提交 971cf705 编写于 作者: Q qiaolongfei

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-optimizer-accumulator

......@@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
#include <stdexcept>
#include <string>
#include <vector>
#include "paddle/fluid/framework/executor.h"
......@@ -53,8 +54,14 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
}
}
}
std::vector<framework::LoDTensor> fetch_data;
std::exception_ptr eptr;
try {
fetch_data = underlying_executor_->Run(fetch_tensors);
} catch (...) {
eptr = std::current_exception();
}
auto fetch_data = underlying_executor_->Run(fetch_tensors);
drop_scope_counter_ += 1;
if (!fetch_tensors.empty() ||
drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
......@@ -69,7 +76,11 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
scope->DeleteScope(local_scope);
}
}
return fetch_data;
if (eptr) {
std::rethrow_exception(eptr);
} else {
return fetch_data;
}
}
} // namespace details
} // namespace framework
......
......@@ -78,6 +78,10 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
set.clear();
};
// Clean run context
run_op_futures_.clear();
exception_.reset();
// Step 3. Execution
while (!pending_vars.empty()) {
// 1. Run All Ready ops
......@@ -96,16 +100,19 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
if (timeout) {
std::lock_guard<std::mutex> l(exception_mu_);
std::unique_lock<std::mutex> l(exception_mu_);
if (exception_) {
l.unlock();
for (auto &run_op_future : run_op_futures_) {
run_op_future.wait();
}
l.lock();
std::exception *exp = exception_.get();
if (dynamic_cast<platform::EOFException *>(exp)) {
auto e = *static_cast<platform::EOFException *>(exp);
exception_.reset();
throw e;
} else if (dynamic_cast<platform::EnforceNotMet *>(exp)) {
auto e = *static_cast<platform::EnforceNotMet *>(exp);
exception_.reset();
throw e;
} else {
LOG(FATAL) << "Unknown exception.";
......@@ -222,7 +229,7 @@ void ThreadedSSAGraphExecutor::RunOp(
}
};
if (pool_) {
pool_->enqueue(op_run);
run_op_futures_.emplace_back(pool_->enqueue(op_run));
} else {
op_run();
}
......
......@@ -15,6 +15,7 @@
#pragma once
#include <deque>
#include <list>
#include <string>
#include <unordered_set>
#include <utility>
......@@ -77,6 +78,8 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
private:
ExecutionStrategy strategy_;
// use std::list because clear(), push_back, and for_each are O(1)
std::list<std::future<void>> run_op_futures_;
};
} // namespace details
......
......@@ -265,6 +265,8 @@ op_library(recurrent_op DEPS executor)
op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
op_library(cos_sim_op DEPS cos_sim_functor)
op_library(parallel_do_op DEPS executor)
op_library(unsqueeze_op DEPS reshape_op)
op_library(squeeze_op DEPS reshape_op)
if (WITH_GPU)
op_library(conv_op DEPS vol2col depthwise_conv im2col)
......
......@@ -23,7 +23,7 @@ class BatchReader : public framework::DecoratedReader {
BatchReader(const std::shared_ptr<ReaderBase>& reader, int batch_size,
bool discard_leftover)
: DecoratedReader(reader),
batch_size_(batch_size),
batch_size_(static_cast<size_t>(batch_size)),
discard_leftover_(discard_leftover) {
buffer_.reserve(batch_size_);
}
......@@ -31,7 +31,7 @@ class BatchReader : public framework::DecoratedReader {
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override;
private:
int batch_size_;
size_t batch_size_;
bool discard_leftover_;
std::vector<std::vector<framework::LoDTensor>> buffer_;
};
......@@ -78,7 +78,7 @@ class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase {
void BatchReader::ReadNextImpl(std::vector<framework::LoDTensor>* out) {
buffer_.clear();
buffer_.reserve(batch_size_);
for (int i = 0; i < batch_size_; ++i) {
for (size_t i = 0; i < batch_size_; ++i) {
buffer_.push_back(std::vector<framework::LoDTensor>());
reader_->ReadNext(&buffer_.back());
if (buffer_.back().empty()) {
......@@ -95,9 +95,9 @@ void BatchReader::ReadNextImpl(std::vector<framework::LoDTensor>* out) {
// if buffer_ is empty, the 'out' will return as an empty vector.
return;
}
int out_num = buffer_[0].size();
size_t out_num = buffer_[0].size();
out->reserve(out_num);
for (int j = 0; j < out_num; ++j) {
for (size_t j = 0; j < out_num; ++j) {
// Merge shape and check date type
std::type_index batch_type = buffer_[0][j].type();
framework::DDim batch_shape = buffer_[0][j].dims();
......
