未验证 提交 2bd92754 编写于 作者: Y Yu Yang 提交者: GitHub

Merge pull request #8005 from reyoung/feature/make_nmt_normal_unittest

Make NMT as normal python unittests
...@@ -53,6 +53,8 @@ class WhileOp : public framework::OperatorBase { ...@@ -53,6 +53,8 @@ class WhileOp : public framework::OperatorBase {
auto step_scopes = auto step_scopes =
scope.FindVar(Output(kStepScopes))->GetMutable<StepScopeVar>(); scope.FindVar(Output(kStepScopes))->GetMutable<StepScopeVar>();
PADDLE_ENFORCE(platform::is_cpu_place(cond.place()),
"Condition of while op must in CPU memory.");
while (cond.data<bool>()[0]) { while (cond.data<bool>()[0]) {
auto &current_scope = scope.NewScope(); auto &current_scope = scope.NewScope();
step_scopes->push_back(&current_scope); step_scopes->push_back(&current_scope);
......
...@@ -295,7 +295,7 @@ def fill_constant_batch_size_like(input, ...@@ -295,7 +295,7 @@ def fill_constant_batch_size_like(input,
return out return out
def ones(shape, dtype): def ones(shape, dtype, force_cpu=False):
""" """
**ones** **ones**
...@@ -319,7 +319,7 @@ def ones(shape, dtype): ...@@ -319,7 +319,7 @@ def ones(shape, dtype):
return fill_constant(value=1.0, **locals()) return fill_constant(value=1.0, **locals())
def zeros(shape, dtype): def zeros(shape, dtype, force_cpu=False):
""" """
**zeros** **zeros**
......
...@@ -11,21 +11,20 @@ ...@@ -11,21 +11,20 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
import contextlib
import numpy as np import numpy as np
import paddle.v2 as paddle import paddle.v2 as paddle
import paddle.v2.fluid as fluid import paddle.v2.fluid as fluid
import paddle.v2.fluid.core as core
import paddle.v2.fluid.framework as framework import paddle.v2.fluid.framework as framework
import paddle.v2.fluid.layers as pd import paddle.v2.fluid.layers as pd
from paddle.v2.fluid.executor import Executor from paddle.v2.fluid.executor import Executor
import unittest
dict_size = 30000 dict_size = 30000
source_dict_dim = target_dict_dim = dict_size source_dict_dim = target_dict_dim = dict_size
src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
hidden_dim = 32 hidden_dim = 32
word_dim = 16 word_dim = 16
IS_SPARSE = True
batch_size = 2 batch_size = 2
max_length = 8 max_length = 8
topk_size = 50 topk_size = 50
...@@ -34,10 +33,8 @@ beam_size = 2 ...@@ -34,10 +33,8 @@ beam_size = 2
decoder_size = hidden_dim decoder_size = hidden_dim
place = core.CPUPlace()
def encoder(is_sparse):
def encoder():
# encoder # encoder
src_word_id = pd.data( src_word_id = pd.data(
name="src_word_id", shape=[1], dtype='int64', lod_level=1) name="src_word_id", shape=[1], dtype='int64', lod_level=1)
...@@ -45,7 +42,7 @@ def encoder(): ...@@ -45,7 +42,7 @@ def encoder():
input=src_word_id, input=src_word_id,
size=[dict_size, word_dim], size=[dict_size, word_dim],
dtype='float32', dtype='float32',
is_sparse=IS_SPARSE, is_sparse=is_sparse,
param_attr=fluid.ParamAttr(name='vemb')) param_attr=fluid.ParamAttr(name='vemb'))
fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh') fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
...@@ -54,7 +51,7 @@ def encoder(): ...@@ -54,7 +51,7 @@ def encoder():
return encoder_out return encoder_out
def decoder_train(context): def decoder_train(context, is_sparse):
# decoder # decoder
trg_language_word = pd.data( trg_language_word = pd.data(
name="target_language_word", shape=[1], dtype='int64', lod_level=1) name="target_language_word", shape=[1], dtype='int64', lod_level=1)
...@@ -62,7 +59,7 @@ def decoder_train(context): ...@@ -62,7 +59,7 @@ def decoder_train(context):
input=trg_language_word, input=trg_language_word,
size=[dict_size, word_dim], size=[dict_size, word_dim],
dtype='float32', dtype='float32',
is_sparse=IS_SPARSE, is_sparse=is_sparse,
param_attr=fluid.ParamAttr(name='vemb')) param_attr=fluid.ParamAttr(name='vemb'))
rnn = pd.DynamicRNN() rnn = pd.DynamicRNN()
...@@ -82,10 +79,10 @@ def decoder_train(context): ...@@ -82,10 +79,10 @@ def decoder_train(context):
return rnn() return rnn()
def decoder_decode(context): def decoder_decode(context, is_sparse):
init_state = context init_state = context
array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length) array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
