Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 bfd42683 编写于 作者: T tensor-tang
浏览文件
操作

Merge remote-tracking branch 'ups/develop' into refine

上级 0693b414 635099c1
显示空白变更内容
内联 并排
Showing with 1451 addition and 556 deletion
benchmark/fluid/Dockerfile
浏览文件 @ bfd42683
......@@ -19,4 +19,4 @@ ADD *.whl /
RUN pip install /*.whl && rm -f /*.whl && chmod +x /usr/bin/paddle_k8s
ENV LD_LIBRARY_PATH=/usr/local/lib
ADD fluid_benchmark.py dataset.py models/ /workspace/
ADD fluid_benchmark.py recordio_converter.py models/ /workspace/
benchmark/fluid/README.md
浏览文件 @ bfd42683
......@@ -44,6 +44,16 @@ Currently supported `--model` argument include:
PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method nccl2
```
## Prepare the RecordIO file to Achieve Better Performance
Run the following command will generate RecordIO files like "mnist.recordio" under the path
and batch_size you choose, you can use batch_size=1 so that later reader can change the batch_size
at any time using `fluid.batch`.
```bash
python -c 'from recordio_converter import *; prepare_mnist("data", 1)'
```
## Run Distributed Benchmark on Kubernetes Cluster
You may need to build a Docker image before submitting a cluster job onto Kubernetes, or you will
......
benchmark/fluid/fluid_benchmark.py
浏览文件 @ bfd42683
......@@ -38,10 +38,12 @@ def parse_args():
default='resnet',
help='The model to run benchmark with.')
parser.add_argument(
'--batch_size', type=int, default=32, help='The minibatch size.')
'--batch_size',
type=int,
default=32,
help='The batch size on each gpu.')
parser.add_argument(
'--learning_rate', type=float, default=0.001, help='The learning rate.')
# TODO(wuyi): add "--use_fake_data" option back.
parser.add_argument(
'--skip_batch_num',
type=int,
......@@ -49,7 +51,10 @@ def parse_args():
help='The first num of minibatch num to skip, for better performance test'
)
parser.add_argument(
'--iterations', type=int, default=80, help='The number of minibatches.')
'--iterations',
type=int,
default=80,
help='The number of minibatches, set to -1 to run all batches.')
parser.add_argument(
'--pass_num', type=int, default=100, help='The number of passes.')
parser.add_argument(
......@@ -69,6 +74,7 @@ def parse_args():
type=int,
default=1,
help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
# this option is available only for vgg and resnet.
parser.add_argument(
'--cpus',
type=int,
......@@ -78,7 +84,7 @@ def parse_args():
'--data_set',
type=str,
default='flowers',
choices=['cifar10', 'flowers'],
choices=['cifar10', 'flowers', 'imagenet'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--infer_only', action='store_true', help='If set, run forward only.')
......@@ -108,6 +114,16 @@ def parse_args():
default='local',
choices=['local', 'pserver', 'nccl2'],
help='Choose parameter update method, can be local, pserver, nccl2.')
parser.add_argument(
'--use_reader_op',
action='store_true',
help='Whether to use reader op, and must specify the data path if set this to true.'
)
parser.add_argument(
'--data_path',
type=str,
default="",
help='Directory that contains all the training recordio files.')
args = parser.parse_args()
return args
......@@ -210,6 +226,8 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(startup_prog)
if not args.use_reader_op:
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
......@@ -219,16 +237,38 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
train_losses = []
for batch_id, data in enumerate(train_reader()):
if not args.use_reader_op:
reader_generator = train_reader()
batch_id = 0
data = None
while True:
if not args.use_reader_op:
data = next(reader_generator, None)
if data == None:
break
if iters == args.iterations:
break
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
if args.use_reader_op:
try:
loss = exe.run(train_prog, fetch_list=[avg_loss])
except fluid.core.EnforceNotMet as ex:
break
else:
loss = exe.run(train_prog,
feed=feeder.feed(data),
fetch_list=[avg_loss])
iters += 1
batch_id += 1
# FIXME(wuyi): For use_reader_op, if the current
# pass is not the last, the last batch of this pass
# is also equal to args.batch_size.
if args.use_reader_op:
num_samples += args.batch_size * args.gpus
else:
num_samples += len(data)
train_losses.append(loss)
print("Pass: %d, Iter: %d, Loss: %f\n" %
......@@ -250,10 +290,14 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
batch_acc, args, train_prog, startup_prog, nccl_id_var,
num_trainers, trainer_id):
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
if not args.use_reader_op:
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
# generate fake:
if args.use_fake_data:
for var in feed_var_list:
......@@ -270,7 +314,6 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
"value": 1.0,
"dtype": var.dtype})
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
if nccl_id_var and trainer_id == 0:
#FIXME(wuyi): wait other trainer to start listening
time.sleep(30)
......@@ -287,12 +330,21 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
num_trainers=num_trainers,
trainer_id=trainer_id)
feeder = fluid.DataFeeder(feed_var_list, place)
for pass_id in range(args.pass_num):
num_samples = 0
iters = 0
start_time = time.time()
for batch_id, data in enumerate(train_reader()):
if not args.use_reader_op:
reader_generator = train_reader()
batch_id = 0
data = None
while True:
if not args.use_reader_op:
data = next(reader_generator, None)
if data == None:
break
if iters == args.iterations:
break
if args.profile and pass_id == 0 and batch_id == 5:
profiler.start_profiler("All")
elif args.profile and pass_id == 0 and batch_id == 10:
......@@ -301,19 +353,26 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
if args.use_fake_data:
if args.use_fake_data or args.use_reader_op:
try:
loss, = exe.run([avg_loss.name])
except fluid.core.EnforceNotMet as ex:
break
else:
loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
if args.update_method == "pserver":
exe.bcast_params()
if args.use_reader_op:
num_samples += args.batch_size * args.gpus
else:
num_samples += len(data)
iters += 1
if batch_id % 1 == 0:
print("Pass %d, batch %d, loss %s" %
(pass_id, batch_id, np.array(loss)))
batch_id += 1
if args.use_reader_op:
num_samples = num_samples * args.gpus
print_train_time(start_time, time.time(), num_samples)
if not args.no_test and batch_acc:
test_acc = test(startup_exe, infer_prog, test_reader, feeder,
......
benchmark/fluid/models/machine_translation.py
浏览文件 @ bfd42683
......@@ -197,6 +197,8 @@ def lodtensor_to_ndarray(lod_tensor):
def get_model(args):
if args.use_reader_op:
raise Exception("machine_translation do not support reader op for now.")
embedding_dim = 512
encoder_size = 512
decoder_size = 512
......@@ -221,7 +223,7 @@ def get_model(args):
train_batch_generator = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.train(dict_size), buf_size=1000),
batch_size=args.batch_size)
batch_size=args.batch_size * args.gpus)
test_batch_generator = paddle.batch(
paddle.reader.shuffle(
......
benchmark/fluid/models/mnist.py
浏览文件 @ bfd42683
......@@ -20,6 +20,7 @@ import numpy as np
import argparse
import time
import cProfile
import os
import paddle
import paddle.fluid as fluid
......@@ -65,7 +66,22 @@ def cnn_model(data):
def get_model(args):
# Input data
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
data_file = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1, 1, 28, 28], (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=args.gpus,
pass_num=args.pass_num)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file, batch_size=args.batch_size))
images, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
......@@ -103,7 +119,7 @@ def get_model(args):
# Reader
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=args.batch_size)
paddle.dataset.mnist.train(), batch_size=args.batch_size * args.gpus)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=args.batch_size)
return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc
benchmark/fluid/models/resnet.py
浏览文件 @ bfd42683
......@@ -19,6 +19,7 @@ from __future__ import print_function
import functools
import numpy as np
import time
import os
import cProfile, pstats, StringIO
......@@ -26,6 +27,7 @@ import paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
from recordio_converter import imagenet_train, imagenet_test
def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
......@@ -122,14 +124,46 @@ def get_model(args):
else:
dshape = [32, 32, 3]
model = resnet_cifar10
else:
train_reader = paddle.dataset.cifar.train10()
test_reader = paddle.dataset.cifar.test10()
elif args.data_set == "flowers":
class_dim = 102
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
train_reader = paddle.dataset.flowers.train()
test_reader = paddle.dataset.flowers.test()
elif args.data_set == "imagenet":
class_dim = 1000
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
if not args.data_path:
raise Exception(
"Must specify --data_path when training with imagenet")
train_reader = imagenet_train(args.data_path)
test_reader = imagenet_test(args.data_path)
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
data_file = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1] + dshape, (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=args.gpus,
pass_num=args.pass_num)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file, batch_size=args.batch_size))
input, label = fluid.layers.read_file(data_file)
else:
input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
......@@ -162,15 +196,10 @@ def get_model(args):
optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
train_reader = paddle.batch(
batched_train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
test_reader = paddle.batch(
paddle.dataset.cifar.test10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
batch_size=args.batch_size)
return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc
train_reader, buf_size=5120),
batch_size=args.batch_size * args.gpus)
batched_test_reader = paddle.batch(train_reader, batch_size=args.batch_size)
return avg_cost, inference_program, optimizer, batched_train_reader, batched_test_reader, batch_acc
benchmark/fluid/models/stacked_dynamic_lstm.py
浏览文件 @ bfd42683
......@@ -44,6 +44,9 @@ def crop_sentence(reader, crop_size):
def get_model(args):
if args.use_reader_op:
raise Exception(
"stacked_dynamic_lstm do not support reader op for now.")