/* Copyright (c) 2018 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. */
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
class SqueezeOpInferShape : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of SqueezeOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of SqueezeOp should not be null.");
const auto &x_dims = ctx->GetInputDim("X");
// Check input tensor dims (<6) Eigen limit.
PADDLE_ENFORCE(x_dims.size() <= 6,
"Invalid dimnesions, the rank of Input(X) "
"should be in the range of [1, 6] (Eigen limit).");
const auto &axes = ctx->Attrs().Get<std::vector<int>>("axes");
for (int a : axes) {
PADDLE_ENFORCE_LT(a, x_dims.size(),
"The squeeze axis should be less than input "
"tensor's rank.");
}
auto out_dims = GetOutputShape(axes, x_dims);
ctx->SetOutputDim("Out", out_dims);
if (x_dims[0] == out_dims[0]) {
// Only pass LoD when the first dimension of output and Input(X)
// are the same.
ctx->ShareLoD("X", "Out");
}
}
static framework::DDim GetOutputShape(const std::vector<int> squeeze_dims,
const framework::DDim &in_dims) {
size_t num_squeeze_dims = squeeze_dims.size();
int cnt_squeezed_dims = 0;
bool should_squeeze[9] = {false};
// Determines number of dimensions of output tensor after squeeze.
// Mark and count the dimensions need to be squeezed
if (num_squeeze_dims == 0) {
for (int idx = 0; idx < in_dims.size(); ++idx) {
if (in_dims[idx] == 1) {
should_squeeze[idx] = true;
++cnt_squeezed_dims;
}
}
} else {
for (size_t idx = 0; idx < num_squeeze_dims; ++idx) {
int current = squeeze_dims[idx] < 0 ? squeeze_dims[idx] + in_dims.size()
: squeeze_dims[idx];
// Check current index, the upper limit has beed checked in line 36.
PADDLE_ENFORCE(current >= 0,
"Invalid axis, the negative axis is out of range.");
PADDLE_ENFORCE(in_dims[current] == 1,
"Invalid axis index, the axis that will be squeezed "
"should be equal to 1.");
if (!(should_squeeze[current])) {
++cnt_squeezed_dims;
}
should_squeeze[current] = true;
}
}
// Make output dimensions
std::vector<int64_t> output_shape(in_dims.size() - cnt_squeezed_dims, 0);
for (int in_idx = 0, out_idx = 0; in_idx < in_dims.size(); ++in_idx) {
if (!should_squeeze[in_idx]) {
output_shape[out_idx++] = in_dims[in_idx];
}
}
return framework::make_ddim(output_shape);
}
};
class SqueezeOp : public framework::OperatorBase {
public:
using OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope &scope,
const platform::Place &place) const override {
auto &axes = Attr<std::vector<int>>("axes");
auto x_dims = scope.FindVar(Input("X"))->Get<framework::LoDTensor>().dims();
auto out_dims = SqueezeOpInferShape::GetOutputShape(axes, x_dims);
framework::AttributeMap attrs;
attrs["shape"] = framework::vectorize2int(out_dims);
attrs["inplace"] = Attr<bool>("inplace");
// Invoke Reshape Op
auto reshape_op = framework::OpRegistry::CreateOp(
"reshape", {{"X", {Input("X")}}, {"Shape", {}}},
{{"Out", {Output("Out")}}}, attrs);
reshape_op->Run(scope, place);
}
};
class SqueezeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "(Tensor). The input tensor of squeeze operator.");
AddOutput("Out", "(Tensor). The output tensor of squeeze operator.");
AddAttr<std::vector<int>>("axes",
"(std::vector<int>). List of integers,"
" indicating the dimensions to squeeze.")
.SetDefault({});
AddAttr<bool>("inplace",
"(default: false) Squeeze the source tensor's shape without "
"memory copy. When Attr(inplace) is set true, the output "
"tensor shares memory with Input(X), otherwise, a new output "
"tensor is created, and its data are copied from Input(x).")
.SetDefault(false);
AddComment(R"DOC(
Squeeze Operator.
