diff --git a/contrib/float16/float16_inference_report.md b/contrib/float16/README.md similarity index 51% rename from contrib/float16/float16_inference_report.md rename to contrib/float16/README.md index 67623a4d8d58b2ab84031ac4b290931ab540973b..ded959c47cb81b9384abbb9815773e25969344ec 100644 --- a/contrib/float16/float16_inference_report.md +++ b/contrib/float16/README.md @@ -1,33 +1,40 @@ +# Float16 Inference in PaddlePaddle Fluid + +Kexin Zhao + ## Introduction -Working with deep neural networks (DNN) is a two-stage process. First we train DNN using labeled examples of inputs and desired outputs to obtain the model parameters (weights), then we deploy DNN along with the trained weights to run inference on unknown inputs. Typically, these weights are in float data type and hence we run inference in float mode using these weights. This post focuses on the discussion of how to use low precision float16 data type to represent these trained weights and run inference in float16 mode as well as the advantages of float16 inference over its float counterpart by showing some experiment results. +Deep learning is usually a two-stage work: training and inference. The training stage estimates model parameters (weights) from data. The inference stage loads the weights and uses them to interpret inputs. Typically, weights are 32-bit float values (float32). Some new devices, including NVIDIA Volta GPUs, support higher speed computation using 16-bit float values (float16). + +This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16. + ## What is float16? -float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has high tolerance against the loss of precision and range when using float16 to represent the weights and the inference accuracy will only be minimally affected in most cases. This gives us the opportunity to use float16 data type to speedup the inference. +float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float or float32 data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has a high tolerance for the loss of precision and range when using float16 to represent the weights, and the inference accuracy will only be minimally affected in most cases, which gives us the opportunity to use float16 data type to speed up the inference. Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type. ## Why float16? -The trend in today's deep learning community is to use bigger and deeper model. This translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float are correspondingly three-fold: +The trend in today's deep learning community is to use bigger and deeper model, which translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float32 are correspondingly three-fold: -1. We only need half the memory size to load the same model using float16 representations. Moreover, most of the intermediate results generated during float16 inference are also of float16 data type. This makes the whole memory footprint of float16 inference roughly about half of its float counterpart. This is especially useful when deploying inference on mobile devices with limited available memory. Also given the same available memory, the maximum batch size for float16 inference is about twice that for float inference. +1. We only need half the memory size to load the same model using float16 representations. Moreover, most of the intermediate results generated during float16 inference are also of the float16 data type. As a result, the whole memory footprint of float16 inference is roughly half of its float counterpart, which is especially useful when deploying inference on mobile devices with limited available memory. Also given the same available memory, the maximum batch size for float16 inference is about twice that for float inference. -2. Because float16 occupies less memory than float, in theory hardware devices can achieve much higher floating point operators per second (FLOPS) for float16 data than float data. Right now, an outstanding example of hardware devices that actually deliver such advantages is Nvidia's latest Volta architecture GPUs, including Tesla V100 and Titan V. Moreover float16 takes less time to read from or write to memory and hence float16 can make inference more efficient especially in memory-bound applications where the performance is largely affected by how fast it is to read and write data. +2. Because float16 occupies less memory than float, in theory, hardware devices can achieve much higher floating point operators per second (FLOPS) for float16 data than float data. Right now, NVIDIA's latest Volta GPUs, including Tesla V100 and Titan V, can deliver significantly higher FLOPS for float16 using Tensor Cores. Moreover, float16 takes less time to read from or write to memory, and hence float16 can make inference more efficient especially in memory-bound applications where the performance is mostly affected by how fast it is to read and write data. -3. From the energy efficiency perspective, the energy needed to read, write, and compute float16 data is much less that its float counterpart, which can significantly reduce the battery power consumption on mobile devices or the total cost of ownership (TCO) of data centers. +3. From the energy efficiency perspective, the energy needed to read, write, and compute float16 data is much less than its float counterpart, which can significantly reduce