lstm_size = 512
emb_dim = 512
crop_size = 1500
......@@ -114,7 +117,7 @@ def get_model(args):
train_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.train(word_dict), crop_size), buf_size=25000),
batch_size=args.batch_size)
batch_size=args.batch_size * args.gpus)
test_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.test(word_dict), crop_size), buf_size=25000),
......
benchmark/fluid/models/vgg.py
浏览文件 @ bfd42683
......@@ -22,6 +22,7 @@ import paddle.fluid as fluid
import paddle.fluid.core as core
import argparse
import functools
import os
def vgg16_bn_drop(input):
......@@ -65,8 +66,23 @@ def get_model(args):
else:
data_shape = [224, 224, 3]
# Input data
images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
data_file = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1] + data_shape, (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=args.gpus,
pass_num=args.pass_num)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file, batch_size=args.batch_size))
images, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Train program
......@@ -95,7 +111,7 @@ def get_model(args):
paddle.dataset.cifar.train10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
batch_size=args.batch_size * args.gpus)
test_reader = paddle.batch(
paddle.dataset.cifar.test10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
......
benchmark/fluid/recordio_converter.py 0 → 100644
浏览文件 @ bfd42683
# 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 os
import random
import paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
from paddle.dataset import mnist, cifar, flowers, image
def convert_2_recordio(py_reader, outfilepath, batch_size, shape_data,
shape_label):
num_batches = 0
with fluid.program_guard(fluid.Program(), fluid.Program()):
reader = paddle.batch(py_reader(), batch_size=batch_size)
feeder = fluid.DataFeeder(
feed_list=[ # order is image and label
fluid.layers.data(
name='image', shape=shape_data),
fluid.layers.data(
name='label', shape=shape_label, dtype='int64'),
],
place=fluid.CPUPlace())
num_batches = fluid.recordio_writer.convert_reader_to_recordio_file(
outfilepath, reader, feeder)
return num_batches
def prepare_mnist(outpath, batch_size):
outfilepath = os.path.join(outpath, "mnist.recordio")
convert_2_recordio(mnist.train, outfilepath, batch_size, [784], [1])
def prepare_cifar10(outpath, batch_size):
outfilepath = os.path.join(outpath, "cifar.recordio")
convert_2_recordio(cifar.train10, outfilepath, batch_size, [3, 32, 32], [1])
def prepare_flowers(outpath, batch_size):
outfilepath = os.path.join(outpath, "flowers.recordio")
convert_2_recordio(flowers.train, outfilepath, batch_size, [3, 224, 224],
[1])
def default_mapper(sample):
img, label = sample
img = image.simple_transform(
img, 256, 224, True, mean=[103.94, 116.78, 123.68])
return img.flatten().astype('float32'), label
def imagenet_train(data_dir):
contents = os.listdir(data_dir)
if set(contents) != set(
["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]):
raise Exception("Imagenet data contents error!")
img2label = dict()
imgfilelist = []
with open(os.path.join(data_dir, "train.txt")) as fn:
while 1:
l = fn.readline()
if not l:
break
img, lbl = l[:-1].split(" ")
img2label[img] = int(lbl)
imgfilelist.append(img)
# shuffle all, this is slow
random.shuffle(imgfilelist)
def train_reader():
for idx, imgfile in enumerate(imgfilelist):
data = image.load_image(
os.path.join(data_dir, "train", imgfile.lower()))
label = [img2label[imgfile], ]
yield [data, label]
return paddle.reader.map_readers(default_mapper, train_reader)
def imagenet_test(data_dir):
contents = os.listdir(data_dir)
if set(contents) != set(
["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]):
raise Exception("Imagenet data contents error!")
img2label = dict()
imgfilelist = []
with open(os.path.join(data_dir, "val.txt")) as fn:
while 1:
l = fn.readline()
if not l:
break
img, lbl = l[:-1].split(" ")
img2label[img] = int(lbl)
imgfilelist.append(img)
def test_reader():
for idx, imgfile in enumerate(imgfilelist):
base_path = os.path.join(data_dir, "val", imgfile.split(".")[0])
image_path = ".".join([base_path, "jpeg"])
data = image.load_image(image_path)
label = [img2label[imgfile], ]
yield [data, label]
return paddle.reader.map_readers(default_mapper, test_reader)
# FIXME(wuyi): delete this when https://github.com/PaddlePaddle/Paddle/pull/11066 is merged
def convert_reader_to_recordio_files(
filename,
batch_per_file,
reader_creator,
feeder,
compressor=core.RecordIOWriter.Compressor.Snappy,
max_num_records=1000,
feed_order=None):
if feed_order is None:
feed_order = feeder.feed_names
f_name, f_ext = os.path.splitext(filename)
assert (f_ext == ".recordio")
lines = []
f_idx = 0
counter = 0
for idx, batch in enumerate(reader_creator()):
lines.append(batch)
if idx >= batch_per_file and idx % batch_per_file == 0:
filename = "%s-%05d%s" % (f_name, f_idx, f_ext)
with fluid.recordio_writer.create_recordio_writer(
filename, compressor, max_num_records) as writer:
for l in lines:
res = feeder.feed(l)
for each in feed_order:
writer.append_tensor(res[each])
writer.complete_append_tensor()
counter += 1
lines = []
f_idx += 1
print("written file: ", filename)
return counter
def prepare_imagenet(inpath, outpath, batch_size):
r = paddle.batch(imagenet_train(inpath), batch_size=batch_size)
feeder = fluid.DataFeeder(
feed_list=[
fluid.layers.data(
name="image", shape=[3, 224, 224]), fluid.layers.data(
name="label", shape=[1], dtype='int64')
],
place=fluid.CPUPlace())
outpath = os.path.join(outpath, "imagenet.recordio")
convert_reader_to_recordio_files(outpath, 10000, r, feeder)
cmake/configure.cmake
浏览文件 @ bfd42683
......@@ -92,6 +92,9 @@ if(WITH_GPU)
if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
message(FATAL_ERROR "TensorRT needs CUDNN >= 7.0 to compile")
endif()
if(${TENSORRT_MAJOR_VERSION} VERSION_LESS 4)
message(FATAL_ERROR "Paddle needs TensorRT >= 4.0 to compile")
endif()
include_directories(${TENSORRT_INCLUDE_DIR})
endif()
elseif(WITH_AMD_GPU)
......
paddle/fluid/framework/data_layout.h
浏览文件 @ bfd42683
......@@ -27,6 +27,7 @@ enum class DataLayout {
kNHWC = 0,
kNCHW = 1,
kAnyLayout = 2,
kMKLDNN = 3, // all layouts supported by MKLDNN internally
};
inline DataLayout StringToDataLayout(const std::string& str) {
......@@ -41,6 +42,8 @@ inline DataLayout StringToDataLayout(const std::string& str) {
return DataLayout::kNCHW;
} else if (s == "ANYLAYOUT") {
return DataLayout::kAnyLayout;
} else if (s == "MKLDNNLAYOUT") {
return DataLayout::kMKLDNN;
} else {
PADDLE_THROW("Unknown storage order string: %s", s);
}
......@@ -54,8 +57,10 @@ inline std::string DataLayoutToString(const DataLayout& data_layout) {
return "NCHW";
case DataLayout::kAnyLayout:
return "ANY_LAYOUT";
case DataLayout::kMKLDNN:
return "MKLDNNLAYOUT";
default:
PADDLE_THROW("unknown DataLayou %d", data_layout);
PADDLE_THROW("unknown DataLayout %d", data_layout);
}
}
......