Remove single-dimensional entries from the shape of a tensor.
Takes a parameter axes with a list of axes to squeeze.
If axes is not provided, all the single dimensions will be removed from the shape.
If an axis is selected with shape entry not equal to one, an error is raised.
Examples:
Case 1:
Given
X.shape = (1, 3, 1, 5)
and
axes = [0]
we get:
Out.shape = (3, 1, 5)
Case 2:
Given
X.shape = (1, 3, 1, 5)
and
axes = []
we get:
Out.shape = (3, 5)
)DOC");
}
};
class SqueezeGradInferShape : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext *context) const override {
context->SetOutputDim(framework::GradVarName("X"),
context->GetInputDim("X"));
context->ShareLoD("X", framework::GradVarName("X"));
}
};
class SqueezeGradOp : public framework::OperatorBase {
public:
using OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope &scope,
const platform::Place &place) const override {
auto dx_name = Output(framework::GradVarName("X"));
auto dout_name = Input(framework::GradVarName("Out"));
auto x_dims = scope.FindVar(Input("X"))->Get<framework::LoDTensor>().dims();
framework::AttributeMap attrs;
attrs["shape"] = framework::vectorize2int(x_dims);
attrs["inplace"] = Attr<bool>("inplace");
auto reshape_op = framework::OpRegistry::CreateOp(
"reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}},
attrs);
reshape_op->Run(scope, place);
}
};
} // namespace operators
} // namespace paddle
// Tell linker to use reshape op
USE_OP(reshape);
namespace ops = paddle::operators;
REGISTER_OPERATOR(squeeze, ops::SqueezeOp, ops::SqueezeOpMaker,
ops::SqueezeOpInferShape,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(squeeze_grad, ops::SqueezeGradOp, ops::SqueezeGradInferShape);
/* Copyright (c) 2018 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. */
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
class UnsqueezeOpInferShape : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of UnsqueezeOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of UnsqueezeOp should not be null.");
const auto &axes = ctx->Attrs().Get<std::vector<int>>("axes");
const auto &x_dims = ctx->GetInputDim("X");
// Validity Check: input tensor dims (<6).
PADDLE_ENFORCE(x_dims.size() <= 6,
"Invalid dimensions, the rank of Input(X) "
"should be in the range of [1, 6] (Eigen limit)");
auto out_dims = GetOutputShape(axes, x_dims);
ctx->SetOutputDim("Out", out_dims);
if (x_dims[0] == out_dims[0]) {
// Only pass LoD when the first dimension of output and Input(X)
// are the same.
ctx->ShareLoD("X", "Out");
}
}
static framework::DDim GetOutputShape(const std::vector<int> unsqz_dims,
const framework::DDim &in_dims) {
int output_size = in_dims.size() + static_cast<int>(unsqz_dims.size());
int cur_output_size = in_dims.size();
std::vector<int64_t> output_shape(output_size, 0);
// Validity Check: rank range.
PADDLE_ENFORCE(output_size <= 6,
"The output tensor's rank should be less than 6.");
for (int axis : unsqz_dims) {
int cur = axis < 0 ? axis + cur_output_size + 1 : axis;
// Vaildity Check: the axis bound
PADDLE_ENFORCE(
cur >= 0 && cur <= cur_output_size,
"The unsqueeze dims must be within range of current rank.");
// Move old axis, and insert new axis
for (int i = cur_output_size; i >= cur; --i) {
if (output_shape[i] == 1) {
// Move axis
output_shape[i + 1] = 1;
output_shape[i] = 0;
}
}
output_shape[cur] = 1;
// Add the output size.
cur_output_size++;
}
// Make output shape
for (int in_idx = 0, out_idx = 0; out_idx < output_size; ++out_idx) {
if (output_shape[out_idx] == 0) {
output_shape[out_idx] = in_dims[in_idx++];
}
}
return framework::make_ddim(output_shape);
}
};
class UnsqueezeOp : public framework::OperatorBase {
public:
using OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope &scope,
const platform::Place &place) const override {
auto &axes = Attr<std::vector<int>>("axes");
auto x_dims = scope.FindVar(Input("X"))->Get<framework::LoDTensor>().dims();
auto out_dims = UnsqueezeOpInferShape::GetOutputShape(axes, x_dims);
framework::AttributeMap attrs;
attrs["shape"] = framework::vectorize2int(out_dims);
attrs["inplace"] = Attr<bool>("inplace");
// Invoke Reshape op.
