提交 78949c07 编写于 作者: S Siddharth Goyal 提交者: Yiqun Liu

Inference example for image_classification and unit_test for "inference" (#8020)

* First basic implementation

* Add infer example for image_classification

* Address review comments: round 1
上级 84994608
...@@ -3,5 +3,17 @@ cc_test(test_inference_recognize_digits_mlp ...@@ -3,5 +3,17 @@ cc_test(test_inference_recognize_digits_mlp
SRCS test_inference_recognize_digits.cc SRCS test_inference_recognize_digits.cc
DEPS ARCHIVE_START paddle_fluid ARCHIVE_END DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
ARGS --dirname=${PYTHON_TESTS_DIR}/book/recognize_digits_mlp.inference.model) ARGS --dirname=${PYTHON_TESTS_DIR}/book/recognize_digits_mlp.inference.model)
cc_test(test_inference_image_classification_vgg
SRCS test_inference_image_classification.cc
DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
ARGS --dirname=${PYTHON_TESTS_DIR}/book/image_classification_vgg.inference.model)
cc_test(test_inference_image_classification_resnet
SRCS test_inference_image_classification.cc
DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
ARGS --dirname=${PYTHON_TESTS_DIR}/book/image_classification_resnet.inference.model)
set_tests_properties(test_inference_recognize_digits_mlp set_tests_properties(test_inference_recognize_digits_mlp
PROPERTIES DEPENDS test_recognize_digits) PROPERTIES DEPENDS test_recognize_digits)
set_tests_properties(test_inference_image_classification_vgg
PROPERTIES DEPENDS test_image_classification_train)
set_tests_properties(test_inference_image_classification_resnet
PROPERTIES DEPENDS test_image_classification_train)
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
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 <gtest/gtest.h>
#include <time.h>
#include <sstream>
#include "gflags/gflags.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/inference/io.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
template <typename Place, typename T>
void TestInference(const std::string& dirname,
const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
// 1. Define place, executor and scope
auto place = Place();
auto executor = paddle::framework::Executor(place);
auto* scope = new paddle::framework::Scope();
// 2. Initialize the inference_program and load all parameters from file
auto inference_program = paddle::inference::Load(executor, *scope, dirname);
// 3. Get the feed_target_names and fetch_target_names
const std::vector<std::string>& feed_target_names =
inference_program->GetFeedTargetNames();
const std::vector<std::string>& fetch_target_names =
inference_program->GetFetchTargetNames();
// 4. Prepare inputs: set up maps for feed targets
std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
for (size_t i = 0; i < feed_target_names.size(); ++i) {
// Please make sure that cpu_feeds[i] is right for feed_target_names[i]
feed_targets[feed_target_names[i]] = cpu_feeds[i];
}
// 5. Define Tensor to get the outputs: set up maps for fetch targets
std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
for (size_t i = 0; i < fetch_target_names.size(); ++i) {
fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
}
// 6. Run the inference program
executor.Run(*inference_program, scope, feed_targets, fetch_targets);
delete scope;
}
TEST(inference, image_classification) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
std::string dirname = FLAGS_dirname;
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
paddle::framework::LoDTensor input;
srand(time(0));
float* input_ptr =
input.mutable_data<float>({1, 3, 32, 32}, paddle::platform::CPUPlace());
for (int i = 0; i < 3072; ++i) {
input_ptr[i] = rand() / (static_cast<float>(RAND_MAX));
}
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
cpu_feeds.push_back(&input);
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace, float>(
dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace, float>(
dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << output2.dims();
EXPECT_EQ(output1.dims(), output2.dims());
EXPECT_EQ(output1.numel(), output2.numel());
float err = 1E-3;
int count = 0;
for (int64_t i = 0; i < output1.numel(); ++i) {
if (fabs(output1.data<float>()[i] - output2.data<float>()[i]) > err) {
count++;
}
}
EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
#endif
}
...@@ -16,8 +16,9 @@ from __future__ import print_function ...@@ -16,8 +16,9 @@ from __future__ import print_function