counter = pd.zeros(shape=[1], dtype='int64') counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
# fill the first element with init_state # fill the first element with init_state
state_array = pd.create_array('float32') state_array = pd.create_array('float32')
...@@ -117,7 +114,7 @@ def decoder_decode(context): ...@@ -117,7 +114,7 @@ def decoder_decode(context):
input=pre_ids, input=pre_ids,
size=[dict_size, word_dim], size=[dict_size, word_dim],
dtype='float32', dtype='float32',
is_sparse=IS_SPARSE) is_sparse=is_sparse)
# use rnn unit to update rnn # use rnn unit to update rnn
current_state = pd.fc(input=[pre_ids_emb, pre_state_expanded], current_state = pd.fc(input=[pre_ids_emb, pre_state_expanded],
...@@ -150,7 +147,7 @@ def decoder_decode(context): ...@@ -150,7 +147,7 @@ def decoder_decode(context):
def set_init_lod(data, lod, place): def set_init_lod(data, lod, place):
res = core.LoDTensor() res = fluid.LoDTensor()
res.set(data, place) res.set(data, place)
res.set_lod(lod) res.set_lod(lod)
return res return res
...@@ -165,15 +162,19 @@ def to_lodtensor(data, place): ...@@ -165,15 +162,19 @@ def to_lodtensor(data, place):
lod.append(cur_len) lod.append(cur_len)
flattened_data = np.concatenate(data, axis=0).astype("int64") flattened_data = np.concatenate(data, axis=0).astype("int64")
flattened_data = flattened_data.reshape([len(flattened_data), 1]) flattened_data = flattened_data.reshape([len(flattened_data), 1])
res = core.LoDTensor() res = fluid.LoDTensor()
res.set(flattened_data, place) res.set(flattened_data, place)
res.set_lod([lod]) res.set_lod([lod])
return res return res
def train_main(): def train_main(use_cuda, is_sparse):
context = encoder() if use_cuda and not fluid.core.is_compiled_with_cuda():
rnn_out = decoder_train(context) return
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
context = encoder(is_sparse)
rnn_out = decoder_train(context, is_sparse)
label = pd.data( label = pd.data(
name="target_language_next_word", shape=[1], dtype='int64', lod_level=1) name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
cost = pd.cross_entropy(input=rnn_out, label=label) cost = pd.cross_entropy(input=rnn_out, label=label)
...@@ -212,9 +213,13 @@ def train_main(): ...@@ -212,9 +213,13 @@ def train_main():
batch_id += 1 batch_id += 1
def decode_main(): def decode_main(use_cuda, is_sparse):
context = encoder() if use_cuda and not fluid.core.is_compiled_with_cuda():
translation_ids, translation_scores = decoder_decode(context) return
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
context = encoder(is_sparse)
translation_ids, translation_scores = decoder_decode(context, is_sparse)
exe = Executor(place) exe = Executor(place)
exe.run(framework.default_startup_program()) exe.run(framework.default_startup_program())
...@@ -250,6 +255,60 @@ def decode_main(): ...@@ -250,6 +255,60 @@ def decode_main():
break break
class TestMachineTranslation(unittest.TestCase):
pass
@contextlib.contextmanager
def scope_prog_guard():
prog = fluid.Program()
startup_prog = fluid.Program()
scope = fluid.core.Scope()
with fluid.scope_guard(scope):
with fluid.program_guard(prog, startup_prog):
yield
def inject_test_train(use_cuda, is_sparse):
f_name = 'test_{0}_{1}_train'.format('cuda' if use_cuda else 'cpu', 'sparse'
if is_sparse else 'dense')
def f(*args):
with scope_prog_guard():
train_main(use_cuda, is_sparse)
setattr(TestMachineTranslation, f_name, f)
def inject_test_decode(use_cuda, is_sparse, decorator=None):
f_name = 'test_{0}_{1}_decode'.format('cuda'
if use_cuda else 'cpu', 'sparse'
if is_sparse else 'dense')
def f(*args):
with scope_prog_guard():
decode_main(use_cuda, is_sparse)
if decorator is not None:
f = decorator(f)
setattr(TestMachineTranslation, f_name, f)
for _use_cuda_ in (False, True):
for _is_sparse_ in (False, True):
inject_test_train(_use_cuda_, _is_sparse_)
for _use_cuda_ in (False, True):
for _is_sparse_ in (False, True):
_decorator_ = None
if _use_cuda_:
_decorator_ = unittest.skip(
reason='Beam Search does not support CUDA!')
inject_test_decode(
is_sparse=_is_sparse_, use_cuda=_use_cuda_, decorator=_decorator_)
if __name__ == '__main__': if __name__ == '__main__':
# train_main() unittest.main()
decode_main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册