auto reshape_op = framework::OpRegistry::CreateOp(
"reshape", {{"X", {Input("X")}}, {"Shape", {}}},
{{"Out", {Output("Out")}}}, attrs);
reshape_op->Run(scope, place);
}
};
class UnsqueezeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "(Tensor). The input tensor of unsqueeze operator.");
AddOutput("Out", "(Tensor). The output tensor of unsqueeze operator.");
AddAttr<std::vector<int>>("axes",
"(std::vector<int>). List of integers,"
" indicating the dimensions to be inserted")
.AddCustomChecker([](const std::vector<int> &axes) {
PADDLE_ENFORCE(!axes.empty(),
"Invalid axes, The unsqueeze axes is empty.");
// Validity Check: axes dims (<6).
PADDLE_ENFORCE(static_cast<int>(axes.size()) < 6,
"Invalid dimensions, dynamic dimensions should be "
"within [1, 6] dimensions (Eigen limit).");
// Validity Check: the range of unsqueeze aixs.
for (int axis : axes) {
PADDLE_ENFORCE(axis < 6,
"Invalid dimensions, input axis should be"
" within [1, 6] dimensions (Eigen limit).");
}
});
AddAttr<bool>(
"inplace",
"(default: false) Unsqueeze the source tensor's shape without "
"memory copy. When Attr(inplace) is set true, the output "
"tensor shares memory with Input(X), otherwise, a new output "
"tensor is created, and its data are copied from Input(x).")
.SetDefault(false);
AddComment(R"DOC(
Unsqueeze Operator.
Insert single-dimensional entries to the shape of a tensor.
Takes one required argument axes, a list of dimensions that will be inserted.
Dimension indices in axes are as seen in the output tensor.
For example:
Given a tensor such that tensor with shape [3, 4, 5],
then Unsqueeze(tensor, axes=[0, 4]) has shape [1, 3, 4, 5, 1]
)DOC");
}
};
class UnsqueezeGradInferShape : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext *ctx) const override {
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
ctx->ShareLoD("X", framework::GradVarName("X"));
}
};
class UnsqueezeGradOp : public framework::OperatorBase {
public:
using OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope &scope,
const platform::Place &place) const override {
auto dx_name = Output(framework::GradVarName("X"));
auto dout_name = Input(framework::GradVarName("Out"));
auto x_dims = scope.FindVar(Input("X"))->Get<framework::LoDTensor>().dims();
framework::AttributeMap attrs;
attrs["shape"] = framework::vectorize2int(x_dims);
attrs["inplace"] = Attr<bool>("inplace");
auto reshape_op = framework::OpRegistry::CreateOp(
"reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}},
attrs);
reshape_op->Run(scope, place);
}
};
} // namespace operators
} // namespace paddle
// Tell linker to use reshape op.
USE_OP(reshape);
namespace ops = paddle::operators;
REGISTER_OPERATOR(unsqueeze, ops::UnsqueezeOp, ops::UnsqueezeOpMaker,
ops::UnsqueezeOpInferShape,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(unsqueeze_grad, ops::UnsqueezeGradOp,
ops::UnsqueezeGradInferShape);
......@@ -92,8 +92,15 @@ install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR}
DESTINATION opt/paddle/share/wheels
)
find_program(PATCHELF_EXECUTABLE patchelf)
if(NOT PATCHELF_EXECUTABLE)
message(FATAL_ERROR "patchelf not found, please install it.\n"
"For Ubuntu, the command is: apt-get install -y patchelf.")
endif()
if(APPLE)
find_program(INSTALL_NAME_TOOL_EXECUTABLE install_name_tool)
if(NOT INSTALL_NAME_TOOL_EXECUTABLE)
message(FATAL_ERROR "install_name_tool not found, please check.\n")
endif()
else(APPLE)
find_program(PATCHELF_EXECUTABLE patchelf)
if(NOT PATCHELF_EXECUTABLE)
message(FATAL_ERROR "patchelf not found, please install it.\n"
"For Ubuntu, the command is: apt-get install -y patchelf.")
endif()
endif(APPLE)
# Copyright (c) 2018 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.
import functools
import sys
__all__ = ['deprecated']
def deprecated(since, instead, extra_message=""):
def decorator(func):
err_msg = "API {0} is deprecated since {1}. Please use {2} instead.".format(
func.__name__, since, instead)
if len(extra_message) != 0:
err_msg += "\n"
err_msg += extra_message
@functools.wraps(func)
def wrapper(*args, **kwargs):
print >> sys.stderr, err_msg
return func(*args, **kwargs)
wrapper.__doc__ += "\n "
wrapper.__doc__ += err_msg
return wrapper
return decorator
......@@ -18,10 +18,7 @@ import collections
import copy
import unique_name
__all__ = [
'append_backward',
'calc_gradient',
]
__all__ = ['append_backward']
def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
......