import paddle.v2 as paddle import paddle.v2 as paddle
import paddle.v2.fluid as fluid import paddle.v2.fluid as fluid
import unittest
import contextlib import contextlib
import numpy
import unittest
def resnet_cifar10(input, depth=32): def resnet_cifar10(input, depth=32):
...@@ -89,10 +90,7 @@ def vgg16_bn_drop(input): ...@@ -89,10 +90,7 @@ def vgg16_bn_drop(input):
return fc2 return fc2
def main(net_type, use_cuda): def train(net_type, use_cuda, save_dirname):
if use_cuda and not fluid.core.is_compiled_with_cuda():
return
classdim = 10 classdim = 10
data_shape = [3, 32, 32] data_shape = [3, 32, 32]
...@@ -111,12 +109,14 @@ def main(net_type, use_cuda): ...@@ -111,12 +109,14 @@ def main(net_type, use_cuda):
predict = fluid.layers.fc(input=net, size=classdim, act='softmax') predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label) cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost) avg_cost = fluid.layers.mean(x=cost)
acc = fluid.layers.accuracy(input=predict, label=label)
# Test program
test_program = fluid.default_main_program().clone()
optimizer = fluid.optimizer.Adam(learning_rate=0.001) optimizer = fluid.optimizer.Adam(learning_rate=0.001)
optimizer.minimize(avg_cost) optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
BATCH_SIZE = 128 BATCH_SIZE = 128
PASS_NUM = 1 PASS_NUM = 1
...@@ -125,6 +125,9 @@ def main(net_type, use_cuda): ...@@ -125,6 +125,9 @@ def main(net_type, use_cuda):
paddle.dataset.cifar.train10(), buf_size=128 * 10), paddle.dataset.cifar.train10(), buf_size=128 * 10),
batch_size=BATCH_SIZE) batch_size=BATCH_SIZE)
test_reader = paddle.batch(
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place) exe = fluid.Executor(place)
feeder = fluid.DataFeeder(place=place, feed_list=[images, label]) feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
...@@ -132,18 +135,68 @@ def main(net_type, use_cuda): ...@@ -132,18 +135,68 @@ def main(net_type, use_cuda):
loss = 0.0 loss = 0.0
for pass_id in range(PASS_NUM): for pass_id in range(PASS_NUM):
accuracy.reset(exe) for batch_id, data in enumerate(train_reader()):
for data in train_reader(): exe.run(feed=feeder.feed(data))
loss, acc = exe.run(fluid.default_main_program(),
feed=feeder.feed(data), if (batch_id % 10) == 0:
fetch_list=[avg_cost] + accuracy.metrics) acc_list = []
pass_acc = accuracy.eval(exe) avg_loss_list = []
print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str( for tid, test_data in enumerate(test_reader()):
pass_acc)) loss_t, acc_t = exe.run(program=test_program,
return feed=feeder.feed(test_data),
fetch_list=[avg_cost, acc])
raise AssertionError( acc_list.append(float(acc_t))
"Image classification loss is too large, {0:2.2}".format(loss)) avg_loss_list.append(float(loss_t))
break # Use 1 segment for speeding up CI
acc_value = numpy.array(acc_list).mean()
avg_loss_value = numpy.array(avg_loss_list).mean()
print(
'PassID {0:1}, BatchID {1:04}, Test Loss {2:2.2}, Acc {3:2.2}'.
format(pass_id, batch_id + 1,
float(avg_loss_value), float(acc_value)))
if acc_value > 0.01: # Low threshold for speeding up CI
fluid.io.save_inference_model(save_dirname, ["pixel"],
[predict], exe)
return
def infer(use_cuda, save_dirname=None):
if save_dirname is None:
return
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
# Use fluid.io.load_inference_model to obtain the inference program desc,
# the feed_target_names (the names of variables that will be feeded
# data using feed operators), and the fetch_targets (variables that
# we want to obtain data from using fetch operators).
[inference_program, feed_target_names,
fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
# The input's dimension of conv should be 4-D or 5-D.
tensor_img = numpy.random.rand(1, 3, 32, 32).astype("float32")
# Construct feed as a dictionary of {feed_target_name: feed_target_data}
# and results will contain a list of data corresponding to fetch_targets.
results = exe.run(inference_program,
feed={feed_target_names[0]: tensor_img},
fetch_list=fetch_targets)
print("infer results: ", results[0])
def main(net_type, use_cuda):
if use_cuda and not fluid.core.is_compiled_with_cuda():
return
# Directory for saving the trained model
save_dirname = "image_classification_" + net_type + ".inference.model"
train(net_type, use_cuda, save_dirname)
infer(use_cuda, save_dirname)
class TestImageClassification(unittest.TestCase): class TestImageClassification(unittest.TestCase):
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册