Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into format
weixing02
7 years ago
commit
246f613538
@ -0,0 +1,127 @@
|
||||
# How to use RecordIO in Fluid
|
||||
|
||||
If you want to use RecordIO as your training data format, you need to convert to your training data
|
||||
to RecordIO files and reading them in the process of training, PaddlePaddle Fluid provides some
|
||||
interface to deal with the RecordIO files.
|
||||
|
||||
## Generate RecordIO File
|
||||
|
||||
Before start training with RecordIO files, you need to convert your training data
|
||||
to RecordIO format by `fluid.recordio_writer.convert_reader_to_recordio_file`, the sample codes
|
||||
as follows:
|
||||
|
||||
```python
|
||||
reader = paddle.batch(mnist.train(), batch_size=1)
|
||||
feeder = fluid.DataFeeder(
|
||||
feed_list=[ # order is image and label
|
||||
fluid.layers.data(
|
||||
name='image', shape=[784]),
|
||||
fluid.layers.data(
|
||||
name='label', shape=[1], dtype='int64'),
|
||||
],
|
||||
place=fluid.CPUPlace())
|
||||
fluid.recordio_writer.convert_reader_to_recordio_file('./mnist.recordio', reader, feeder)
|
||||
```
|
||||
|
||||
The above code snippet would generate a RecordIO `./mnist.recordio` on your host.
|
||||
|
||||
**NOTE**: we recommend users to set `batch_size=1` when generating the recordio files so that users can
|
||||
adjust it flexibly while reading it.
|
||||
|
||||
## Use the RecordIO file in a Local Training Job
|
||||
|
||||
PaddlePaddle Fluid provides an interface `fluid.layers.io.open_recordio_file` to load your RecordIO file
|
||||
and then you can use them as a Layer in your network configuration, the sample codes as follows:
|
||||
|
||||
```python
|
||||
data_file = fluid.layers.io.open_recordio_file(
|
||||
filename="./mnist.recordio",
|
||||
shapes=[(-1, 784),(-1, 1)],
|
||||
lod_levels=[0, 0],
|
||||
dtypes=["float32", "int32"])
|
||||
data_file = fluid.layers.io.batch(data_file, batch_size=4)
|
||||
|
||||
img, label = fluid.layers.io.read_file(data_file)
|
||||
hidden = fluid.layers.fc(input=img, size=100, act='tanh')
|
||||
prediction = fluid.layers.fc(input=hidden, size=10, act='softmax')
|
||||
loss = fluid.layers.cross_entropy(input=prediction, label=label)
|
||||
avg_loss = fluid.layers.mean(loss)
|
||||
|
||||
fluid.optimizer.Adam(learning_rate=1e-3).minimize(avg_loss)
|
||||
|
||||
place = fluid.CPUPlace()
|
||||
|
||||
exe = fluid.Executor(place)
|
||||
exe.run(fluid.default_startup_program())
|
||||
avg_loss_np = []
|
||||
|
||||
# train a pass
|
||||
batch_id = 0
|
||||
while True:
|
||||
tmp, = exe.run(fetch_list=[avg_loss])
|
||||
|
||||
avg_loss_np.append(tmp)
|
||||
print(batch_id)
|
||||
batch_id += 1
|
||||
```
|
||||
|
||||
## Use the RecordIO files in Distributed Training
|
||||
|
||||
1. generate multiple RecordIO files
|
||||
|
||||
For a distributed training job, you may have multiple trainer nodes,
|
||||
and one or more RecordIO files for one trainer node, you can use the interface
|
||||
`fluid.recordio_writer.convert_reader_to_recordio_files` to convert your training data
|
||||
into multiple RecordIO files, the sample codes as follows:
|
||||
|
||||
```python
|
||||
reader = paddle.batch(mnist.train(), batch_size=1)
|
||||
feeder = fluid.DataFeeder(
|
||||
feed_list=[ # order is image and label
|
||||
fluid.layers.data(
|
||||
name='image', shape=[784]),
|
||||
fluid.layers.data(
|
||||
name='label', shape=[1], dtype='int64'),
|
||||
],
|
||||
place=fluid.CPUPlace())
|
||||
fluid.recordio_writer.convert_reader_to_recordio_files(
|
||||
filename_suffix='./mnist.recordio', batch_per_file=100, reader, feeder)
|
||||
```
|
||||
|
||||
The above codes would generate multiple RecordIO files on your host like:
|
||||
|
||||
```bash
|
||||
.
|
||||
\_mnist-00000.recordio
|
||||
|-mnist-00001.recordio
|
||||
|-mnist-00002.recordio
|
||||
|-mnist-00003.recordio
|
||||
|-mnist-00004.recordio
|
||||
```
|
||||
|
||||
2. open multiple RecordIO files by `fluid.layers.io.open_files`
|
||||
|
||||
For a distributed training job, the distributed operator system will schedule trainer process on multiple nodes,
|
||||
each trainer process reads parts of the whole training data, we usually take the following approach to make the training
|
||||
data allocated by each trainer process as uniform as possiable:
|
||||
|
||||
```python
|
||||
def gen_train_list(file_pattern, trainers, trainer_id):
|
||||
file_list = glob.glob(file_pattern)
|
||||
ret_list = []
|
||||
for idx, f in enumerate(file_list):
|
||||
if (idx + trainers) % trainers == trainer_id:
|
||||
ret_list.append(f)
|
||||
return ret_list
|
||||
|
||||
trainers = int(os.getenv("TRAINERS"))
|
||||
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
|
||||
data_file = fluid.layers.io.open_files(
|
||||
filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0),
|
||||
thread_num=1,
|
||||
shapes=[(-1, 784),(-1, 1)],
|
||||
lod_levels=[0, 0],
|
||||
dtypes=["float32", "int32"])
|
||||
img, label = fluid.layers.io.read_file(data_files)
|
||||
...