......@@ -18,10 +18,12 @@ All util layers.
from layer_function_generator import autodoc
from ..framework import unique_name
from ..layer_helper import LayerHelper
from ..annotations import deprecated
__all__ = ['get_places']
__all__ = []
@deprecated(since='0.15.0', instead="ParallelExecutor")
@autodoc()
def get_places(device_count=None, device_type=None):
helper = LayerHelper('get_places', **locals())
......
......@@ -29,7 +29,7 @@ __all__ = [
'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer',
'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'Optimizer', 'RMSPropOptimizer'
'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'RMSPropOptimizer'
]
......@@ -67,7 +67,7 @@ class Optimizer(object):
self._LARS_weight_decay = LARS_weight_decay
def _create_global_learning_rate(self):
lr = self.global_learning_rate()
lr = self._global_learning_rate()
if isinstance(lr, framework.Variable):
return
......@@ -86,7 +86,7 @@ class Optimizer(object):
dtype='float32' if self._dtype == None else self._dtype,
persistable=True)
def global_learning_rate(self, program=None):
def _global_learning_rate(self, program=None):
"""
get global decayed learning rate
:return:
......@@ -110,9 +110,9 @@ class Optimizer(object):
return param_lr
else:
if param_lr == 1.0:
return self.global_learning_rate()
return self._global_learning_rate()
else:
return self.global_learning_rate() * param_lr
return self._global_learning_rate() * param_lr
def _create_accumulators(self, block, parameters):
"""Create all accumulators needed by the parameters
......@@ -185,10 +185,10 @@ class Optimizer(object):
format(name, param.name))
return self._accumulators[name][param.name]
def create_optimization_pass(self,
parameters_and_grads,
loss,
startup_program=None):
def _create_optimization_pass(self,
parameters_and_grads,
loss,
startup_program=None):
"""Add optimization operators to update gradients to variables.
Args:
......@@ -221,7 +221,7 @@ class Optimizer(object):
self._create_global_learning_rate()
if self._LARS_weight_decay > 0.0:
layers.append_LARS(parameters_and_grads,
self.global_learning_rate(),
self._global_learning_rate(),
self._LARS_weight_decay)
optimize_ops = []
......@@ -263,8 +263,8 @@ class Optimizer(object):
params_grads = append_regularization_ops(params_grads,
self.regularization)
optimize_ops = self.create_optimization_pass(params_grads, loss,
startup_program)
optimize_ops = self._create_optimization_pass(params_grads, loss,
startup_program)
return optimize_ops, params_grads
......
......@@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
from paddle.fluid.layers.device import get_places
import unittest
import paddle.fluid as fluid
import paddle
......@@ -144,7 +144,7 @@ def train(word_dict,
cost, acc_out, prediction = net_method(
data, label, input_dim=dict_dim, class_dim=class_dim)
else:
places = fluid.layers.get_places()
places = get_places()
pd = fluid.layers.ParallelDo(places)
with pd.do():
cost, acc, _ = net_method(
......
......@@ -12,15 +12,17 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
import argparse
import paddle.fluid as fluid
import paddle
import sys
import numpy
import unittest
import math
import sys
import os
import sys
import unittest
import numpy
import paddle
import paddle.fluid as fluid
from paddle.fluid.layers.device import get_places
BATCH_SIZE = 64
......@@ -76,7 +78,7 @@ def train(nn_type,
net_conf = conv_net
if parallel:
places = fluid.layers.get_places()
places = get_places()
pd = fluid.layers.ParallelDo(places)
with pd.do():
img_ = pd.read_input(img)
......
......@@ -14,6 +14,7 @@
import paddle
import paddle.fluid as fluid
from paddle.fluid.layers.device import get_places
import unittest
import os
import numpy as np
......@@ -80,7 +81,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
avg_cost, predict_word = __network__(
[first_word, second_word, third_word, forth_word, next_word])
else:
places = fluid.layers.get_places()
places = get_places()
pd = fluid.layers.ParallelDo(places)
with pd.do():
avg_cost, predict_word = __network__(
......