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ bfd42683
......@@ -16,6 +16,9 @@
#include <vector>
#include "paddle/fluid/operators/math/math_function.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace framework {
......@@ -88,5 +91,85 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var,
out->set_layout(expected_kernel_type.data_layout_);
}
#ifdef PADDLE_WITH_MKLDNN
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::reorder;
void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) {
switch (type) {
case mkldnn::memory::data_type::f32:
return platform::to_void_cast(tensor.data<float>());
case mkldnn::memory::data_type::s8:
return platform::to_void_cast(tensor.data<char>());
case mkldnn::memory::data_type::u8:
return platform::to_void_cast(tensor.data<unsigned char>());
case mkldnn::memory::data_type::s16:
return platform::to_void_cast(tensor.data<int16_t>());
case mkldnn::memory::data_type::s32:
return platform::to_void_cast(tensor.data<int32_t>());
default:
PADDLE_THROW("wrong mkldnn type provided");
}
}
#endif
void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
const OpKernelType& expected_kernel_type,
const Tensor& in, Tensor* out) {
auto in_layout = kernel_type_for_var.data_layout_;
auto out_layout = expected_kernel_type.data_layout_;
PADDLE_ENFORCE(
in_layout == DataLayout::kMKLDNN && out_layout != DataLayout::kMKLDNN,
"TransDataLayoutFromMKLDNN only supports transform from MKLDNN to "
"non-MKLDNN");
#ifdef PADDLE_WITH_MKLDNN
PADDLE_ENFORCE(in.format() != memory::format::format_undef &&
in.format() != memory::format::any,
"Input tensor should have specified memory format");
// Set default as NCHW in case not specified
out_layout =
out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout;
auto& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx = dynamic_cast<platform::MKLDNNDeviceContext*>(
pool.Get(expected_kernel_type.place_));
auto& cpu_engine = dev_ctx->GetEngine();
std::vector<int> in_tz = paddle::framework::vectorize2int(in.dims());
std::vector<int> out_tz = in_tz;
memory::data_type in_type = ToMKLDNNDataType(in.type());
PADDLE_ENFORCE(in_type != memory::data_type::data_undef,
"Input tensor type is not supported: ", in.type().name());
memory::data_type out_type = in_type;
memory::format in_format =
in_tz.size() == 2 ? memory::format::nc : in.format();
memory::format out_format =
out_tz.size() == 2 ? memory::format::nc : ToMKLDNNFormat(out_layout);
void* in_data = GetDataFromTensor(in, in_type);
// output tensor has the same dims as input. Reorder don't change dims
out->Resize(in.dims());
auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
auto out_memory =
memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
platform::Reorder(in_memory, out_memory);
out->set_layout(out_layout);
// reset format since the out tensor will be feed to non-MKLDNN OPkernel
out->set_format(memory::format::format_undef);
#endif
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/data_layout_transform.h
浏览文件 @ bfd42683
......@@ -14,6 +14,7 @@
#pragma once
#include <map>
#include <vector>
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/tensor.h"
......@@ -22,6 +23,50 @@
namespace paddle {
namespace framework {
#ifdef PADDLE_WITH_MKLDNN
using MKLDNNFormat = mkldnn::memory::format;
using MKLDNNDataType = mkldnn::memory::data_type;
inline MKLDNNFormat ToMKLDNNFormat(const DataLayout& layout) {
switch (layout) {
case DataLayout::kNHWC:
return MKLDNNFormat::nhwc;
case DataLayout::kNCHW:
return MKLDNNFormat::nchw;
default:
PADDLE_THROW("Fail to convert layout %s to MKLDNN format",
DataLayoutToString(layout));
}
}
inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) {
switch (format) {
case MKLDNNFormat::nhwc:
return DataLayout::kNHWC;
case MKLDNNFormat::nchw:
return DataLayout::kNCHW;
default:
PADDLE_THROW("Fail to convert MKLDNN format to paddle layout");
}
}
inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
static const std::map<std::type_index, MKLDNNDataType> dict{
{std::type_index(typeid(float)), MKLDNNDataType::f32}, // NOLINT
{std::type_index(typeid(char)), MKLDNNDataType::s8}, // NOLINT
{std::type_index(typeid(unsigned char)), MKLDNNDataType::u8},
{std::type_index(typeid(int16_t)), MKLDNNDataType::s16},
{std::type_index(typeid(int32_t)), MKLDNNDataType::s32}};
auto iter = dict.find(type);
if (iter != dict.end()) return iter->second;
return MKLDNNDataType::data_undef;
}
#endif
void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
const OpKernelType& expected_kernel_type,
const Tensor& in, Tensor* out);
std::vector<int> GetAxis(const DataLayout& from, const DataLayout& to);
void TransDataLayout(const OpKernelType& kernel_type_for_var,
......
paddle/fluid/framework/data_transform.cc
浏览文件 @ bfd42683
......@@ -33,11 +33,34 @@ void DataTransform(const OpKernelType& expected_kernel_type,
Tensor in;
in.ShareDataWith(input_tensor);
Tensor out;
DataLayout lin = kernel_type_for_var.data_layout_;
DataLayout lout = expected_kernel_type.data_layout_;
// do layout transform
if (NeedTransformLayout(expected_kernel_type.data_layout_,
kernel_type_for_var.data_layout_)) {
if (NeedTransformLayout(lout, lin)) {
if (lin == DataLayout::kMKLDNN || lout == DataLayout::kMKLDNN) {
PADDLE_ENFORCE(
!(lin == DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN),
"No layout transform needed between two MKLDNN OPKernels");
if (lin != DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN) {
#ifdef PADDLE_WITH_MKLDNN
// Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
// Just set layout/format. No real transform occur
out.ShareDataWith(input_tensor);
out.set_layout(DataLayout::kMKLDNN);
out.set_format(ToMKLDNNFormat(lin));
#endif
} else {
// Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel
// Do transform via MKLDNN lib
TransDataLayoutFromMKLDNN(kernel_type_for_var, expected_kernel_type, in,
&out);
}
} else {
// Case3 - transfrom between Non-MKLDNN OPKernels
TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out);
}
transformed = true;
PassTensorData(&out, &in);
}
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ bfd42683
......@@ -13,7 +13,7 @@ cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_ro
if(WITH_GPU)
nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda)
dynload_cuda variable_visitor)
set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
......@@ -25,6 +25,7 @@ else()
endif()
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle)
......
paddle/fluid/framework/details/fuse_vars_op_handle.cc 0 → 100644
浏览文件 @ bfd42683
// 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/fluid/framework/details/fuse_vars_op_handle.h"
namespace paddle {
namespace framework {
namespace details {
void FuseVarsOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(in_var_handles.size(), 0);
PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), "");
auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto out_var_handle = out_var_handles[0];
auto out_var = scope->Var(out_var_handle->name_);
auto out_tensor = out_var->GetMutable<LoDTensor>();
out_tensor->Resize({total_numel_}).mutable_data(this->place_, type_);
int64_t s = 0;
for (size_t i = 1; i < out_var_handles.size(); ++i) {
auto out_name = out_var_handles[i]->name_;
auto out_t = scope->Var(out_name)->GetMutable<LoDTensor>();
auto numel = this->inputs_numel_.at(out_name);
out_t->ShareDataWith(out_tensor->Slice(s, s + numel));
s += numel;
}
this->RunAndRecordEvent([this] {});
}
std::string FuseVarsOpHandle::Name() const { return "fuse vars"; }
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fuse_vars_op_handle.h 0 → 100644
浏览文件 @ bfd42683
// 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.
#pragma once
#include <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace framework {
namespace details {
struct FuseVarsOpHandle : public OpHandleBase {
public:
FuseVarsOpHandle(Scope *local_scope, const platform::Place &place,
const std::unordered_map<std::string, int64_t> &inputs_numel,
const std::type_index &var_type)
: local_scope_(local_scope),
place_(place),
inputs_numel_(inputs_numel),
type_(var_type) {
total_numel_ = 0;
for (auto in_numel : inputs_numel) {
PADDLE_ENFORCE_GT(in_numel.second, 0);
total_numel_ += in_numel.second;
}
}
std::string Name() const override;
bool IsMultiDeviceTransfer() override { return false; };
protected:
void RunImpl() override;
private:
Scope *local_scope_;
const platform::Place place_;
const std::unordered_map<std::string, int64_t> inputs_numel_;
const std::type_index type_;
int64_t total_numel_;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc
浏览文件 @ bfd42683
......@@ -11,10 +11,12 @@
// 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 <algorithm>
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h"
#include <algorithm>
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
namespace paddle {
namespace framework {
......@@ -30,27 +32,34 @@ NCCLAllReduceOpHandle::NCCLAllReduceOpHandle(
}
void NCCLAllReduceOpHandle::RunImpl() {
if (inputs_.size() == 1) {
if (NoDummyInputSize() == 1) {
return; // No need to all reduce when GPU count = 1;
} else {
// Wait input done
WaitInputVarGenerated();
auto &var_name = static_cast<VarHandle *>(this->inputs_[0])->name_;
int dtype = -1;
size_t numel = 0;
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(
in_var_handles.size(), places_.size(),
"The NoDummyInputSize should be equal to the number of places.");
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
std::vector<const LoDTensor *> lod_tensors;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto *s = local_scopes_[i];
auto &local_scope = *s->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto &lod_tensor = local_scope.FindVar(var_name)->Get<LoDTensor>();
auto &lod_tensor =
local_scope.FindVar(in_var_handles[i]->name_)->Get<LoDTensor>();
lod_tensors.emplace_back(&lod_tensor);
PADDLE_ENFORCE_EQ(in_var_handles[i]->name_, out_var_handles[i]->name_,
"The name of input and output should be equal.");
}
if (platform::is_gpu_place(lod_tensors[0]->place())) {
int dtype = -1;
size_t numel = 0;
std::vector<std::function<void()>> all_reduce_calls;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &p = places_[i];
......@@ -96,7 +105,7 @@ void NCCLAllReduceOpHandle::RunImpl() {
auto &scope =
*local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto &p = places_[i];
auto *var = scope.FindVar(var_name);
auto *var = scope.FindVar(in_var_handles[i]->name_);
auto *dev_ctx = dev_ctxes_[p];
RunAndRecordEvent(p, [&trg, var, dev_ctx, p] {
......
paddle/fluid/framework/details/op_handle_base.cc
浏览文件 @ bfd42683
......@@ -104,6 +104,16 @@ void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
}
}
size_t OpHandleBase::NoDummyInputSize() const {
size_t cnt = 0;
for (auto *in : inputs_) {
if (dynamic_cast<DummyVarHandle *>(in) == nullptr) {
++cnt;
}
}
return cnt;
}
bool OpHandleBase::NeedWait(VarHandleBase *in_var) {
return in_var && in_var->generated_op_;
}
......
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @ bfd42683
......@@ -80,6 +80,8 @@ class OpHandleBase {
const std::vector<VarHandleBase *> &Outputs() const { return outputs_; }
size_t NoDummyInputSize() const;
protected:
void RunAndRecordEvent(const std::function<void()> &callback);
......
paddle/fluid/framework/details/ssa_graph_printer.h
浏览文件 @ bfd42683
......@@ -20,7 +20,7 @@
namespace paddle {
namespace framework {
namespace details {
class SSAGraph;
struct SSAGraph;
class SSAGraphPrinter {
public:
virtual ~SSAGraphPrinter() {}
......
paddle/fluid/framework/op_kernel_type.h
浏览文件 @ bfd42683
......@@ -87,7 +87,14 @@ inline std::string KernelTypeToString(const OpKernelType& kernel_key) {
}
inline bool NeedTransformLayout(const DataLayout& l, const DataLayout& r) {
return l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r;
bool ret =
(l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r);
#ifdef PADDLE_WITH_MKLDNN
// Layout transform needed for either non-MKLDNN to MKLDNN or vice versa
ret |= (l != DataLayout::kMKLDNN && r == DataLayout::kMKLDNN);
ret |= (l == DataLayout::kMKLDNN && r != DataLayout::kMKLDNN);
#endif
return ret;
}
inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) {
......
paddle/fluid/framework/op_registry.h
浏览文件 @ bfd42683
......@@ -83,8 +83,14 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
void operator()(const char* op_type, const char* library_type) const {
using T = typename KERNEL_TYPE::ELEMENT_TYPE;
std::string library(library_type);
std::string data_layout = "ANYLAYOUT";
if (library == "MKLDNN") {
data_layout = "MKLDNNLAYOUT";
}
OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
DataLayout::kAnyLayout, StringToLibraryType(library_type));
StringToDataLayout(data_layout),
StringToLibraryType(library_type));
OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
......@@ -99,7 +105,8 @@ struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
void operator()(const char* op_type, const char* library_type) const {}
};
// User can register many kernel in one place. The data type could be different.