the battery power consumption on mobile devices or the total cost of ownership (TCO) of data centers. ## Fluid implementation of float16 inference ### Overview Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block. ### Basic requirement -When an operator is run by an executor, it uses a kernel to perform computations on tensors contained in the input variables, and then write the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs. +When an executor runs an operator, it uses a kernel to perform computations on tensors contained in the input variables, and then writes the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs. -This means that if we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will makes the program run in float mode and gives us a final output of float data type. +If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type. -The same principle applies if we want a program to run in float16 mode. We provide input variable of float16 data type to the first operator and every subsequent operator will invoke the float16 kernel until we get the final output in float16 data type. So the preliminary requirements for float16 inference is to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax. +The same principle applies if we want a program to run in float16 mode. We provide input variable of the float16 data type to the first operator, and every subsequent operator will invoke the float16 kernel until we get the final output in float16. So the preliminary requirements for float16 inference are to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax. ### float16 transpiler -Furthermore, we need a float16 transpiler to achieve the following usage code: +Furthermore, we need a transpiler to write float16 inference code similar to the following: ```python # Get the float32 inference program and load the associated float32 weights @@ -64,14 +71,15 @@ fluid.io.save_inference_model(fp16_save_dirname, feed_target_names, float16_inference_program) ``` -In this scenario, we already have a float32 inference program and some associated float32 weights that can do float32 inference. We can easily use the `transpile` method of the `Float16Transpiler` class to do certain modifications to the existing program and weights so that we have a new float16 program and the associated float16 weights. +In this scenario, we already have a float32 inference program and some associated float32 weights. We can simply use the `transpile` method of the `Float16Transpiler` class to do certain modifications to the existing program and weights so that we have a new float16 program and the associated float16 weights. -We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. This requires us to add some cast operators in the program to convert between float16 tensor and float32 tensor. +We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. Consequently, we need to add cast operators in the float16 inference program for conversions between the float16 tensor and float32 tensor. The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference). ### Experiment results -We provide demo codes that can be used to reproduce the experiment results by doing: +Simply running the following commands to reproduce the experiment results presented in this section: + ```bash git clone https://github.com/PaddlePaddle/Paddle.git cd Paddle @@ -84,8 +92,8 @@ nvidia-docker build -t paddle:float16 . nvidia-docker run -it -v $PWD:/paddle paddle:float16 /paddle/contrib/float16/run_float16_demo.sh ``` -#### Correctness -As is mentioned before, DNN inference has been found to be tolerant against the loss of precision and range incured by float16 and we want to see how good this tolerance is. +#### Accuracy +As is mentioned before, DNN inference has been found to be tolerant against the loss of precision and range incurred by float16, and we want to see how good this tolerance is. We train a resnet32 model using cifar10 data set, save it when test set accuracy is above 60%, and then test the inference accuracy on the 10000 examples of the cifar10 test set in float16 and float32 mode, respectively. @@ -105,18 +113,18 @@ We repeat the test ten times and get the following results: | #10 | 62.53% | 62.48% | | average| 62.63% | 62.62% | -We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over 10 tests. +We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests. #### Performance benchmark -Currently, Fluid inference in float16 mode is only supported on Nvidia GPU device. There is no motivation to support float16 inference on non-ARM CPUs because float16 is not natively supported there and float16 calculation will only be slower than its float counterpart. +Currently, Fluid only supports float16 inference on NVIDIA GPUs. There is no motivation to support float16 inference on non-ARM CPUs where float16 is not natively supported, and float16 calculation will only be slower than its float32 counterpart. -Nvidia started to support its native float16 data type (which has the same internal memory representation as Fluid float16 class) on CUDA 7.5. Moreover, float16 speedups on common computational intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cublas 7.5 and cuDNN 5.0. +NVIDIA started to support its native float16 data type (which has the same internal memory representation as Fluid's float16 class) on CUDA 7.5. Moreover, float16 