.data_field = {"x", "h_boot"}
self.feed_data_field = {"x", "h_boot"}
self.grad_data_field = self.feed_data_field
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
@ -161,7 +162,7 @@ class RecurrentOpTest1(unittest.TestCase):
def forward(self):
self.feed_map = {
x: create_tensor(getattr(self.py_rnn, x), self.place)
for x in self.data_field
for x in self.feed_data_field
}
exe = Executor(self.place)
out = exe.run(self.main_program,
@ -173,11 +174,11 @@ class RecurrentOpTest1(unittest.TestCase):
def backward(self):
self.feed_map = {
x: create_tensor(getattr(self.py_rnn, x), self.place)
for x in self.data_field
for x in self.feed_data_field
}
fetch_list = [
self.main_program.global_block().var(grad_var_name(x))
for x in self.data_field
for x in self.grad_data_field
]
exe = Executor(self.place)
@ -195,7 +196,7 @@ class RecurrentOpTest1(unittest.TestCase):
ana_grad = [np.array(x) for x in self.backward()]
num_grad = self.get_numerical_gradient()
for idx, name in enumerate(self.data_field):
for idx, name in enumerate(self.grad_data_field):
self.assertEqual(num_grad[idx].shape, ana_grad[idx].shape)
self.assertTrue(
np.isclose(
@ -212,7 +213,7 @@ class RecurrentOpTest1(unittest.TestCase):
def get_numerical_gradient(self, delta=0.005):
dloss_dout = 1.0
feed_list = [getattr(self.py_rnn, x) for x in self.data_field]
feed_list = [getattr(self.py_rnn, x) for x in self.grad_data_field]
grad_list = [np.zeros_like(x) for x in feed_list]
for feed, grad in zip(feed_list, grad_list):
for f, g in np.nditer([feed, grad], op_flags=['readwrite']):
@ -253,7 +254,8 @@ class RecurrentOpTest2(RecurrentOpTest1):
def setUp(self):
self.setup_program()
self.data_field = {"x", "h_boot", "W", "U"}
self.feed_data_field = {"x", "h_boot", "W", "U"}
self.grad_data_field = self.feed_data_field
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
@ -352,7 +354,8 @@ class RecurrentOpMultipleMemoryTest(RecurrentOpTest1):
def setUp(self):
self.setup_program()
self.data_field = {"x", "h_boot1", "h_boot2"}
self.feed_data_field = {"x", "h_boot1", "h_boot2"}
self.grad_data_field = self.feed_data_field
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
@ -435,7 +438,8 @@ class RecurrentOpNoMemBootTest(RecurrentOpTest1):
def setUp(self):
self.setup_program()
self.data_field = {"x"}
self.feed_data_field = {"x"}
self.grad_data_field = self.feed_data_field
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
@ -535,7 +539,8 @@ class RecurrentOpSubBlockTest(RecurrentOpTest1):
def setUp(self):
self.setup_program()
self.data_field = {"x", "emb", "w1", "w2"}
self.feed_data_field = {"x", "emb", "w1", "w2"}
self.grad_data_field = self.feed_data_field
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
@ -602,5 +607,76 @@ class RecurrentOpSubBlockTest(RecurrentOpTest1):
return rnn()
class RecurrentOpStopGradientTest(RecurrentOpTest1):
"""
Test RNNOp with stop_gradient = True
equation:
h_t = \sigma (W x_t + U h_{t-1})
weights:
- W
- U
vars:
- x
memories:
- h
output:
- h
"""
input_dim = 2
batch_size = 10
sent_len = 2
def setUp(self):
self.setup_program()
self.feed_data_field = {"x", "h_boot", "W", "U"}
self.grad_data_field = {"x", "W", "U"}
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
self.py_rnn = PySimpleRNN2(self.input_shape, self.output_shape)
with fluid.program_guard(self.main_program, self.startup_program):
self.output = layers.mean(self.create_rnn_op())
def create_rnn_op(self):
x = layers.data(
shape=[self.sent_len, self.batch_size, self.input_dim],
dtype="float32",
name="x",
append_batch_size=False)
x.stop_gradient = False
h_boot = layers.data(
shape=[self.input_dim], dtype="float32", name="h_boot")
h_boot.stop_gradient = True
rnn = layers.StaticRNN()
with rnn.step():
h_pre = rnn.memory(init=h_boot) # init doesn't have gradient
x_t = rnn.step_input(x)
temp_l = layers.fc(
input=x_t,
size=self.input_dim,
param_attr=ParamAttr(
name="W",
initializer=fluid.initializer.ConstantInitializer(1.0)),
bias_attr=False)
temp_r = layers.fc(
input=h_pre,
size=self.input_dim,
param_attr=ParamAttr(
name="U",
initializer=fluid.initializer.ConstantInitializer(0.0)),
bias_attr=False)
h = layers.sigmoid(x=layers.elementwise_add(temp_l, temp_r))
rnn.update_memory(h_pre, h)
rnn.output(h)
return rnn()
if __name__ == '__main__':
unittest.main()

Write Preview
Loading…
Cancel
Save