// User can register many kernel in one place. The data type could be
// different.
template <typename PlaceType, typename... KernelType>
class OpKernelRegistrar : public Registrar {
public:
......
paddle/fluid/framework/operator.cc
浏览文件 @ bfd42683
......@@ -444,9 +444,24 @@ class RuntimeInferShapeContext : public InferShapeContext {
auto* out_tensor = out_var->GetMutable<LoDTensor>();
out_tensor->set_lod(in_tensor.lod());
// TODO(dzhwinter) : reuse ShareLoD in most operators.
// Need to call ShareLayout explicitly in sequence related ops.
// Shall we have a better method to shared info between in/out Tensor?
// TODO(dzhwinter) : reuse ShareLoD in most operators.
// Need to call ShareLayout explicitly in sequence related ops.
// Shall we have a better method to shared info between in/out Tensor?
#ifdef PADDLE_WITH_MKLDNN
// Fix me: ugly workaround below
// Correct solution:
// set_layout() should NOT be called here (i.e. ShareLoD). Instead,
// layout of output tensor should be set "manually" in Compute()
// of each OPKernel. The reason layout should NOT be shared between
// input and output "automatically" (now by InferShape()->ShareLoD())
// is that layout transform may occur after InferShape().
// Workaround:
// Skip set_layout() when input layout is kMKLDNN
// This is to avoid kMKLDNN is populated wrongly into a non-MKLDNN
// OPKernel. In all MKLDNN OPkernel, set_layout(kMKLDNN) should be called
// in Compute()
if (in_tensor.layout() != DataLayout::kMKLDNN)
#endif
out_tensor->set_layout(in_tensor.layout());
}
......@@ -646,8 +661,10 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
}
if (t != nullptr) {
int tmp = static_cast<int>(ToDataType(t->type()));
PADDLE_ENFORCE(tmp == data_type || data_type == -1,
"DataType of Paddle Op %s must be the same.", Type());
PADDLE_ENFORCE(
tmp == data_type || data_type == -1,
"DataType of Paddle Op %s must be the same. Get %d != %d", Type(),
data_type, tmp);
data_type = tmp;
}
}
......@@ -665,7 +682,8 @@ OpKernelType OperatorWithKernel::GetExpectedKernelType(
OpKernelType OperatorWithKernel::GetKernelTypeForVar(
const std::string& var_name, const Tensor& tensor,
const OpKernelType& expected_kernel_type) const {
return OpKernelType(expected_kernel_type.data_type_, tensor.place());
return OpKernelType(expected_kernel_type.data_type_, tensor.place(),
tensor.layout());
}
} // namespace framework
......
paddle/fluid/framework/tensor.h
浏览文件 @ bfd42683
......@@ -34,6 +34,28 @@ namespace framework {
class LoDTensor;
class Tensor {
#ifdef PADDLE_WITH_MKLDNN
public:
inline mkldnn::memory::format format() const { return format_; }
inline void set_format(const mkldnn::memory::format format) {
format_ = format;
}
protected:
/**
* @brief the detail format of memory block which have layout as kMKLDNN
*
* @note MKLDNN lib support various memory format like nchw, nhwc, nChw8C,
* nChw16c, etc. For a MKLDNN memory block, layout will be set as
* DataLayout::kMKLDNN meanwhile detail memory format will be kept in
* this field.
*/
mkldnn::memory::format format_ = mkldnn::memory::format::format_undef;
#endif
public:
template <typename T, size_t D, int MajorType, typename IndexType>
friend struct EigenTensor;
......@@ -195,8 +217,10 @@ class Tensor {
* N,C,H,W for respectively the batch size, the number of
* feature maps, the height.
*/
DataLayout layout_ = DataLayout::kNHWC;
// Fix me: here just change the default layout to kNCHW
// it doesn't fix the real issue, i.e. feeder should set up tensor layout
// according to actual input data
DataLayout layout_ = DataLayout::kNCHW;
/**
* @brief A PlaceHolder may be shared by more than one tensor.
......
paddle/fluid/framework/tensor_test.cc
浏览文件 @ bfd42683
......@@ -209,7 +209,7 @@ TEST(Tensor, ReshapeToMatrix) {
TEST(Tensor, Layout) {
framework::Tensor src;
ASSERT_EQ(src.layout(), framework::DataLayout::kNHWC);
ASSERT_EQ(src.layout(), framework::DataLayout::kNCHW);
src.set_layout(framework::DataLayout::kAnyLayout);
ASSERT_EQ(src.layout(), framework::DataLayout::kAnyLayout);
}
paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
浏览文件 @ bfd42683
# Add TRT tests
# This test is not stable
# See https://paddleci.ngrok.io/viewLog.html?tab=buildLog&buildTypeId=Paddle_PrCi2&buildId=36834&_focus=8828
#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
# DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine
# SERIAL)
nv_library(tensorrt_converter
SRCS mul_op.cc conv2d_op.cc fc_op.cc
DEPS tensorrt_engine mul_op)
......@@ -16,3 +11,5 @@ nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL)
nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
DEPS ${FLUID_CORE_MODULES} tensorrt_engine activation_op SERIAL)
paddle/fluid/inference/tensorrt/convert/activation_op.cc
浏览文件 @ bfd42683
......@@ -22,7 +22,8 @@ namespace tensorrt {
class ReluOpConverter : public OpConverter {
public:
ReluOpConverter() {}
void operator()(const framework::proto::OpDesc& op) override {
void operator()(const framework::proto::OpDesc& op,
const framework::Scope& scope, bool test_mode) override {
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework::OpDesc op_desc(op, nullptr);
......@@ -33,7 +34,12 @@ class ReluOpConverter : public OpConverter {
nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER(
engine_, Activation, *const_cast<nvinfer1::ITensor*>(input_tensor),
nvinfer1::ActivationType::kRELU);
engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0));
auto output_name = op_desc.Output("Out")[0];
engine_->SetITensor(output_name, layer->getOutput(0));
if (test_mode) { // the test framework can not determine which is the
// output, so place the declaration inside.
engine_->DeclareOutput(output_name);
}
}
};
......
paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
浏览文件 @ bfd42683
/* 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
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. */
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 <gtest/gtest.h>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/place.h"
USE_OP(relu);
#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
namespace paddle {
namespace inference {
namespace tensorrt {
void Compare(const std::string op_type, float input, float expect) {
TEST(ReluOpConverter, main) {
framework::Scope scope;
platform::CUDAPlace place;
platform::CUDADeviceContext ctx(place);
// init fluid op and variable
auto x_var = scope.Var("X");
auto x_tensor = x_var->GetMutable<framework::LoDTensor>();
x_tensor->Resize({1, 1});
x_tensor->mutable_data<float>(place);
std::vector<float> init;
init.push_back(input);
framework::TensorFromVector(init, ctx, x_tensor);
auto out_var = scope.Var("Out");
auto out_tensor = out_var->GetMutable<framework::LoDTensor>();
out_tensor->Resize({1, 1});
out_tensor->mutable_data<float>(place);
framework::OpDesc op_desc;
op_desc.SetType(op_type);
op_desc.SetInput("X", {"X"});
op_desc.SetOutput("Out", {"Out"});
auto op = framework::OpRegistry::CreateOp(*op_desc.Proto());
// run fluid op
op->Run(scope, place);
// get fluid output
std::vector<float> out1;
framework::TensorToVector(*out_tensor, ctx, &out1);
// init tensorrt op
cudaStream_t stream;
ASSERT_EQ(0, cudaStreamCreate(&stream));
TensorRTEngine* engine = new TensorRTEngine(1, 1 << 10, &stream);
engine->InitNetwork();
engine->DeclareInput("X", nvinfer1::DataType::kFLOAT,
nvinfer1::DimsCHW{1, 1, 1});
// convert op
OpConverter op_converter;
op_converter.ConvertOp(*op_desc.Proto(), engine);
engine->DeclareOutput("Out");
engine->FreezeNetwork();
// convert LoDTensor to ITensor
size_t size = x_tensor->memory_size();
EngineIOConverter::ConvertInput(op_type, *x_tensor,
engine->buffer("X").buffer, size, &stream);
// run tensorrt Outp
engine->Execute(1);
// convert ITensor to LoDTensor
EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer,
out_tensor, size, &stream);
// get tensorrt output
std::vector<float> out2;
framework::TensorToVector(*out_tensor, ctx, &out2);
// compare
ASSERT_EQ(out1[0], out2[0]);
ASSERT_EQ(out1[0], expect);
delete engine;
cudaStreamDestroy(stream);
}
TEST(OpConverter, ConvertRelu) {
Compare("relu", 1, 1); // relu(1) = 1
Compare("relu", -5, 0); // relu(-5) = 0
std::unordered_set<std::string> parameters;
TRTConvertValidation validator(10, parameters, scope, 1000);
validator.DeclInputVar("relu-X", nvinfer1::Dims2(10, 6));
validator.DeclOutputVar("relu-Out", nvinfer1::Dims2(10, 6));
// Prepare Op description
framework::OpDesc desc;
desc.SetType("relu");
desc.SetInput("X", {"relu-X"});
desc.SetOutput("Out", {"relu-Out"});
LOG(INFO) << "set OP";
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(10);
}
} // namespace tensorrt
} // namespace inference
} // namespace paddle
USE_OP(activation);
USE_OP(relu);
paddle/fluid/operators/activation_op.cc
浏览文件 @ bfd42683
......@@ -58,14 +58,16 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
const framework::OperatorWithKernel& oper,
const std::string& name) {
framework::LibraryType library{framework::LibraryType::kPlain};
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN
auto it = oper.Attrs().find("use_mkldnn");
if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() &&
platform::CanMKLDNNBeUsed(ctx)) {
library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
return framework::OpKernelType(
framework::ToDataType(ctx.Input<framework::Tensor>(name)->type()),
ctx.GetPlace(), layout, library);
......