speedups on computationally intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cuBLAS 7.5 and cuDNN 5.0. -Recently, the introduction of [tensor core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in volta architecture GPUs and the support of tensor core calculation in CUDA 9.0 and cuDNN 7 make float16 truly superior to float in certain deep learning applications. +Recently, the introduction of [Tensor Core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in Volta architecture GPUs and the support of Tensor Core computation in CUDA 9.0 and cuDNN 7 make float16 genuinely superior to float in some deep learning applications. -We thus benchmark the float16 inference performance on a single Nvidia Tesla V100 GPU (volta architecture and with tensor cores) and compare it with its float32 counterpart. All the following results are in ms (millisecond) averaged over 1000 mini-batches with respective to different mini-batch(mb) sizes. +We thus benchmark the float16 inference performance on a single NVIDIA Tesla V100 GPU (Volta architecture and with Tensor Cores) and compare it with its float32 counterpart. All the following results are in ms (millisecond) averaged over 1000 mini-batches with respective to different mini-batch(mb) sizes. -Average inference time for one mini-batch on Vgg16 model tested on imagenet data set: +Average inference time for one mini-batch on Vgg16 model tested on ImageNet dataset: | total | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | |-------|-----: |-----: |-----: |-----: |------: |------:|-------:| @@ -124,7 +132,7 @@ Average inference time for one mini-batch on Vgg16 model tested on imagenet data |float16| 3.32 | 4.11 | 5.88 | 9.41 | 16.54 | 30.47 | 60.23 | |Speedup| 4.22 | 2.36  | 3.91 | 3.00 | 3.26  | 2.77 | 2.97 | -We can see that float16 inference provides 2x ~ 4x speedup on different batch sizes. +We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes. Convolution operation is ususally the computational bottleneck of CNN, so we also check the average time spent on the Fluid convolution operators for one mini-batch as follows: @@ -134,9 +142,9 @@ Convolution operation is ususally the computational bottleneck of CNN, so we als |float16| 1.78 | 2.10 | 2.93 | 4.55 | 7.99 | 14.63 | 28.67 | |Speedup| 6.71 | 3.31  | 6.37 | 4.71 | 5.18  | 4.14 | 4.54 | -Fluid convolution operator uses cuDNN 7 to implement the kernel and we can see that with the help of tensor core, float16 convolution is significantly faster than its float32 counterpart, which makes the overall float16 inference performance much better. +Fluid convolution operator uses cuDNN 7 to implement the kernel, and we can see that with the help of Tensor Core, float16 convolution is significantly faster than its float32 counterpart, which makes the overall float16 inference performance much better. -Similarly, we also list the benchmark results of Resnet50 model tested on imagenet data set: +Similarly, we also list the benchmark results of Resnet50 model tested on the ImageNet dataset: | total | mb=1 | mb=2 | mb=4 | mb=8 | mb=16 | mb=32 | mb=64 | mb=128 | |-------|-----: |-----: |-----: |-----: |------: |------:|-------:|-------:| @@ -150,14 +158,14 @@ Similarly, we also list the benchmark results of Resnet50 model tested on imagen |float16| 4.19 | 4.30 | 3.96 | 4.21 | 5.63 | 8.77 | 15.24 | 28.40 | |Speedup| 1.30 | 1.27  | 1.64  | 1.99 | 2.45  | 2.79 | 2.70 | 2.59 | -We find that the speedup provided by float16 inference starts relatively small at 1.15x for batch size 1 and gradually increase to about 2x for larger batch sizes. Similar trend can be found for the time spent on the convolution operator. Note that right now the tensor core will only be utilized in the convolution operation when certain dimentional requirements are met for the input data and filter. The speedup by float16 inference for Resnet50 is smaller than the Vgg16 counterpart partially because the convolution operation in Resnet is much simpler than the Vgg counterpart and this makes the tensor core less utilized in Resnet than in Vgg. +We find that the speedup provided by float16 inference starts relatively small at 1.15x for batch size 1 and gradually increases to about 2x for larger batch sizes. A similar trend can be found for the time spent on the convolution operator. Note that right now Tensor Cores will only be utilized in the convolution operation when the input data and filter meet specific dimensional requirements. The speedup by float16 inference for Resnet50 is smaller than the Vgg16 counterpart partially because the convolution operation in Resnet is much simpler than its Vgg counterpart and this makes the tensor core less utilized in Resnet than in Vgg. -We also did the same benchmark on a Nvidia GeForce GTX 1080 Ti GPU that does not support tensor core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then deliver around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that tensor core, which is specialized for float16 computations, is a critical component for high performance float16 inference. +We also did the same benchmark on a single NVIDIA GeForce GTX 1080 Ti GPU that does not support Tensor Core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then delivers around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that Tensor Core, which is specialized for float16 computations, is a critical