......@@ -12,12 +12,13 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle
import paddle.fluid as fluid
import math
import sys
import paddle
import paddle.fluid as fluid
from paddle.fluid.layers.device import get_places
# need to fix random seed and training data to compare the loss
# value accurately calculated by the default and the memory optimization
# version.
......@@ -34,7 +35,7 @@ if fluid.core.is_compiled_with_cuda():
use_nccl = False
place = fluid.CUDAPlace(0)
places = fluid.layers.get_places(device_count=0, device_type=device_type)
places = get_places(device_count=0, device_type=device_type)
pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl)
with pd.do():
x_ = pd.read_input(x)
......
......@@ -16,8 +16,6 @@ import unittest
import paddle.fluid as fluid
import paddle.fluid.layers as layers
import paddle.fluid.framework as framework
import paddle.fluid.optimizer as optimizer
from paddle.fluid.backward import calc_gradient
......
......@@ -13,6 +13,7 @@
# limitations under the License.
import paddle.fluid as fluid
from paddle.fluid.layers.device import get_places
import decorators
import unittest
......@@ -20,7 +21,7 @@ import unittest
class TestGetPlaces(unittest.TestCase):
@decorators.prog_scope()
def test_get_places(self):
places = fluid.layers.get_places()
places = get_places()
cpu = fluid.CPUPlace()
exe = fluid.Executor(cpu)
exe.run(fluid.default_main_program())
......
......@@ -16,6 +16,7 @@ from __future__ import print_function
import unittest
import paddle.fluid.layers as layers
from paddle.fluid.layers.device import get_places
import paddle.fluid.nets as nets
from paddle.fluid.framework import Program, program_guard, default_main_program
from paddle.fluid.param_attr import ParamAttr
......@@ -238,7 +239,7 @@ class TestBook(unittest.TestCase):
def test_get_places(self):
program = Program()
with program_guard(program):
x = layers.get_places(device_count=4)
x = get_places(device_count=4)
self.assertIsNotNone(x)
print(str(program))
......
......@@ -97,7 +97,7 @@ class TestMomentumOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
opts = momentum_optimizer.create_optimization_pass(
opts = momentum_optimizer._create_optimization_pass(
params_grads, mul_out, init_program)
self.assertEqual(len(opts), 3)
sgd_op = opts[-1]
......@@ -151,7 +151,7 @@ class TestMomentumOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
opts = momentum_optimizer.create_optimization_pass(
opts = momentum_optimizer._create_optimization_pass(
params_grads, mul_out, init_program)
self.assertEqual(len(opts), 3)
sgd_op = opts[-1]
......@@ -214,8 +214,8 @@ class TestAdagradOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0)
opts = adagrad_optimizer.create_optimization_pass(params_grads, mul_out,
init_program)
opts = adagrad_optimizer._create_optimization_pass(
params_grads, mul_out, init_program)
self.assertEqual(len(opts), 3)
self.assertEqual([op.type for op in opts],
["fill_constant", "elementwise_mul", "adagrad"])
......@@ -278,8 +278,8 @@ class TestAdamOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(adam_optimizer.get_accumulators()), 0)
opts = adam_optimizer.create_optimization_pass(params_grads, mul_out,
init_program)
opts = adam_optimizer._create_optimization_pass(params_grads, mul_out,
init_program)
self.assertEqual(len(opts), 5)
self.assertEqual(
[op.type for op in opts],
......@@ -345,8 +345,8 @@ class TestAdamaxOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(adamax_optimizer.get_accumulators()), 0)
opts = adamax_optimizer.create_optimization_pass(params_grads, mul_out,
init_program)
opts = adamax_optimizer._create_optimization_pass(params_grads, mul_out,
init_program)
self.assertEqual(len(opts), 4)
self.assertEqual(
[op.type for op in opts],
......@@ -409,7 +409,7 @@ class TestDecayedAdagradOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0)
opts = decayed_adagrad_optimizer.create_optimization_pass(
opts = decayed_adagrad_optimizer._create_optimization_pass(
params_grads, mul_out, init_program)
self.assertEqual(len(opts), 3)
self.assertEqual(
......@@ -475,8 +475,8 @@ class TestFtrlOptimizer(unittest.TestCase):
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0)
opts = ftrl_optimizer.create_optimization_pass(params_grads, mul_out,
init_program)
opts = ftrl_optimizer._create_optimization_pass(params_grads, mul_out,
init_program)
self.assertEqual(len(opts), 3)
self.assertEqual([op.type for op in opts],
["fill_constant", "elementwise_mul", "ftrl"])
......