paddle/fluid/operators/batch_norm_op.cc
浏览文件 @ bfd42683
......@@ -111,14 +111,16 @@ class BatchNormOp : public framework::OperatorWithKernel {
"Variance input should be of float type");
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
library_);
}
......@@ -367,17 +369,18 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
}
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout, library_);
layout_, library_);
}
};
......
paddle/fluid/operators/conv_op.cc
浏览文件 @ bfd42683
......@@ -75,6 +75,11 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
framework::LibraryType library{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library = framework::LibraryType::kCUDNN;
......@@ -84,6 +89,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
......@@ -99,9 +105,6 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
"float16 can only be used when CUDNN is used");
}
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout = framework::StringToDataLayout(data_format);
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
library);
}
......@@ -309,6 +312,10 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
......@@ -318,12 +325,10 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
layout_, library_);
......
paddle/fluid/operators/fc_op.cc
浏览文件 @ bfd42683
......@@ -43,7 +43,7 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType FCOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
framework::LibraryType library{framework::LibraryType::kMKLDNN};
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
framework::DataLayout layout{framework::DataLayout::kMKLDNN};
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
......@@ -65,7 +65,7 @@ void FCOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType FCOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
framework::LibraryType library{framework::LibraryType::kMKLDNN};
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
framework::DataLayout layout{framework::DataLayout::kMKLDNN};
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
......
paddle/fluid/operators/lrn_op.cc
浏览文件 @ bfd42683
......@@ -124,16 +124,17 @@ namespace {
framework::OpKernelType GetExpectedLRNKernel(
const framework::ExecutionContext& ctx) {
framework::LibraryType library_{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout_, library_);
......
paddle/fluid/operators/pool_mkldnn_op.cc
浏览文件 @ bfd42683
......@@ -24,10 +24,13 @@ using mkldnn::pooling_backward;
// Generate keys for storing/retriving primitives for this operator
// TODO(jczaja): Make hashing function more optimial
static std::string gethash(memory::dims& input_dims, std::string& pooling_type,
std::vector<int>& ksize, std::vector<int>& strides,
std::vector<int>& paddings, std::string suffix) {
auto dims2str = [](memory::dims& operand_dims) {
static std::string gethash(const memory::dims& input_dims,
const std::string& pooling_type,
const std::vector<int>& ksize,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const std::string& suffix) {
auto dims2str = [](const memory::dims& operand_dims) {
std::string dstr = "";
for (size_t i = 0; i < operand_dims.size(); ++i) {
dstr += std::to_string(operand_dims[i]) + "-";
......
paddle/fluid/operators/pool_op.cc
浏览文件 @ bfd42683
......@@ -83,6 +83,9 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
framework::OpKernelType PoolOp::GetExpectedKernelType(
const framework::ExecutionContext &ctx) const {
framework::LibraryType library_{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
......@@ -92,11 +95,10 @@ framework::OpKernelType PoolOp::GetExpectedKernelType(
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout_, library_);
......@@ -112,6 +114,9 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const {
framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
const framework::ExecutionContext &ctx) const {
framework::LibraryType library_{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
......@@ -121,6 +126,7 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
......@@ -129,8 +135,6 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN,
"float16 can only be used when CUDNN is used");
}
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_,
library_);
}
......
paddle/fluid/operators/softmax_op.cc
浏览文件 @ bfd42683
......@@ -49,6 +49,9 @@ class SoftmaxOp : public framework::OperatorWithKernel {
const framework::ExecutionContext& ctx) const override {
// choose cudnn kernel if the runtime supported.
framework::LibraryType library_{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
......@@ -58,6 +61,7 @@ class SoftmaxOp : public framework::OperatorWithKernel {
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
}
#endif
......@@ -68,9 +72,7 @@ class SoftmaxOp : public framework::OperatorWithKernel {
"float16 can only be used on GPU place");
}
std::string data_format = ctx.Attr<std::string>("data_format");
return framework::OpKernelType(input_data_type, ctx.GetPlace(),
framework::StringToDataLayout(data_format),
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_,
library_);
}
};
......@@ -142,6 +144,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
const framework::ExecutionContext& ctx) const override {
// choose cudnn kernel if the runtime supported.
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
......
paddle/fluid/platform/mkldnn_helper.h
浏览文件 @ bfd42683
......@@ -16,6 +16,7 @@ limitations under the License. */
#include <mkldnn.h>
#include <vector>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace platform {
......@@ -86,5 +87,17 @@ inline mkldnn::memory::data_type MKLDNNGetDataType<float>() {
return mkldnn::memory::f32;
}
inline void Reorder(const mkldnn::memory& src, const mkldnn::memory& dst) {
auto reorder_prim = mkldnn::reorder(src, dst);
std::vector<mkldnn::primitive> pipeline;
pipeline.push_back(reorder_prim);
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}
inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) {
return static_cast<mkldnn::memory::format>(
memory.get_primitive_desc().desc().data.format);
}
} // namespace platform
} // namespace paddle
python/paddle/fluid/executor.py
浏览文件 @ bfd42683
......@@ -170,6 +170,8 @@ def get_program_cache_key(feed, fetch_list):
return var.desc.name()
elif isinstance(var, str):
return var
elif isinstance(var, basestring):
return str(var)
else:
raise TypeError(str(var) + " should be Variable or str")
......
python/paddle/fluid/framework.py
浏览文件 @ bfd42683
......@@ -72,6 +72,8 @@ def convert_np_dtype_to_dtype_(np_dtype):
return core.VarDesc.VarType.INT64
elif dtype == np.bool:
return core.VarDesc.VarType.BOOL
elif dtype == np.uint16:
return core.VarDesc.VarType.INT16
elif dtype == np.uint8:
return core.VarDesc.VarType.UINT8
else:
......@@ -368,6 +370,13 @@ class Operator(object):
Block. Users can use the build in instructions to describe their neural
network.
"""
OP_WITHOUT_KERNEL_SET = {
'feed', 'fetch', 'save', 'load', 'recurrent', 'go',
'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv',
'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine',
'ncclInit', 'channel_create', 'channel_close', 'channel_send',
'channel_recv', 'select'
}
def __init__(self,
block,
......@@ -504,17 +513,13 @@ class Operator(object):
else:
self.desc.set_attr(attr_name, self.attrs[attr_name])
self.desc.check_attrs()
no_kernel_op_set = {
'feed', 'fetch', 'save', 'load', 'recurrent', 'go',
'rnn_memory_helper_grad', 'conditional_block', 'while', 'send',
'recv', 'listen_and_serv', 'parallel_do', 'save_combine',
'load_combine', 'ncclInit', 'channel_create', 'channel_close',
'channel_send', 'channel_recv', 'select', 'gen_nccl_id'
}
if type not in no_kernel_op_set:
if self.has_kernel(type):
self.desc.infer_var_type(self.block.desc)
self.desc.infer_shape(self.block.desc)
def has_kernel(self, op_type):
return op_type not in self.OP_WITHOUT_KERNEL_SET
def to_string(self, throw_on_error):
"""
To debug string.
......@@ -742,7 +747,9 @@ class Block(object):
def var(self, name):
if not isinstance(name, basestring):
raise TypeError()
raise TypeError(
"var require string as parameter, but get %s instead." %
(type(name)))
v = self.vars.get(name, None)
if v is None:
raise ValueError("var %s not in this block" % name)
......
python/paddle/fluid/layers/io.py
浏览文件 @ bfd42683
......@@ -434,7 +434,7 @@ def open_files(filenames,
shapes,
lod_levels,
dtypes,
thread_num,
thread_num=1,
buffer_size=None,
pass_num=1,
for_parallel=True):
......
python/paddle/fluid/tests/unittests/benchmark.py 0 → 100644
浏览文件 @ bfd42683
# 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 numpy as np
import unittest
import time
import itertools
import paddle.fluid as fluid
import paddle.fluid.core as core
from paddle.fluid.op import Operator
from op_test import OpTest
class BenchmarkSuite(OpTest):
def timeit_function(self, callback, iters, *args, **kwargs):
assert iters != 0, "Iters should >= 1"
start = time.time()
for i in range(iters):
callback(*args, **kwargs)
elapse = time.time() - start
return elapse / iters
def _assert_cpu_gpu_same(self, cpu_outs, gpu_outs, fetch_list, atol):
for item_cpu_out, item_gpu_out, variable in zip(cpu_outs, gpu_outs,
fetch_list):
# the cpu version is baseline, expect gpu version keep same with cpu version.
expect = item_cpu_out
expect_t = np.array(item_cpu_out)
actual = item_gpu_out
actual_t = np.array(item_gpu_out)
var_name = variable if isinstance(variable,
basestring) else variable.name
self.assertTrue(
np.allclose(
actual_t, expect_t, atol=atol),
"Output (" + var_name + ") has diff" + str(actual_t) + "\n" +
str(expect_t))
self.assertListEqual(actual.lod(),
expect.lod(),
"Output (" + var_name + ") has different lod")
def _get_input_names(self):
inputs = []
for name, value in self.inputs.iteritems():
if isinstance(value, list):
inputs.extend([sub_name for sub_name, _ in value])
inputs.append(name)
return inputs
def _get_output_names(self):
outputs = []
for var_name, var in self.outputs.iteritems():
if isinstance(var, list):
for sub_var_name, sub_var in var:
outputs.append(sub_var_name)
else:
outputs.append(var_name)
if len(outputs) == 0:
for out_name, out_dup in Operator.get_op_outputs(self.op_type):
outputs.append(str(out_name))
return outputs
def check_output_stability(self, atol=1e-8):
places = self._get_places()
if len(places) < 2:
return
cpu_outs, fetch_list = self._calc_output(places[0])
gpu_outs, _ = self._calc_output(places[1])
self._assert_cpu_gpu_same(cpu_outs, gpu_outs, fetch_list, atol)
def timeit_output_with_place(self, place, iters):
return self.timeit_function(self.calc_output, iters, place)
def timeit_output(self, iters=100):
places = self._get_places()
elapses = []
for place in places:
elapses.append(self.timeit_output_with_place(place, iters))
for place, elapse in zip(places, elapses):
print("One pass of ({2}_op) at {0} cost {1}".format(
str(place), elapse, self.op_type))
def timeit_grad_with_place(self, place, iters=100):
inputs_to_check = self._get_input_names()
output_names = self._get_output_names()
return self.timeit_function(
self._get_gradient,
iters,
inputs_to_check,
place,
output_names,
no_grad_set=None)
def timeit_grad(self, iters=100):
places = self._get_places()
elapses = []
for place in places:
elapses.append(self.timeit_grad_with_place(place, iters))
for place, elapse in zip(places, elapses):
print("One pass of ({2}_grad_op) at {0} cost {1}".format(
str(place), elapse, self.op_type))
python/paddle/fluid/tests/unittests/benchmark_sum_op.py 0 → 100644
浏览文件 @ bfd42683
# 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
import paddle.fluid as fluid
from benchmark import BenchmarkSuite
from op_test import OpTest
# This is a demo op test case for operator benchmarking and high resolution number stability alignment.
class TestSumOp(BenchmarkSuite):
def setUp(self):
self.op_type = "sum"
self.customize_testcase()
self.customize_fetch_list()
def customize_fetch_list(self):
"""
customize fetch list, configure the wanted variables.
>>> self.fetch_list = ["Out"]
"""
self.fetch_list = ["Out"]
# pass
def customize_testcase(self):
# a test case
x0 = np.random.random((300, 400)).astype('float32')
x1 = np.random.random((300, 400)).astype('float32')
x2 = np.random.random((300, 400)).astype('float32')
# NOTE: if the output is empty, then it will autofilled by benchmarkSuite.
# only the output dtype is used, the shape, lod and data is computed from input.
self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]}
self.outputs = {"Out": x0 + x1 + x2}
def test_check_output(self):
"""
compare the output with customized output. In this case,
you should set the correct output by hands.
>>> self.outputs = {"Out": x0 + x1 + x2}
"""
self.check_output(atol=1e-8)
def test_output_stability(self):
# compare the cpu gpu output in high resolution.
self.check_output_stability()
def test_timeit_output(self):
"""
perf the op, time cost will be averged in iters.
output example
>>> One pass of (sum_op) at CPUPlace cost 0.000461330413818
>>> One pass of (sum_op) at CUDAPlace(0) cost 0.000556070804596
"""
self.timeit_output(iters=100)
def test_timeit_grad(self):
"""
perf the op gradient, time cost will be averged in iters.
output example
>>> One pass of (sum_grad_op) at CPUPlace cost 0.00279935121536
>>> One pass of (sum_grad_op) at CUDAPlace(0) cost 0.00500632047653
"""
self.timeit_grad(iters=100)
if __name__ == "__main__":
unittest.main()
python/paddle/fluid/tests/unittests/op_test.py
浏览文件 @ bfd42683
......@@ -15,13 +15,17 @@
import unittest
import numpy as np
import random
import time
import itertools
import paddle.fluid.core as core
import collections
import paddle.fluid as fluid
import paddle.fluid.core as core
from paddle.fluid.backward import append_backward
from paddle.fluid.op import Operator
from paddle.fluid.executor import Executor
from paddle.fluid.framework import Program, OpProtoHolder
from paddle.fluid.framework import Program, OpProtoHolder, Variable
from testsuite import create_op, set_input, append_input_output, append_loss_ops
def randomize_probability(batch_size, class_num, dtype='float32'):
......@@ -33,73 +37,6 @@ def randomize_probability(batch_size, class_num, dtype='float32'):
return prob
def create_op(scope, op_type, inputs, outputs, attrs):
kwargs = dict()
op_maker = core.op_proto_and_checker_maker
op_role_attr_name = op_maker.kOpRoleAttrName()
if op_role_attr_name not in attrs:
attrs[op_role_attr_name] = int(op_maker.OpRole.Forward)
def __create_var__(name, var_name):
scope.var(var_name).get_tensor()
kwargs[name].append(var_name)
for in_name, in_dup in Operator.get_op_inputs(op_type):
if in_name in inputs:
kwargs[in_name] = []
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, _ = item[0], item[1]
__create_var__(in_name, sub_in_name)
else:
__create_var__(in_name, in_name)
for out_name, out_dup in Operator.get_op_outputs(op_type):
if out_name in outputs:
kwargs[out_name] = []
if out_dup:
sub_out = outputs[out_name]
for item in sub_out:
sub_out_name, _ = item[0], item[1]
__create_var__(out_name, sub_out_name)
else:
__create_var__(out_name, out_name)
for attr_name in Operator.get_op_attr_names(op_type):
if attr_name in attrs:
kwargs[attr_name] = attrs[attr_name]
return Operator(op_type, **kwargs)
def set_input(scope, op, inputs, place):
def __set_input__(var_name, var):
if isinstance(var, tuple) or isinstance(var, np.ndarray):
tensor = scope.find_var(var_name).get_tensor()
if isinstance(var, tuple):
tensor.set_lod(var[1])
var = var[0]
tensor.set_dims(var.shape)
tensor.set(var, place)
elif isinstance(var, float):
scope.find_var(var_name).set_float(var)
elif isinstance(var, int):
scope.find_var(var_name).set_int(var)
for in_name, in_dup in Operator.get_op_inputs(op.type()):
if in_name in inputs:
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, sub_in_val = item[0], item[1]
__set_input__(sub_in_name, sub_in_val)
else:
__set_input__(in_name, inputs[in_name])
def get_numeric_gradient(place,
scope,
op,
......@@ -173,54 +110,15 @@ def get_numeric_gradient(place,
return gradient_flat.reshape(tensor_to_check.get_dims())
def append_input_output(block, op_proto, np_list, is_input):
'''Insert VarDesc and generate Python variable instance'''
proto_list = op_proto.inputs if is_input else op_proto.outputs
def create_var(block, name, np_list, var_proto):
if name not in np_list:
assert var_proto.intermediate, "{} not found".format(name)
shape = None
lod_level = None
else:
np_value = np_list[name]
if isinstance(np_value, tuple):
shape = list(np_value[0].shape)
lod_level = len(np_value[1])
else:
shape = list(np_value.shape)
lod_level = 0
return block.create_var(
dtype="float32", shape=shape, lod_level=lod_level, name=name)
var_dict = {}
for var_proto in proto_list:
var_name = str(var_proto.name)
if is_input:
if (var_name not in np_list) and var_proto.dispensable:
continue
assert (var_name in np_list) or (var_proto.dispensable), \
"Missing {} as input".format(var_name)
if var_proto.duplicable:
assert isinstance(np_list[var_name], list), \
"Duplicable {} should be set as list".format(var_name)
var_list = []
for (name, np_value) in np_list[var_name]:
var_list.append(
create_var(block, name, {name: np_value}, var_proto))
var_dict[var_name] = var_list
else:
var_dict[var_name] = create_var(block, var_name, np_list, var_proto)
return var_dict
class OpTest(unittest.TestCase):
@classmethod
def setUpClass(cls):
'''Fix random seeds to remove randomness from tests'''
cls._np_rand_state = np.random.get_state()
cls._py_rand_state = random.getstate()
cls.call_once = False
cls.dtype = "float32"
cls.outputs = {}
np.random.seed(123)
random.seed(124)
......@@ -231,6 +129,31 @@ class OpTest(unittest.TestCase):
np.random.set_state(cls._np_rand_state)
random.setstate(cls._py_rand_state)
def try_call_once(self, data_type):
if not self.call_once:
self.call_once = True
self.dtype = data_type
def infer_dtype_from_inputs_outputs(self, inputs, outputs):
def infer_dtype(numpy_dict):
assert isinstance(
numpy_dict,
dict), "self.inputs, self.outputs must be numpy_dict"
for var_name, var_value in numpy_dict.iteritems():
if isinstance(var_value, (np.ndarray, np.generic)):
self.try_call_once(var_value.dtype)
elif isinstance(var_value, (list, tuple)):
# the case of self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]}
if len(var_value) > 1 and isinstance(var_value[1], (
np.ndarray, np.generic)):
instance = var_value[1]
self.try_call_once(instance[1].dtype)
else:
self.try_call_once("float32")
infer_dtype(inputs)
infer_dtype(outputs)
def feed_var(self, input_vars, place):
feed_map = {}
for var_name in input_vars:
......@@ -254,18 +177,14 @@ class OpTest(unittest.TestCase):
return feed_map
def calc_output(self, place):
outs, _ = self._calc_output(place)
return outs
def _calc_output(self, place):
def _append_ops(self, block):
op_proto = OpProtoHolder.instance().get_op_proto(self.op_type)
program = Program()
block = program.global_block()
inputs = append_input_output(block, op_proto, self.inputs, True)
outputs = append_input_output(block, op_proto, self.outputs, False)
"infer datatype from inputs and outputs for this test case"
self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs)
inputs = append_input_output(block, op_proto, self.inputs, True,
self.dtype)
outputs = append_input_output(block, op_proto, self.outputs, False,
self.dtype)
op = block.append_op(
type=self.op_type,
inputs=inputs,
......@@ -275,19 +194,65 @@ class OpTest(unittest.TestCase):
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
fetch_list = []
def _get_io_vars(self, block, numpy_inputs):
inputs = {}
for name, value in numpy_inputs.iteritems():
if isinstance(value, list):
var_list = [
block.var(sub_name) for sub_name, sub_value in value
]
inputs[name] = var_list
else:
inputs[name] = block.var(name)
return inputs
def _get_inputs(self, block):
return self._get_io_vars(block, self.inputs)
def _get_outputs(self, block):
return self._get_io_vars(block, self.outputs)
def calc_output(self, place):
outs, _ = self._calc_output(place)
return outs
def _calc_output(self, place, parallel=False):
program = Program()
block = program.global_block()
self._append_ops(block)
inputs = self._get_inputs(block)
outputs = self._get_outputs(block)
feed_map = self.feed_var(inputs, place)
if parallel:
use_cuda = False
if isinstance(place, fluid.CUDAPlace(0)):
use_cuda = True
executor = fluid.ParallelExecutor(
use_cuda=use_cuda, loss_name=loss.name, main_program=program)
else:
executor = Executor(place)
fetch_list = getattr(self, "fetch_list", [])