component of high performance float16 inference. -Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for comprehensive benchmark results. +Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for complete benchmark results. ### Summary 1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode. -2. The accuracy of float16 inference is verified to be almost identical to the float32 counterpart at least on CNNs. -3. float16 inference provides significant speedup on large and computationally intensive Vgg16 network on image net data set. For the much smaller and simpler Resnet50, the speedup provided by float16 inference is less significant than on Vgg16 but still favorable especially for large batch size. -4. We cannot achieve the superior float16 inference performance without the help of the newly introduced tensor cores on the Nvidia Volta architecture GPUs. +2. The accuracy of float16 inference is verified to be almost identical to its float32 counterpart at least on CNN models. +3. float16 inference provides a significant speedup on large and computationally intensive Vgg16 model on ImageNet dataset. For the much smaller and simpler Resnet50 model, the speedup provided by float16 inference is less significant than for Vgg16 model but still favorable, especially for large batch sizes. +4. We cannot achieve the superior float16 inference performance without the help of the newly introduced Tensor Cores on NVIDIA Volta architecture GPUs. diff --git a/doc/v2/api/config/layer.rst b/doc/v2/api/config/layer.rst index 29388f5005bf779a1bfa63c0d46d35996c0c792d..1a6496968cae1fef88142ba9ca3f9e63a81b196d 100644 --- a/doc/v2/api/config/layer.rst +++ b/doc/v2/api/config/layer.rst @@ -142,7 +142,7 @@ gated_unit ----------- .. autoclass:: paddle.v2.layer.gated_unit :noindex: - + Recurrent Layer Group ===================== @@ -354,7 +354,7 @@ dropout -------- .. autoclass:: paddle.v2.layer.dropout :noindex: - + dot_prod --------- .. autoclass:: paddle.v2.layer.dot_prod @@ -460,6 +460,11 @@ multi_binary_label_cross_entropy_cost .. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost :noindex: +classification_cost +------------------- +.. autoclass:: paddle.v2.layer.classification_cost + :noindex: + huber_regression_cost ------------------------- .. autoclass:: paddle.v2.layer.huber_regression_cost @@ -534,7 +539,7 @@ detection_output ---------------- .. autoclass:: paddle.v2.layer.detection_output :noindex: - + Check Layer ============ diff --git a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md index dee1b7554f97af17989c3f7739d8feea3b6b8e3f..b2dc4da8451af317df76c5b3df328b6f58429610 100644 --- a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md +++ b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md @@ -41,7 +41,7 @@ Training docker image needs to package the paddle pserver and paddle trainer run - Generating the initialization arguments for `Paddle PServer` and `Paddle Training` processes. Since the paddlepaddle official docker image already has the runtimes we need, we'll take it as the base image and pack some additional scripts for the processes mentioned above to build our training image. for more detail, please find from the following link: -- https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile +- https://github.com/PaddlePaddle/Paddle/tree/develop/doc/v2/howto/cluster/multi_cluster/src/k8s_train/Dockerfile ```bash @@ -62,7 +62,7 @@ represent the Docker Image which built in this step. ### Prepare Training Data We can download and split the training job by creating a Kubernetes Job, or custom your image -by editing [k8s_train](./src/k8s_train/). +by editing [k8s_train](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/v2/howto/cluster/multi_cluster/src/k8s_train). Before creating a Job, we need to bind a [persistenVolumeClaim](https://kubernetes.io/docs/user-guide/persistent-volumes) by the different type of the different file system, the generated dataset would be saved on this volume. diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index c8b656394b403c4965e01e96c9215d9406091907..288789d6e484100820c937e6081701f1e9245706 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,5 +1,4 @@ nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) -nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) set(ENGINE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/engine.cc) add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 572e29515f89685c528023cf6569ef8e559d6874..3c5909c0be1c690d5148ecfb32b1b6c2dd6f3211 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,3 +1,4 @@ -nv_test(test_tensorrt_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES}) -nv_test(test_tensorrt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} activation_op.cc +nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES}) +nv_test(test_trt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} activation_op.cc DEPS ${FLUID_CORE_MODULES} activation_op) +nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) diff --git a/paddle/fluid/inference/tensorrt/io_converter.cc b/paddle/fluid/inference/tensorrt/convert/io_converter.cc similarity index 93% rename from paddle/fluid/inference/tensorrt/io_converter.cc rename to paddle/fluid/inference/tensorrt/convert/io_converter.cc index 2baac96c26453af7e70e541d80b437df3d5c2657..32e8631fde3f748669d2008b4a060455a37e154e 100644 --- a/paddle/fluid/inference/tensorrt/io_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.cc @@ -12,7 +12,7 @@ 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/inference/tensorrt/io_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" #include #include "paddle/fluid/platform/enforce.h" @@ -50,7 +50,7 @@ class DefaultInputConverter : public EngineInputConverter { } }; -REGISTER_TENSORRT_INPUT_CONVERTER(mul, DefaultInputConverter); +REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter); } // namespace tensorrt } // namespace inference diff --git a/paddle/fluid/inference/tensorrt/io_converter.h b/paddle/fluid/inference/tensorrt/convert/io_converter.h similarity index 95% rename from paddle/fluid/inference/tensorrt/io_converter.h rename to paddle/fluid/inference/tensorrt/convert/io_converter.h index 6ea61cbbac05f106f736b7d6a13912157c5ef48c..8972dae92be2c2d261a13c48d98e675f64e51d31 100644 --- a/paddle/fluid/inference/tensorrt/io_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.h @@ -40,7 +40,8 @@ class EngineInputConverter { static void Run(const std::string& in_op_type, const LoDTensor& in, void* out, size_t max_size, cudaStream_t* stream) { PADDLE_ENFORCE(stream != nullptr); - auto* converter = Registry::Lookup(in_op_type); + auto* converter = Registry::Lookup( + in_op_type, "default" /* default_type */); PADDLE_ENFORCE_NOT_NULL(converter); converter->SetStream(stream); (*converter)(in, out, max_size); diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index f8ca219bb83c6db3ba640464233c6cbd337926a6..77c788550b2c7df1f483b926661789b2a54d8fff 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -19,6 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace inference { @@ -32,34 +33,23 @@ class OpConverter { OpConverter() {} virtual void operator()(const framework::OpDesc& op) {} - void Execute(const framework::OpDesc& op, TensorRTEngine* engine) { + void Run(const framework::OpDesc& op, TensorRTEngine* engine) { std::string type = op.Type(); - auto it = converters_.find(type); - PADDLE_ENFORCE(it != converters_.end(), "no OpConverter for optype [%s]", - type); - it->second->SetEngine(engine); - (*it->second)(op); - } - - static OpConverter& Global() { - static auto* x = new OpConverter; - return *x; - } - - template - void Register(const std::string& key) { - converters_[key] = new T; + auto* it = Registry::Lookup(type); + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); + it->SetEngine(engine); + (*it)(op); } // convert fluid op to tensorrt layer void ConvertOp(const framework::OpDesc& op, TensorRTEngine* engine) { - OpConverter::Global().Execute(op, engine); + OpConverter::Run(op, engine); } // convert fluid block to tensorrt network void ConvertBlock(const framework::BlockDesc& block, TensorRTEngine* engine) { for (auto op : block.AllOps()) { - OpConverter::Global().Execute(*op, engine); + OpConverter::Run(*op, engine); } } @@ -78,12 +68,12 @@ class OpConverter { framework::Scope* scope_{nullptr}; }; -#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ - struct trt_##op_type__##_converter { \ - trt_##op_type__##_converter() { \ - OpConverter::Global().Register(#op_type__); \ - } \ - }; \ +#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ + struct trt_##op_type__##_converter { \ + trt_##op_type__##_converter() { \ + Registry::Register(#op_type__); \ + } \ + }; \ trt_##op_type__##_converter trt_##op_type__##_converter__; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc index 0f390bee1f5967547ea21e6d42a1b4300de58d04..23e3435c21725328d3765fae0d158a83ac21478b 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -26,7 +26,7 @@ namespace paddle { namespace inference { namespace tensorrt { -void compare(float input, float expect) { +void Compare(float input, float expect) { framework::Scope scope; platform::CUDAPlace place; platform::CUDADeviceContext ctx(place); @@ -85,8 +85,8 @@ void compare(float input, float expect) { } TEST(OpConverter, ConvertRelu) { - compare(1, 1); // relu(1) = 1 - compare(-5, 0); // relu(-5) = 0 + Compare(1, 1); // relu(1) = 1 + Compare(-5, 0); // relu(-5) = 0 } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/test_io_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc similarity index 87% rename from paddle/fluid/inference/tensorrt/test_io_converter.cc rename to paddle/fluid/inference/tensorrt/convert/test_io_converter.cc index 365e9366862bee25c70dba0cdd92f318ab3ee90f..afcc516e6b76d58e37ce0e60746704cf3933fac7 100644 --- a/paddle/fluid/inference/tensorrt/test_io_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/inference/tensorrt/io_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" #include @@ -34,7 +34,7 @@ TEST_F(EngineInputConverterTester, DefaultCPU) { ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0); cudaStream_t stream; - EngineInputConverter::Run("mul", tensor, buffer, tensor.memory_size(), + EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(), &stream); } @@ -44,7 +44,7 @@ TEST_F(EngineInputConverterTester, DefaultGPU) { ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0); cudaStream_t stream; - EngineInputConverter::Run("mul", tensor, buffer, tensor.memory_size(), + EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(), &stream); } diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index 5c5ac10394d3cb76ac4b8409c7fa79eb5e9a90b2..aa5fb726f1129eda65a6f39791330b795aad660d 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -20,7 +20,7 @@ namespace paddle { namespace inference { namespace tensorrt { -TEST(BlockConverter, ConvertBlock) { +TEST(OpConverter, ConvertBlock) { framework::ProgramDesc prog; auto* block = prog.MutableBlock(0); auto* mul_op = block->AppendOp(); diff --git a/paddle/fluid/inference/utils/singleton.h b/paddle/fluid/inference/utils/singleton.h index f05921067c45f156319375b613f51101cfda8e90..cfb89e704457a11a3cd6e89dba5efad5acae0bce 100644 --- a/paddle/fluid/inference/utils/singleton.h +++ b/paddle/fluid/inference/utils/singleton.