提交 212f6eae 编写于 作者: L Luo Tao

modify the test config for test_CompareTwoNets.cpp

上级 3654e1e0
......@@ -111,3 +111,12 @@ if(NOT ON_TRAVIS)
${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
endif()
################ test_CompareTwoNets ######################
add_unittest_without_exec(test_CompareTwoNets
test_CompareTwoNets.cpp)
add_test(NAME test_CompareTwoNets
COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
#!/usr/bin/env python
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers import *
######################## data source ################################
dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
dict_file = dict()
for line_count, line in enumerate(open(dict_path, "r")):
dict_file[line.strip()] = line_count
define_py_data_sources2(
train_list='gserver/tests/Sequence/train.list',
test_list=None,
module='sequenceGen',
obj='process',
args={"dict_file": dict_file})
settings(batch_size=5)
######################## network configure ################################
dict_dim = len(open(dict_path, 'r').readlines())
word_dim = 128
hidden_dim = 128
label_dim = 3
# This config is designed to be equivalent with sequence_recurrent_group.py
data = data_layer(name="word", size=dict_dim)
emb = embedding_layer(
input=data, size=word_dim, param_attr=ParamAttr(name="emb"))
recurrent = recurrent_layer(input=emb, bias_attr=False, act=SoftmaxActivation())
recurrent_last = last_seq(input=recurrent)
with mixed_layer(
size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
output += full_matrix_projection(input=recurrent_last)
outputs(
classification_cost(
input=output, label=data_layer(
name="label", size=1)))
#!/usr/bin/env python
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers import *
######################## data source ################################
dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
dict_file = dict()
for line_count, line in enumerate(open(dict_path, "r")):
dict_file[line.strip()] = line_count
define_py_data_sources2(
train_list='gserver/tests/Sequence/train.list',
test_list=None,
module='sequenceGen',
obj='process',
args={"dict_file": dict_file})
settings(batch_size=5)
######################## network configure ################################
dict_dim = len(open(dict_path, 'r').readlines())
word_dim = 128
hidden_dim = 128
label_dim = 3
# This config is designed to be equivalent with sequence_recurrent.py
data = data_layer(name="word", size=dict_dim)
emb = embedding_layer(
input=data, size=word_dim, param_attr=ParamAttr(name="emb"))
def step(y):
mem = memory(name="rnn_state", size=hidden_dim)
with mixed_layer(
name="rnn_state",
size=hidden_dim,
bias_attr=False,
act=SoftmaxActivation()) as out:
out += identity_projection(input=y)
out += full_matrix_projection(
input=mem, param_attr=ParamAttr(name="___recurrent_layer_0__"))
return out
recurrent = recurrent_group(name="rnn", step=step, input=emb)
recurrent_last = last_seq(input=recurrent)
with mixed_layer(
size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
output += full_matrix_projection(input=recurrent_last)
outputs(
classification_cost(
input=output, label=data_layer(
name="label", size=1)))
......@@ -30,8 +30,6 @@ DECLARE_bool(use_gpu);
DECLARE_string(config);
DECLARE_string(nics);
DEFINE_string(config_file_a, "", "config of one network to compare");
DEFINE_string(config_file_b, "", "config of another network to compare");
DEFINE_bool(need_high_accuracy,
false,
"whether need to run in double accuracy");
......@@ -42,6 +40,10 @@ DEFINE_double(
DECLARE_bool(thread_local_rand_use_global_seed);
DECLARE_int32(seed);
static const string& config_file_a = "gserver/tests/sequence_recurrent.py";
static const string& config_file_b =
"gserver/tests/sequence_recurrent_group.py";
struct ComData {
vector<Argument> outArgs;
vector<ParameterPtr> parameters;
......@@ -66,6 +68,7 @@ void calcGradient(ComData& data, const string configFile) {
DataBatch dataBatch;
int32_t batchSize = trainer.getConfig().opt_config().batch_size();
trainer.getDataProvider()->reset();
trainer.getDataProvider()->setSkipShuffle();
trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch);
......@@ -167,11 +170,11 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
TEST(Trainer, create) {
ComData dataA;
calcGradient(dataA, FLAGS_config_file_a);
calcGradient(dataA, config_file_a);
LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n";
ComData dataB;
calcGradient(dataB, FLAGS_config_file_b);
calcGradient(dataB, config_file_b);
LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n";
compareGradient(dataA, dataB);
......