# if the fetch_list is customized by user, we use it directly.
# if not, fill the fetch_list by the user configured outputs in test.
if len(fetch_list) == 0:
for var_name, var in outputs.iteritems():
if var_name in self.outputs:
if isinstance(var, list):
for v in var:
fetch_list.append(v)
else:
fetch_list.append(var)
feed_map = self.feed_var(inputs, place)
exe = Executor(place)
outs = exe.run(program,
# if the fetch_list still empty, fill the fetch_list by the operator output.
if len(fetch_list) == 0:
for out_name, out_dup in Operator.get_op_outputs(self.op_type):
fetch_list.append(str(out_name))
# fetch_list = map(block.var, fetch_list)
if not isinstance(fetch_list[0], Variable):
fetch_list = map(block.var, fetch_list)
outs = executor.run(program,
feed=feed_map,
fetch_list=fetch_list,
return_numpy=False)
......@@ -346,17 +311,19 @@ class OpTest(unittest.TestCase):
"Output (" + out_name +
") has different lod at " + str(place))
def check_output(self, atol=1e-5):
places = [core.CPUPlace()]
def _get_places(self):
places = [fluid.CPUPlace()]
if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type):
places.append(core.CUDAPlace(0))
return places
def check_output(self, atol=1e-5):
places = self._get_places()
for place in places:
self.check_output_with_place(place, atol)
def check_output_customized(self, checker):
places = [core.CPUPlace()]
if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type):
places.append(core.CUDAPlace(0))
places = self._get_places()
for place in places:
outs = self.calc_output(place)
outs = [np.array(out) for out in outs]
......@@ -389,9 +356,7 @@ class OpTest(unittest.TestCase):
in_place=False,
max_relative_error=0.005,
user_defined_grads=None):
places = [core.CPUPlace()]
if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type):
places.append(core.CUDAPlace(0))
places = self._get_places()
for place in places:
self.check_grad_with_place(place, inputs_to_check, output_names,
no_grad_set, numeric_grad_delta,
......@@ -438,35 +403,6 @@ class OpTest(unittest.TestCase):
max_relative_error,
"Gradient Check On %s" % str(place))
@staticmethod
def _create_var_descs_(block, var_dict):
# FIXME: Try unify with `append_input_output`
for param_name in var_dict:
var = var_dict[param_name]
if not isinstance(var, list) and not isinstance(var, tuple):
var = [(param_name, var, None)]
if not isinstance(var[0], list) and not isinstance(var[0], tuple):
var = [(param_name, var[0], var[1])]
for i, item in enumerate(var):
if not isinstance(item[0], basestring):
item = [[param_name] + list(item)]
if len(item) == 2:
if isinstance(item[1], tuple):
var[i] = [item[0], item[1][0], item[1][1]]
else:
# only set var name and value, set lod to None
var[i] = list(item) + [None]
var_descs = [(block.create_var(
name=name, shape=each.shape, dtype=each.dtype), each, lod)
for name, each, lod in var]
yield param_name, var_descs
@staticmethod
def _merge_list(iterable):
return reduce(lambda a, b: list(a) + list(b), iterable, [])
@staticmethod
def _numpy_to_lod_tensor(np_value, lod, place):
tensor = core.LoDTensor()
......@@ -497,82 +433,30 @@ class OpTest(unittest.TestCase):
input.dtype = np.uint16
return input
def _get_gradient(self, input_to_check, place, output_names, no_grad_set):
def _get_gradient(self,
input_to_check,
place,
output_names,
no_grad_set,
parallel=False):
prog = Program()
block = prog.global_block()
inputs_with_np = {
key: value
for (key, value) in OpTest._create_var_descs_(
block, getattr(self, 'inputs', {}))
}
outputs_with_np = {
key: val
for (key, val) in OpTest._create_var_descs_(
block, getattr(self, 'outputs', {}))
}
inputs = {
k: [item[0] for item in inputs_with_np[k]]
for k in inputs_with_np
}
outputs = {
k: [item[0] for item in outputs_with_np[k]]
for k in outputs_with_np
}
op = block.append_op(
type=self.op_type,
inputs=inputs,
outputs=outputs,
attrs=getattr(self, 'attrs', {}))
# infer variable type and infer shape in compile-time
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
mean_inputs = map(block.var, output_names)
if len(mean_inputs) == 1:
loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1])
op = block.append_op(
inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean')
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
else:
avg_sum = []
for cur_loss in mean_inputs:
cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1])
op = block.append_op(
inputs={"X": [cur_loss]},
outputs={"Out": [cur_avg_loss]},
type="mean")
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
avg_sum.append(cur_avg_loss)
loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1])
op_sum = block.append_op(
inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum')
op_sum.desc.infer_var_type(block.desc)
op_sum.desc.infer_shape(block.desc)
loss = block.create_var(dtype=loss_sum.dtype, shape=[1])
op_loss = block.append_op(
inputs={"X": loss_sum},
outputs={"Out": loss},
type='scale',
attrs={'scale': 1.0 / float(len(avg_sum))})
op_loss.desc.infer_var_type(block.desc)
op_loss.desc.infer_shape(block.desc)
self._append_ops(block)
loss = append_loss_ops(block, output_names)
param_grad_list = append_backward(
loss=loss, parameter_list=input_to_check, no_grad_set=no_grad_set)
feed_dict = {
item[0].name: OpTest._numpy_to_lod_tensor(item[1], item[2], place)
for p_name in inputs_with_np for item in inputs_with_np[p_name]
}
inputs = self._get_inputs(block)
feed_dict = self.feed_var(inputs, place)
fetch_list = [g for p, g in param_grad_list]
if parallel:
use_cuda = False
if isinstance(place, fluid.CUDAPlace(0)):
use_cuda = True
executor = fluid.ParallelExecutor(
use_cuda=use_cuda, loss_name=loss.name, main_program=program)
else:
executor = Executor(place)
return map(np.array,
executor.run(prog, feed_dict, fetch_list,
......
python/paddle/fluid/tests/unittests/test_lstm_op.py
浏览文件 @ bfd42683
......@@ -194,107 +194,104 @@ class TestLstmOp(OpTest):
['Input', 'Weight', 'Bias'], ['Hidden'], max_relative_error=5e-4)
class TestLstmOpHasInitial(TestLstmOp):
def set_argument(self):
self.lod = [[0, 2, 5, 7]]
self.D = 16
self.act_gate = 'sigmoid'
self.act_cell = 'tanh'
self.act_cand = 'tanh'
self.has_initial_state = True
self.is_reverse = True
self.use_peepholes = True
def test_check_grad(self):
