CompiledProgram

class paddle.static. CompiledProgram ( program_or_graph, build_strategy=None ) [source]
Api_attr

Static Graph

The CompiledProgram is used to transform a program or graph for various optimizations according to the configuration of build_strategy, for example, the operators’ fusion in the computation graph, memory optimization during the execution of the computation graph, etc. For more information about build_strategy, please refer to paddle.static.BuildStrategy.

Parameters

  • program_or_graph (Graph|Program) – This argument is the Program or Graph being executed.

  • build_strategy (BuildStrategy) – This argument is used to compile the program or graph with the specified options, such as operators’ fusion in the computational graph and memory optimization during the execution of the computational graph. For more information about build_strategy, please refer to paddle.static.BuildStrategy. The default is None.

Returns

CompiledProgram

Example

import numpy
import paddle
import paddle.static as static

paddle.enable_static()

place = paddle.CUDAPlace(0) # paddle.CPUPlace()
exe = static.Executor(place)

data = static.data(name='X', shape=[None, 1], dtype='float32')
hidden = static.nn.fc(x=data, size=10)
loss = paddle.mean(hidden)
paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)

exe.run(static.default_startup_program())
compiled_prog = static.CompiledProgram(
    static.default_main_program())

x = numpy.random.random(size=(10, 1)).astype('float32')
loss_data, = exe.run(compiled_prog,
                    feed={"X": x},
                    fetch_list=[loss.name])
with_data_parallel ( loss_name=None, build_strategy=None, exec_strategy=None, share_vars_from=None, places=None )

with_data_parallel

This interface is used to transform the input Program or Graph to a multi-graph to run the model in data parallel mode. Users can use the build_strategy and exec_strategy to set some optimizations that can be applied during the construction and computation of the Graph, such as reducing the number of AllReduce operations, specifying the size of the thread pool used in the computation Graph running the model, and so on.

Note

If build_strategy is specified when building CompiledProgram and calling with_data_parallel, build_strategy in CompiledProgram will be overwritten, therefore, if it is data parallel training, it is recommended to set build_strategy when calling with_data_parallel interface.

Parameters

  • loss_name (str) – This parameter is the name of the loss Tensor of the model. Note: If it is model training, you must set loss_name, otherwise the result may be problematic. The default is None.

  • build_strategy (BuildStrategy) – This parameter is used to compile the program or graph with the specified options, such as operators’ fusion in the computational graph and memory optimization during the execution of the computational graph. For more information about build_strategy, please refer to fluid.BuildStrategy. The default is None.

  • exec_strategy (ExecutionStrategy) – exec_strategy specifies the options that can be changed when running the current model, such as the thread pool size. For more information about exec_strategy, please refer to fluid.ExecutionStrategy. The default is None.

  • share_vars_from (CompiledProgram) – If share_vars_from is set, the current CompiledProgram will share the parameter value with the CompiledProgram specified by share_vars_from. This parameter needs to be set when model testing is required during model training, and the data parallel mode is used for training and testing. Since CompiledProgram will only distribute parameter Tensors to other devices when it is first executed, the CompiledProgram specified by share_vars_from must be run before the current CompiledProgram. The default is None.

  • places (list(CUDAPlace)|list(CPUPlace)|list(str)|None) – This parameter specifies the device on which the model is running. If you want to run on GPU0 and GPU1, places are [fluid.CUDAPlace(0), fluid.CUDAPlace(1)]; if you want to run with 2 CPUs, places are [fluid.CPUPlace()] * 2. If the parameter is not set, i.e. the parameter is None, the available device will be obtained from the environment variable when the model is executed: If the GPU is used, the currently available device ID is obtained from the environment variable FLAGS_selected_gpus or CUDA_VISIBLE_DEVICES when the model is executed; CPU, when the model is executed, the currently available CPU number is obtained from the environment variable CPU_NUM. For example, export CPU_NUM=4, if the environment variable is not set, the executor will add the variable to the environment variable and set its value to 1. The default is None. If places is the list of string, the string in the list can be cpu, gpu:x, where x is the index of the GPUs.

Returns

CompiledProgram

Example

import numpy
import os
import paddle
import paddle.static as static

paddle.enable_static()

use_cuda = True
place = paddle.CUDAPlace(0) if use_cuda else paddle.CPUPlace()
parallel_places = [paddle.CUDAPlace(0), paddle.CUDAPlace(1)] if use_cuda else [paddle.CPUPlace()] * 2

# NOTE: If you use CPU to run the program, you need
# to specify the CPU_NUM, otherwise, paddle will use
# all the number of the logic core as the CPU_NUM,
# in that case, the batch size of the input should be
# greater than CPU_NUM, if not, the process will be
# failed by an exception.
if not use_cuda:
    os.environ['CPU_NUM'] = str(2)

exe = static.Executor(place)

data = static.data(name='X', shape=[None, 1], dtype='float32')
hidden = static.nn.fc(x=data, size=10)
loss = paddle.mean(hidden)

test_program = static.default_main_program().clone(for_test=True)
paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)

exe.run(static.default_startup_program())
compiled_train_prog = static.CompiledProgram(
    static.default_main_program()).with_data_parallel(
            loss_name=loss.name, places=parallel_places)
# NOTE: if not set share_vars_from=compiled_train_prog,
# the parameters used in test process are different with
# the parameters used by train process
compiled_test_prog = static.CompiledProgram(
    test_program).with_data_parallel(
            share_vars_from=compiled_train_prog,
            places=parallel_places)

train_data = numpy.random.random(size=(10, 1)).astype('float32')
loss_data, = exe.run(compiled_train_prog,
                feed={"X": train_data},
                fetch_list=[loss.name])
test_data = numpy.random.random(size=(10, 1)).astype('float32')
loss_data, = exe.run(compiled_test_prog,
                feed={"X": test_data},
                fetch_list=[loss.name])