......@@ -28,14 +28,6 @@ if(WITH_PYTHON)
${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
endif()
################ test_CompareTwoNets ######################
add_unittest_without_exec(test_CompareTwoNets
test_CompareTwoNets.cpp)
add_test(NAME test_CompareTwoNets
COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
--config_file_a=trainer/tests/sample_trainer_config_qb_rnn.conf --config_file_b=trainer/tests/sample_trainer_config_rnn.conf
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
############### test_CompareTwoOpts ###################
add_unittest_without_exec(test_CompareTwoOpts
......
#edit-mode: -*- python -*-
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#Todo(luotao02) This config is only used for unitest. It is out of date now, and will be updated later.
# Note: when making change to this file, please make sure
# sample_trainer_config_rnn.conf is changed accordingly so that the uniitest
# for comparing these two nets can pass (test_CompareTwoNets)
default_initial_std(0.1)
default_device(0)
word_dim = 1451594
l1 = 0
l2 = 0
model_type("nn")
sparse_update = get_config_arg("sparse_update", bool, False)
TrainData(ProtoData(
type = "proto_sequence",
files = ('trainer/tests/train.list'),
))
Settings(
algorithm='sgd',
batch_size=100,
learning_rate=0.0001,
learning_rate_decay_a=4e-08,
learning_rate_decay_b=0.0,
learning_rate_schedule='poly',
)
wordvec_dim = 128
layer2_dim = 96
layer3_dim = 96
hidden_dim = 128
slot_names = ["qb", "qw", "tb", "tw"]
def ltr_network(network_name,
word_dim=word_dim,
wordvec_dim=wordvec_dim,
layer2_dim=layer2_dim,
layer3_dim=layer3_dim,
hidden_dim=hidden_dim,
slot_names=slot_names,
l1=l1,
l2=l2):
slotnum = len(slot_names)
for i in xrange(slotnum):
Inputs(slot_names[i] + network_name)
for i in xrange(slotnum):
Layer(
name = slot_names[i] + network_name,
type = "data",
size = word_dim,
device = -1,
)
Layer(
name = slot_names[i] + "_embedding_" + network_name,
type = "mixed",
size = wordvec_dim,
bias = False,
device = -1,
inputs = TableProjection(slot_names[i] + network_name,
parameter_name = "embedding.w0",
decay_rate_l1=l1,
sparse_remote_update = True,
sparse_update = sparse_update,
),
)
Layer(
name = slot_names[i] + "_rnn1_" + network_name,
type = "recurrent",
active_type = "tanh",
bias = Bias(initial_std = 0,
parameter_name = "rnn1.bias"),
inputs = Input(slot_names[i] + "_embedding_" + network_name,
parameter_name = "rnn1.w0")
)
Layer(
name = slot_names[i] + "_rnnlast_" + network_name,
type = "seqlastins",
inputs = [
slot_names[i] + "_rnn1_" + network_name,
],
)
Layer(
name = "layer2_" + network_name,
type = "fc",
active_type = "tanh",
size = layer2_dim,
bias = Bias(parameter_name = "layer2.bias"),
inputs = [Input(slot_name + "_rnnlast_" + network_name,
parameter_name = "_layer2_" + slot_name + ".w",
decay_rate = l2,
initial_smart = True) for slot_name in slot_names]
)
Layer(
name = "layer3_" + network_name,
type = "fc",
active_type = "tanh",
size = layer3_dim,
bias = Bias(parameter_name = "layer3.bias"),
inputs = [
Input("layer2_" + network_name,
parameter_name = "_layer3.w",
decay_rate = l2,
initial_smart = True),
]
)
Layer(
name = "output_" + network_name,
type = "fc",
size = 1,
bias = False,
inputs = [
Input("layer3_" + network_name,
parameter_name = "_layerO.w"),
],
)
ltr_network("left")
ltr_network("right")
Inputs("label")
Layer(
name = "label",
type = "data",
size = 1,
)
Outputs("cost", "qb_rnnlast_left")
Layer(
name = "cost",
type = "rank-cost",
inputs = ["output_left", "output_right", "label"],
)
#edit-mode: -*- python -*-
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#Todo(luotao02) This config is only used for unitest. It is out of date now, and will be updated later.