# TODO(qingqing) remove folowing lines after the check_grad is refined.
N = len(self.lod[0]) - 1
self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
self.outputs['BatchCellPreAct'] = np.zeros(
(N, self.D)).astype('float64')
self.check_grad(
['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'],
max_relative_error=5e-4)
def test_check_grad_ingore_bias(self):
N = len(self.lod[0]) - 1
self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
self.outputs['BatchCellPreAct'] = np.zeros(
(N, self.D)).astype('float64')
self.check_grad(
['Input', 'Weight'], ['Hidden'],
max_relative_error=5e-4,
no_grad_set=set('Bias'))
def test_check_grad_ingore_weight(self):
N = len(self.lod[0]) - 1
self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
self.outputs['BatchCellPreAct'] = np.zeros(
(N, self.D)).astype('float64')
self.check_grad(
['Input', 'Bias'], ['Hidden'],
max_relative_error=5e-4,
no_grad_set=set('Weight'))
def test_check_grad_ingore_input(self):
N = len(self.lod[0]) - 1
self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
self.outputs['BatchCellPreAct'] = np.zeros(
(N, self.D)).astype('float64')
self.check_grad(
['Weight', 'Bias'], ['Hidden'],
max_relative_error=5e-4,
no_grad_set=set('Input'))
def test_check_grad_ingore_h0(self):
N = len(self.lod[0]) - 1
self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
self.outputs['BatchCellPreAct'] = np.zeros(
(N, self.D)).astype('float64')
self.check_grad(
['Input', 'Weight', 'Bias', 'C0'], ['Hidden'],
max_relative_error=5e-4,
no_grad_set=set('H0'))
def test_check_grad_ingore_c0(self):
N = len(self.lod[0]) - 1
self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
self.outputs['BatchCellPreAct'] = np.zeros(
(N, self.D)).astype('float64')
self.check_grad(
['Input', 'Weight', 'Bias', 'H0'], ['Hidden'],
max_relative_error=5e-4,
no_grad_set=set('C0'))
class TestLstmOpRerverse(TestLstmOp):
def set_argument(self):
self.lod = [[0, 2, 5, 7]]
self.D = 16
self.act_gate = 'sigmoid'
self.act_cell = 'tanh'
self.act_cand = 'tanh'
self.has_initial_state = False
self.is_reverse = True
self.use_peepholes = True
class TestLstmOpNotUsePeepholes(TestLstmOp):
def set_argument(self):
self.lod = [[0, 2, 5, 7]]
self.D = 16
self.act_gate = 'sigmoid'
self.act_cell = 'tanh'
self.act_cand = 'tanh'
self.has_initial_state = False
self.is_reverse = True
self.use_peepholes = False
# class TestLstmOpHasInitial(TestLstmOp):
# def set_argument(self):
# self.lod = [[0, 2, 5, 7]]
# self.D = 16
# self.act_gate = 'sigmoid'
# self.act_cell = 'tanh'
# self.act_cand = 'tanh'
# self.has_initial_state = True
# self.is_reverse = True
# self.use_peepholes = True
# def test_check_grad(self):
# # TODO(qingqing) remove folowing lines after the check_grad is refined.
# N = len(self.lod[0]) - 1
# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
# self.outputs['BatchCellPreAct'] = np.zeros(
# (N, self.D)).astype('float64')
# self.check_grad(
# ['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'],
# max_relative_error=5e-4)
# def test_check_grad_ingore_bias(self):
# N = len(self.lod[0]) - 1
# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
# self.outputs['BatchCellPreAct'] = np.zeros(
# (N, self.D)).astype('float64')
# self.check_grad(
# ['Input', 'Weight'], ['Hidden'],
# max_relative_error=5e-4,
# no_grad_set=set('Bias'))
# def test_check_grad_ingore_weight(self):
# N = len(self.lod[0]) - 1
# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
# self.outputs['BatchCellPreAct'] = np.zeros(
# (N, self.D)).astype('float64')
# self.check_grad(
# ['Input', 'Bias'], ['Hidden'],
# max_relative_error=5e-4,
# no_grad_set=set('Weight'))
# def test_check_grad_ingore_input(self):
# N = len(self.lod[0]) - 1
# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
# self.outputs['BatchCellPreAct'] = np.zeros(
# (N, self.D)).astype('float64')
# self.check_grad(
# ['Weight', 'Bias'], ['Hidden'],
# max_relative_error=5e-4,
# no_grad_set=set('Input'))
# def test_check_grad_ingore_h0(self):
# N = len(self.lod[0]) - 1
# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
# self.outputs['BatchCellPreAct'] = np.zeros(
# (N, self.D)).astype('float64')
# self.check_grad(
# ['Input', 'Weight', 'Bias', 'C0'], ['Hidden'],
# max_relative_error=5e-4,
# no_grad_set=set('H0'))
# def test_check_grad_ingore_c0(self):
# N = len(self.lod[0]) - 1
# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64')
# self.outputs['BatchCellPreAct'] = np.zeros(
# (N, self.D)).astype('float64')
# self.check_grad(
# ['Input', 'Weight', 'Bias', 'H0'], ['Hidden'],
# max_relative_error=5e-4,
# no_grad_set=set('C0'))
# class TestLstmOpRerverse(TestLstmOp):
# def set_argument(self):
# self.lod = [[0, 2, 5, 7]]
# self.D = 16
# self.act_gate = 'sigmoid'
# self.act_cell = 'tanh'
# self.act_cand = 'tanh'
# self.has_initial_state = False
# self.is_reverse = True
# self.use_peepholes = True
# class TestLstmOpNotUsePeepholes(TestLstmOp):
# def set_argument(self):
# self.lod = [[0, 2, 5, 7]]
# self.D = 16
# self.act_gate = 'sigmoid'
# self.act_cell = 'tanh'
# self.act_cand = 'tanh'
# self.has_initial_state = False
# self.is_reverse = True
# self.use_peepholes = False
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/testsuite.py 0 → 100644
浏览文件 @ bfd42683
# 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 numpy as np
import paddle.fluid.core as core
from paddle.fluid.op import Operator
def as_lodtensor(np_array, lod, place):
tensor = core.LoDTensor()
tensor.set(np_value, place)
if lod is not None:
tensor.set_lod(lod)
return tensor
def create_op(scope, op_type, inputs, outputs, attrs):
kwargs = dict()
op_maker = core.op_proto_and_checker_maker
op_role_attr_name = op_maker.kOpRoleAttrName()
if op_role_attr_name not in attrs:
attrs[op_role_attr_name] = int(op_maker.OpRole.Forward)
def __create_var__(name, var_name):
scope.var(var_name).get_tensor()
kwargs[name].append(var_name)
for in_name, in_dup in Operator.get_op_inputs(op_type):
if in_name in inputs:
kwargs[in_name] = []
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, _ = item[0], item[1]
__create_var__(in_name, sub_in_name)
else:
__create_var__(in_name, in_name)
for out_name, out_dup in Operator.get_op_outputs(op_type):
if out_name in outputs:
kwargs[out_name] = []
if out_dup:
sub_out = outputs[out_name]
for item in sub_out:
sub_out_name, _ = item[0], item[1]
__create_var__(out_name, sub_out_name)
else:
__create_var__(out_name, out_name)
for attr_name in Operator.get_op_attr_names(op_type):
if attr_name in attrs:
kwargs[attr_name] = attrs[attr_name]
return Operator(op_type, **kwargs)
def set_input(scope, op, inputs, place):
def __set_input__(var_name, var):
if isinstance(var, tuple) or isinstance(var, np.ndarray):
tensor = scope.find_var(var_name).get_tensor()
if isinstance(var, tuple):
tensor.set_lod(var[1])
var = var[0]
tensor.set_dims(var.shape)
tensor.set(var, place)
elif isinstance(var, float):
scope.find_var(var_name).set_float(var)
elif isinstance(var, int):
scope.find_var(var_name).set_int(var)
for in_name, in_dup in Operator.get_op_inputs(op.type()):
if in_name in inputs:
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, sub_in_val = item[0], item[1]
__set_input__(sub_in_name, sub_in_val)
else:
__set_input__(in_name, inputs[in_name])
def append_input_output(block, op_proto, np_list, is_input, dtype):
'''Insert VarDesc and generate Python variable instance'''
proto_list = op_proto.inputs if is_input else op_proto.outputs
def create_var(block, name, np_list, var_proto):
dtype = None
shape = None
lod_level = None
if name not in np_list:
assert var_proto.intermediate, "{} not found".format(name)
else:
np_value = np_list[name]
if isinstance(np_value, tuple):
dtype = np_value[0].dtype
# output shape, lod should be infered from input.
if is_input:
shape = list(np_value[0].shape)
lod_level = len(np_value[1])
else:
dtype = np_value.dtype
if is_input:
shape = list(np_value.shape)
lod_level = 0
return block.create_var(
dtype=dtype, shape=shape, lod_level=lod_level, name=name)
var_dict = {}
for var_proto in proto_list:
var_name = str(var_proto.name)
if is_input:
if (var_name not in np_list) and var_proto.dispensable:
continue
assert (var_name in np_list) or (var_proto.dispensable), \
"Missing {} as input".format(var_name)
if var_proto.duplicable:
assert isinstance(np_list[var_name], list), \
"Duplicable {} should be set as list".format(var_name)
var_list = []
for (name, np_value) in np_list[var_name]:
var_list.append(
create_var(block, name, {name: np_value}, var_proto))
var_dict[var_name] = var_list
else:
var_dict[var_name] = create_var(block, var_name, np_list, var_proto)
return var_dict
def append_loss_ops(block, output_names):
mean_inputs = map(block.var, output_names)
# for item in mean_inputs:
# print(item)
# print("Item", item.dtype)
if len(mean_inputs) == 1:
loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1])
op = block.append_op(
inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean')
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
else:
avg_sum = []
for cur_loss in mean_inputs:
cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1])
op = block.append_op(
inputs={"X": [cur_loss]},
outputs={"Out": [cur_avg_loss]},
type="mean")
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
avg_sum.append(cur_avg_loss)
loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1])
op_sum = block.append_op(
inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum')
op_sum.desc.infer_var_type(block.desc)
op_sum.desc.infer_shape(block.desc)
loss = block.create_var(dtype=loss_sum.dtype, shape=[1])
op_loss = block.append_op(
inputs={"X": loss_sum},
outputs={"Out": loss},
type='scale',
attrs={'scale': 1.0 / float(len(avg_sum))})
op_loss.desc.infer_var_type(block.desc)
op_loss.desc.infer_shape(block.desc)
return loss
Markdown is supported
0% 或 .
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 或