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include "paddle/fluid/platform/enforce.h" @@ -49,9 +50,15 @@ struct Registry { items_[name] = new ItemChild; } - static ItemParent* Lookup(const std::string& name) { + static ItemParent* Lookup(const std::string& name, + const std::string& default_name = "") { auto it = items_.find(name); - if (it == items_.end()) return nullptr; + if (it == items_.end()) { + if (default_name == "") + return nullptr; + else + return items_.find(default_name)->second; + } return it->second; } diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index 37d368946770978700abe49eef6825e1d96839f0..c8a435748dc5b51bf9e57b5b597e1422f0380e8e 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -60,6 +60,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ +\ 'io', 'initializer', 'layers', + 'transpiler' 'nets', 'optimizer', 'learning_rate_decay', diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index ce9b880aeb31f7706820b89fb7e1edc6c51ba69b..d7eda619c34a20fa09a30afdcf90047d66a05cbf 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -1042,13 +1042,14 @@ class Program(object): Returns(Program): The cloned Program object. """ - p = Program() if for_test: - p.desc = core.inference_optimize(self.desc) + p = self.inference_optimize() else: + p = Program() p.desc = core.ProgramDesc(self.desc) - p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] - p.sync_with_cpp() + p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] + p.sync_with_cpp() + p.copy_param_info_from(self) return p @@ -1061,7 +1062,7 @@ class Program(object): if isinstance(t, Variable): # After transpiler processing, the op that output this # variable maybe has been changed, so t.op is not reliable - # and we need to find the current op that generate this + # and we need to find the current op that generate this # variable here. t.op = None global_block = self.global_block() @@ -1087,8 +1088,16 @@ class Program(object): return res def inference_optimize(self): + # this is an alternative implement before + # core.inference_optimize being fixed. res = Program() - res.desc = core.inference_optimize(self.desc) + res.desc = core.ProgramDesc(self.desc) + for i in xrange(res.desc.num_blocks()): + block = res.desc.block(i) + for j in xrange(block.op_size()): + op = block.op(j) + if op.has_attr('is_test'): + op.set_attr('is_test', True) res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())] res.sync_with_cpp() return res diff --git a/python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py b/python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py new file mode 100755 index 0000000000000000000000000000000000000000..fe36e55bb5380975ae322eccbcd8ad41e1e6748a --- /dev/null +++ b/python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py @@ -0,0 +1,228 @@ +# 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. + +from __future__ import print_function + +import paddle +import paddle.fluid as fluid +import numpy + +WORD_DICT, VERB_DICT, LABEL_DICT = paddle.dataset.conll05.get_dict() +WORD_DICT_LEN = len(WORD_DICT) +LABEL_DICT_LEN = len(LABEL_DICT) +PRED_DICT_LEN = len(VERB_DICT) +MARK_DICT_LEN = 2 + + +def lstm_net(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark): + WORD_DIM = 32 + MARK_DIM = 5 + HIDDEN_DIM = 512 + DEPTH = 8 + EMBEDDING_NAME = 'emb' + + # Data definitions + word = fluid.layers.data( + name='word_data', shape=[1], dtype='int64', lod_level=1) + predicate = fluid.layers.data( + name='verb_data', shape=[1], dtype='int64', lod_level=1) + ctx_n2 = fluid.layers.data( + name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1) + ctx_n1 = fluid.layers.data( + name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1) + ctx_0 = fluid.layers.data( + name='ctx_0_data', shape=[1], dtype='int64', lod_level=1) + ctx_p1 = fluid.layers.data( + name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1) + ctx_p2 = fluid.layers.data( + name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1) + mark = fluid.layers.data( + name='mark_data', shape=[1], dtype='int64', lod_level=1) + + # 8 features + predicate_embedding = fluid.layers.embedding( + input=predicate, + size=[PRED_DICT_LEN, WORD_DIM], + dtype='float32', + is_sparse=IS_SPARSE, + param_attr='vemb') + + mark_embedding = fluid.layers.embedding( + input=mark, + size=[MARK_DICT_LEN, MARK_DIM], + dtype='float32', + is_sparse=IS_SPARSE) + + word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] + emb_layers = [ + fluid.layers.embedding( + size=[WORD_DICT_LEN, WORD_DIM], + input=x, + param_attr=fluid.ParamAttr( + name=EMBEDDING_NAME, trainable=False)) for x in word_input + ] + emb_layers.append(predicate_embedding) + emb_layers.append(mark_embedding) + + hidden_0_layers = [ + fluid.layers.fc(input=emb, size=HIDDEN_DIM, act='tanh') + for emb in emb_layers + ] + + hidden_0 = fluid.layers.sums(input=hidden_0_layers) + + lstm_0 = fluid.layers.dynamic_lstm( + input=hidden_0, + size=HIDDEN_DIM, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid') + + # stack L-LSTM and R-LSTM with direct edges + input_tmp = [hidden_0, lstm_0] + + for i in range(1, DEPTH): + mix_hidden = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=HIDDEN_DIM, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=HIDDEN_DIM, act='tanh') + ]) + + lstm = fluid.layers.dynamic_lstm( + input=mix_hidden, + size=HIDDEN_DIM, + candidate_activation='relu', + gate_activation='sigmoid', + cell_activation='sigmoid', + is_reverse=((i % 2) == 1)) + + input_tmp = [mix_hidden, lstm] + + feature_out = fluid.layers.sums(input=[ + fluid.layers.fc(input=input_tmp[0], size=LABEL_DICT_LEN, act='tanh'), + fluid.layers.fc(input=input_tmp[1], size=LABEL_DICT_LEN, act='tanh') + ]) + + return feature_out + + +def inference_network(): + predict = lstm_net(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, + mark) + + crf_decode = fluid.layers.crf_decoding( + input=feature_out, param_attr=fluid.ParamAttr(name='crfw')) + + return crf_decode + + +def train_network(): + MIX_HIDDEN_LR = 1e-3 + + predict = lstm_net(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, + mark) + target = fluid.layers.data( + name='target', shape=[1], dtype='int64', lod_level=1) + crf_cost = fluid.layers.linear_chain_crf( + input=predict, + label=target, + param_attr=fluid.ParamAttr( + name='crfw', learning_rate=MIX_HIDDEN_LR)) + avg_cost = fluid.layers.mean(crf_cost) + + return avg_cost + + +def train(use_cuda, save_path): + BATCH_SIZE = 128 + EPOCH_NUM = 1 + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.conll05.train(), buf_size=8192), + batch_size=BATCH_SIZE) + test_reader = paddle.batch( + paddle.dataset.conll05.test(), batch_size=BATCH_SIZE) + + def event_handler(event): + if isinstance(event, fluid.EndIteration): + if (event.batch_id % 10) == 0: + avg_cost = trainer.test(reader=test_reader) + + print('BatchID {0:04}, Loss {1:2.2}'.format(event.batch_id + 1, + avg_cost)) + + if avg_cost > 0.01: # Low threshold for speeding up CI + trainer.save_params(save_path) + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + sgd_optimizer = fluid.optimizer.SGD( + learning_rate=fluid.layers.exponential_decay( + learning_rate=0.01, + decay_steps=100000, + decay_rate=0.5, + staircase=True)) + trainer = fluid.Trainer(train_network, optimizer=sgd_optimizer, place=place) + trainer.train(train_reader, EPOCH_NUM, event_handler=event_handler) + + +def infer(use_cuda, save_path): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + inference_program, param_path=save_path, place=place) + + def create_random_lodtensor(lod, place, low, high): + data = np.random.random_integers(low, high, + [lod[-1], 1]).astype("int64") + res = fluid.LoDTensor() + res.set(data, place) + res.set_lod([lod]) + return res + + # Create an input example + lod = [0, 4, 10] + word = create_random_lodtensor(lod, place, low=0, high=WORD_DICT_LEN - 1) + pred = create_random_lodtensor(lod, place, low=0, high=PRED_DICT_LEN - 1) + ctx_n2 = create_random_lodtensor(lod, place, low=0, high=WORD_DICT_LEN - 1) + ctx_n1 = create_random_lodtensor(lod, place, low=0, high=WORD_DICT_LEN - 1) + ctx_0 = create_random_lodtensor(lod, place, low=0, high=WORD_DICT_LEN - 1) + ctx_p1 = create_random_lodtensor(lod, place, low=0, high=WORD_DICT_LEN - 1) + ctx_p2 = create_random_lodtensor(lod, place, low=0, high=WORD_DICT_LEN - 1) + mark = create_random_lodtensor(lod, place, low=0, high=MARK_DICT_LEN - 1) + + results = inferencer.infer({ + 'word_data': word, + 'verb_data': pred, + 'ctx_n2_data': ctx_n2, + 'ctx_n1_data': ctx_n1, + 'ctx_0_data': ctx_0, + 'ctx_p1_data': ctx_p1, + 'ctx_p2_data': ctx_p2, + 'mark_data': mark + }) + + print("infer results: ", results) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + save_path = "label_semantic_roles.inference.model" + train(use_cuda, save_path) + infer(use_cuda, save_path) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py b/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py new file mode 100644 index 0000000000000000000000000000000000000000..a8282c71f895718930ea14a1e0bff777441c4c57 --- /dev/null +++ b/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py @@ -0,0 +1,118 @@ +# 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. +from __future__ import print_function +import argparse +import paddle.fluid as fluid +import paddle +import sys +import numpy +import unittest +import math +import sys +import os +import paddle.v2.dataset as dataset + +BATCH_SIZE = 64 + + +def inference_program(): + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + conv_pool_1 = fluid.layers.batch_norm(conv_pool_1) + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu") + prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') + return prediction + + +def train_program(): + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + predict = inference_program() + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + acc = fluid.layers.accuracy(input=predict, label=label) + return avg_cost, acc + + +def train(use_cuda, save_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + optimizer = fluid.optimizer.Adam(learning_rate=0.001) + trainer = fluid.Trainer(train_program, place=place, optimizer=optimizer) + + def event_handler(event): + if isinstance(event, fluid.EndIteration): + avg_cost, acc = event.values + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if (event.batch_id + 1) % 10 == 0: + test_metrics = trainer.test(reader=dataset.mnist.test()) + avg_cost_set = test_metrics[0] + acc_set = test_metrics[1] + + # get test acc and loss + acc = numpy.array(acc_set).mean() + avg_cost = numpy.array(avg_cost_set).mean() + if float(acc) > 0.2: # Smaller value to increase CI speed + trainer.save_params(save_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.batch_id + 1, float(avg_cost), float(acc))) + if math.isnan(float(avg_cost)): + sys.exit("got NaN loss, training failed.") + + trainer.train( + reader=dataset.mnist.train(), num_pass=100, event_handler=event_handler) + + +def infer(use_cuda, save_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + inferencer = fluid.Inferencer( + inference_program, param_path=save_dirname, place=place) + + batch_size = 1 + tensor_img = numpy.random.uniform(-1.0, 1.0, + [batch_size, 1, 28, 28]).astype("float32") + + results = inferencer.infer({'img': tensor_img}) + + print("infer results: ", results[0]) + + +def main(use_cuda): + save_dirname = "recognize_digits_conv.inference.model" + + # call train() with is_local argument to run distributed train + train(use_cuda=use_cuda, save_dirname=save_dirname) + infer(use_cuda=use_cuda, save_dirname=save_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py new file mode 100644 index 0000000000000000000000000000000000000000..3efa931d5886e565d2d876a31309883ee1660389 --- /dev/null +++ b/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py @@ -0,0 +1,105 @@ +# 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. +from __future__ import print_function +import argparse +import paddle.fluid as fluid +import paddle +import sys +import numpy +import unittest +import math +import sys +import os +import paddle.v2.dataset as dataset + +BATCH_SIZE = 64 + + +def inference_program(): + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + + hidden = fluid.layers.fc(input=img, size=200, act='tanh') + hidden = fluid.layers.fc(input=hidden, size=200, act='tanh') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + return prediction + + +def train_program(): + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + predict = inference_program() + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(cost) + acc = fluid.layers.accuracy(input=predict, label=label) + return avg_cost, acc + + +def train(use_cuda, save_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + optimizer = fluid.optimizer.Adam(learning_rate=0.001) + trainer = fluid.Trainer(train_program, place=place, optimizer=optimizer) + + def event_handler(event): + if isinstance(event, fluid.EndIteration): + avg_cost, acc = event.values + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if (event.batch_id + 1) % 10 == 0: + test_metrics = trainer.test(reader=dataset.mnist.test()) + avg_cost_set = test_metrics[0] + acc_set = test_metrics[1] + + # get test acc and loss + acc = numpy.array(acc_set).mean() + avg_cost = numpy.array(avg_cost_set).mean() + if float(acc) > 0.2: # Smaller value to increase CI speed + trainer.save_params(save_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.batch_id + 1, float(avg_cost), float(acc))) + if math.isnan(float(avg_cost)): + sys.exit("got NaN loss, training failed.") + + trainer.train( + reader=dataset.mnist.train(), num_pass=100, event_handler=event_handler) + + +def infer(use_cuda, save_dirname=None): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + inferencer = fluid.Inferencer( + inference_program, param_path=save_dirname, place=place) + + batch_size = 1 + tensor_img = numpy.random.uniform(-1.0, 1.0, + [batch_size, 1, 28, 28]).astype("float32") + + results = inferencer.infer({'img': tensor_img}) + + print("infer results: ", results[0]) + + +def main(use_cuda): + save_dirname = "recognize_digits_mlp.inference.model" + + # call train() with is_local argument to run distributed train + train(use_cuda=use_cuda, save_dirname=save_dirname) + infer(use_cuda=use_cuda, save_dirname=save_dirname) + + +if __name__ == '__main__': + for use_cuda in (False, True): + main(use_cuda=use_cuda) diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index a9fa2359e019d0301110dfc3573e2d2534c6a47a..d44cb16bfb1545fc840d1a38155ec407afd4473d 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -19,6 +19,7 @@ import executor import data_feeder import contextlib import io +import transpiler # optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module import optimizer as opt_module @@ -172,9 +173,9 @@ class Trainer(object): def save_params(self, param_path): # reference: save_persistables in io.py - exe = executor.Executor(self.place) - io.save_persistables( - exe, dirname=param_path, main_program=self.startup_program) + with self._prog_and_scope_guard(): + exe = executor.Executor(self.place) + io.save_persistables(exe, dirname=param_path) @staticmethod def _check_and_get_place(place): diff --git a/python/setup.py.in b/python/setup.py.in index a811b509a90b8b0d84451f54462a0308c062d022..c42601d335f01491156dc3591341c1a3213aecfe 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -68,7 +68,8 @@ packages=['paddle', 'paddle.fluid', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', - 'paddle.fluid.layers'] + 'paddle.fluid.layers', + 'paddle.fluid.transpiler'] if '${WITH_FLUID_ONLY}'== 'OFF': packages+=['paddle.proto',