......@@ -15,6 +15,7 @@
import unittest
import paddle.fluid as fluid
from paddle.fluid.layers.device import get_places
import paddle.fluid.profiler as profiler
import numpy
......@@ -115,7 +116,7 @@ class BaseParallelForTest(unittest.TestCase):
if use_parallel:
thread_num = fluid.core.get_cuda_device_count(
) if use_gpu else 8
places = fluid.layers.get_places(thread_num)
places = get_places(thread_num)
pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl)
data = next(generator)
......
# Copyright (c) 2018 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.
import unittest
import numpy as np
from op_test import OpTest
# Correct: General.
class TestSqueezeOp(OpTest):
def setUp(self):
self.op_type = "squeeze"
self.init_test_case()
self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")}
self.init_attrs()
self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(["X"], "Out")
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = (0, 2)
self.new_shape = (3, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": False}
# Correct: There is mins axis.
class TestSqueezeOp1(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = (0, -2)
self.new_shape = (3, 5)
# Correct: No axes input.
class TestSqueezeOp2(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = ()
self.new_shape = (3, 5)
# Correct: Just part of axes be squeezed.
class TestSqueezeOp3(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 1, 5, 1, 4, 1)
self.axes = (1, -1)
self.new_shape = (3, 5, 1, 4)
# Correct: Inplace.
class TestSqueezeOpInplace1(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = (0, 2)
self.new_shape = (3, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
# Correct: Inplace. There is mins axis.
class TestSqueezeOpInplace2(TestSqueezeOp):
def inti_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = (0, -2)
self.new_shape = (3, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
# Correct: Inplace. No axes input.
class TestSqueezeOpInplace3(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (1, 3, 1, 5)
self.axes = ()
self.new_shape = (3, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
# Correct: Inpalce. Just part of axes be squeezed.
class TestSqueezeOpInplace4(TestSqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 1, 5, 1, 4, 1)
self.axes = (1, -1)
self.new_shape = (3, 5, 1, 4)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
if __name__ == "__main__":
unittest.main()
# Copyright (c) 2018 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.
import unittest
import numpy as np
from op_test import OpTest
# Correct: General.
class TestUnsqueezeOp(OpTest):
def setUp(self):
self.init_test_case()
self.op_type = "unsqueeze"
self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")}
self.init_attrs()
self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(["X"], "Out")
def init_test_case(self):
self.ori_shape = (3, 5)
self.axes = (1, 2)
self.new_shape = (3, 1, 1, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": False}
# Correct: Single input index.
class TestUnsqueezeOp1(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 5)
self.axes = (-1, )
self.new_shape = (3, 5, 1)
# Correct: Mixed input axis.
class TestUnsqueezeOp2(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 5)
self.axes = (0, -1)
self.new_shape = (1, 3, 5, 1)
# Correct: There is duplicated axis.
class TestUnsqueezeOp3(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 2, 5)
self.axes = (0, 3, 3)
self.new_shape = (1, 3, 2, 1, 1, 5)
# Correct: Reversed axes.
class TestUnsqueezeOp4(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 2, 5)
self.axes = (3, 1, 1)
self.new_shape = (3, 1, 1, 2, 5, 1)
# Correct: Inplace.
class TestUnsqueezeOpInplace1(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 5)
self.axes = (0, 2)
self.new_shape = (1, 3, 1, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
# Correct: Inplace. There is mins index.
class TestUnsqueezeOpInplace2(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 5)
self.axes = (0, -2)
self.new_shape = (1, 3, 1, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
# Correct: Inplace. There is duplicated axis.
class TestUnsqueezeOpInplace3(TestUnsqueezeOp):
def init_test_case(self):
self.ori_shape = (3, 2, 5)
self.axes = (0, 3, 3)
self.new_shape = (1, 3, 2, 1, 1, 5)
def init_attrs(self):
self.attrs = {"axes": self.axes, "inplace": True}
if __name__ == "__main__":
unittest.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