# Note: when making change to this file, please make sure
# sample_trainer_config_qb_rnn.conf is changed accordingly so that the uniitest
# for comparing these two nets can pass (test_CompareTwoNets)
default_initial_std(0.1)
default_device(0)
word_dim = 1451594
l1 = 0
l2 = 0
model_type("recurrent_nn")
sparse_update = get_config_arg("sparse_update", bool, False)
TrainData(ProtoData(
type = "proto_sequence",
files = ('trainer/tests/train.list'),
))
Settings(
algorithm='sgd',
batch_size=100,
learning_rate=0.0001,
learning_rate_decay_a=4e-08,
learning_rate_decay_b=0.0,
learning_rate_schedule='poly',
)
wordvec_dim = 128
layer2_dim = 96
layer3_dim = 96
hidden_dim = 128
slot_names = ["qb", "qw", "tb", "tw"]
def SimpleRecurrentLayer(name,
size,
active_type,
bias,
input_layer_name,
parameter_name,
seq_reversed = False):
RecurrentLayerGroupBegin(name + "_layer_group",
in_links=[input_layer_name],
out_links=[name],
seq_reversed=seq_reversed)
memory_name = Memory(name=name, size=size)
Layer(
name = name,
type = "mixed",
size = size,
active_type = active_type,
bias = bias,
inputs = [IdentityProjection(input_layer_name),
FullMatrixProjection(memory_name,
parameter_name = parameter_name,
),
]
)
RecurrentLayerGroupEnd(name + "_layer_group")
def ltr_network(network_name,
word_dim=word_dim,
wordvec_dim=wordvec_dim,
layer2_dim=layer2_dim,
layer3_dim=layer3_dim,
hidden_dim=hidden_dim,
slot_names=slot_names,
l1=l1,
l2=l2):
slotnum = len(slot_names)
for i in xrange(slotnum):
Inputs(slot_names[i] + network_name)
for i in xrange(slotnum):
Layer(
name = slot_names[i] + network_name,
type = "data",
size = word_dim,
device = -1,
)
Layer(
name = slot_names[i] + "_embedding_" + network_name,
type = "mixed",
size = wordvec_dim,
bias = False,
device = -1,
inputs = TableProjection(slot_names[i] + network_name,
parameter_name = "embedding.w0",
decay_rate_l1=l1,
sparse_remote_update = True,
sparse_update = sparse_update,
),
)
SimpleRecurrentLayer(
name = slot_names[i] + "_rnn1_" + network_name,
size = hidden_dim,
active_type = "tanh",
bias = Bias(initial_std = 0,
parameter_name = "rnn1.bias"),
input_layer_name = slot_names[i] + "_embedding_" + network_name,
parameter_name = "rnn1.w0",
)
Layer(
name = slot_names[i] + "_rnnlast_" + network_name,
type = "seqlastins",
inputs = [
slot_names[i] + "_rnn1_" + network_name,
],
)
Layer(
name = "layer2_" + network_name,
type = "fc",
active_type = "tanh",
size = layer2_dim,
bias = Bias(parameter_name = "layer2.bias"),
inputs = [Input(slot_name + "_rnnlast_" + network_name,
parameter_name = "_layer2_" + slot_name + ".w",
decay_rate = l2,
initial_smart = True) for slot_name in slot_names]
)
Layer(
name = "layer3_" + network_name,
type = "fc",
active_type = "tanh",
size = layer3_dim,
bias = Bias(parameter_name = "layer3.bias"),
inputs = [
Input("layer2_" + network_name,
parameter_name = "_layer3.w",
decay_rate = l2,
initial_smart = True),
]
)
Layer(
name = "output_" + network_name,
type = "fc",
size = 1,
bias = False,
inputs = [
Input("layer3_" + network_name,
parameter_name = "_layerO.w"),
],
)
ltr_network("left")
ltr_network("right")
Inputs("label")
Layer(
name = "label",
type = "data",
size = 1,
)
Outputs("cost", "qb_rnnlast_left")
Layer(
name = "cost",
type = "rank-cost",
inputs = ["output_left", "output_right", "label"],
)
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