|
||||
```
|
||||
@ -1 +0,0 @@
|
||||
../../../../../benchmark/cluster/README.md
|
||||
@ -1 +0,0 @@
|
||||
../../../../../../benchmark/cluster/vgg16/README.md
|
||||
@ -0,0 +1,82 @@
|
||||
// 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 <cuda.h>
|
||||
|
||||
#include "paddle/contrib/inference/paddle_inference_api_anakin_engine.h"
|
||||
|
||||
namespace paddle {
|
||||
|
||||
PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor(
|
||||
const AnakinConfig &config) {
|
||||
CHECK(Init(config));
|
||||
}
|
||||
|
||||
bool PaddleInferenceAnakinPredictor::Init(const AnakinConfig &config) {
|
||||
// TODO(Superjomn) Tell anakin to support return code.
|
||||
engine_.Build(config.model_file, config.max_batch_size);
|
||||
return true;
|
||||
}
|
||||
|
||||
bool PaddleInferenceAnakinPredictor::Run(
|
||||
const std::vector<PaddleTensor> &inputs,
|
||||
std::vector<PaddleTensor> *output_data) {
|
||||
for (const auto &input : inputs) {
|
||||
if (input.dtype != PaddleDType::FLOAT32) {
|
||||
LOG(ERROR) << "Only support float type inputs. " << input.name
|
||||
<< "'s type is not float";
|
||||
return false;
|
||||
}
|
||||
engine_.SetInputFromCPU(
|
||||
input.name, static_cast<float *>(input.data.data), input.data.length);
|
||||
}
|
||||
|
||||
// TODO(Superjomn) Tell anakin to support return code.
|
||||
engine_.Execute();
|
||||
|
||||
if (output_data->empty()) {
|
||||
LOG(ERROR) << "At least one output should be set with tensors' names.";
|
||||
return false;
|
||||
}
|
||||
for (auto &output : *output_data) {
|
||||
auto *tensor = engine_.GetOutputInGPU(output.name);
|
||||
output.shape = tensor->shape();
|
||||
// Copy data from GPU -> CPU
|
||||
if (cudaMemcpy(output.data.data,
|
||||
tensor->data(),
|
||||
tensor->size(),
|
||||
cudaMemcpyDeviceToHost) != 0) {
|
||||
LOG(ERROR) << "copy data from GPU to CPU error";
|
||||
return false;
|
||||
}
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
// TODO(Superjomn) To implement latter.
|
||||
std::unique_ptr<PaddlePredictor> PaddleInferenceAnakinPredictor::Clone() {
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
// A factory to help create difference predictor.
|
||||
template <>
|
||||
std::unique_ptr<PaddlePredictor>
|
||||
CreatePaddlePredictor<AnakinConfig, PaddleEngineKind::kAnakin>(
|
||||
const AnakinConfig &config) {
|
||||
std::unique_ptr<PaddlePredictor> x(
|
||||
new PaddleInferenceAnakinPredictor(config));
|
||||
return x;
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,51 @@
|
||||
/* 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. */
|
||||
|
||||
/*
|
||||
* This file contains the implementation of inference API with Anakin engine
|
||||
* embeded, this API can only support Anakin models.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
// NOTE This header file do not have namespace.
|
||||
// TODO(Superjomn) Tell Anakin to provide better APIs.
|
||||
#include <test/framework/net/paddle_api.h>
|
||||
#include "paddle/contrib/inference/paddle_inference_api.h"
|
||||
|
||||
namespace paddle {
|
||||
|
||||
class PaddleInferenceAnakinPredictor : public PaddlePredictor {
|
||||
public:
|
||||
PaddleInferenceAnakinPredictor(const AnakinConfig& config);
|
||||
|
||||
// NOTE Unlike the native engine, the buffers of anakin engine's output_data
|
||||
// should be allocated first.
|
||||
// TODO(Superjomn) should unify all the behaviors of output_data accross all
|
||||
// the engines.
|
||||
bool Run(const std::vector<PaddleTensor>& inputs,
|
||||
std::vector<PaddleTensor>* output_data) override;
|
||||
|
||||
std::unique_ptr<PaddlePredictor> Clone() override;
|
||||
|
||||
private:
|
||||
bool Init(const AnakinConfig& config);
|
||||
|
||||
anakin::AnakinEngine<anakin::NV,
|
||||
anakin::saber::AK_FLOAT,
|
||||
anakin::Precision::FP32>
|
||||
engine_;
|
||||
};
|
||||
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,27 @@
|
||||
/* 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 "paddle/contrib/inference/paddle_inference_api.h"
|
||||
#include <gtest/gtest.h>
|
||||
|
||||
namespace paddle {
|
||||
|
||||
TEST(inference, anakin) {
|
||||
AnakinConfig config;
|
||||
|
||||
auto engine =
|
||||
CreatePaddlePredictor<AnakinConfig, PaddleEngineKind::kAnakin>(config);
|
||||
}
|
||||
|
||||
} // namespace paddle
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue