diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index 5af5bc9c4731317075b3912a4749a0b358bdd56e..45af83708ea63fc1b6aa86f1e8423bb44b7388a6 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -44,7 +44,7 @@ import metrics import transpiler from param_attr import ParamAttr, WeightNormParamAttr from data_feeder import DataFeeder -from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace +from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope from transpiler import DistributeTranspiler, InferenceTranspiler, \ memory_optimize, release_memory from concurrency import (Go, make_channel, channel_send, channel_recv, @@ -83,6 +83,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \ 'profiler', 'unique_name', 'recordio_writer', + 'Scope', ] diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 81e94096144f58df51315c5a7f597925dda735c3..159b0ca39eed547e4f3448e7ebf4807299d465b2 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -25,6 +25,13 @@ g_scope = core.Scope() def global_scope(): + """ + Get the global/default scope instance. There are a lot of APIs use + :code:`global_scope` as its default value, e.g., :code:`Executor.run` + + Returns: + Scope: The global/default scope instance. + """ return g_scope @@ -37,6 +44,19 @@ def switch_scope(scope): @contextlib.contextmanager def scope_guard(scope): + """ + Change the global/default scope instance by Python `with` statement. All + variable in runtime will assigned to the new scope. + + Examples: + >>> import paddle.fluid as fluid + >>> new_scope = fluid.Scope() + >>> with fluid.scope_guard(new_scope): + >>> ... + + Args: + scope: The new global/default scope. + """ ex = switch_scope(scope) yield switch_scope(ex) diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index ff19a88ec475ecb8f76119910960edf84bba578c..db21b1f3c03c40d79084b0dbb57d22f6d41fa270 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -30,8 +30,6 @@ __all__ = [ 'default_startup_program', 'default_main_program', 'program_guard', - 'switch_startup_program', - 'switch_main_program', 'get_var', ] @@ -1238,23 +1236,18 @@ class Program(object): Notes: we have default_startup_program and default_main_program by default, a pair of them will shared the parameters. The default_startup_program only run once to initialize parameters, - default_main_program run in every minibatch and adjust the weights. - - Args: - None + default_main_program run in every mini batch and adjust the weights. Returns: - Python Program + A empty program. Examples: - .. code-block:: python - - main_program = Program() - startup_program = Program() - with fluid.program_guard(main_program=main_program, startup_program=startup_program): - fluid.layers.data(name="x", shape=[-1, 784], dtype='float32') - fluid.layers.data(name="y", shape=[-1, 1], dtype='int32') - fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu") + >>> main_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(main_program=main_program, startup_program=startup_program): + >>> fluid.layers.data(name="x", shape=[-1, 784], dtype='float32') + >>> fluid.layers.data(name="y", shape=[-1, 1], dtype='int32') + >>> fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu") """ @@ -1268,6 +1261,19 @@ class Program(object): @property def op_role(self): + """ + The operator role. In a enum {Forward, Backward, Optimize}. + + Notes: this is a low level API. It is used only for ParallelExecutor to + duplicate or schedule operator to devices. + + For example, the forward operator should be executed on every device. + The backward operator should be executed on every device and the + parameter gradient of backward (use :code:`op_role_var` to get this + variable) operator should be merged to one device. The optimization + operators should be executed on only one device and broadcast the + optimization result, i.e., the new parameter, to every other device. + """ return self._current_role @op_role.setter @@ -1276,6 +1282,13 @@ class Program(object): @property def op_role_var(self): + """ + The auxiliary variables for :code:`op_role` property. + + See Also: :code:`Program.op_role`'s documentation for details. + + Notes: This is a very low-level API. Users should not use it directly. + """ return self._op_role_var @op_role_var.setter @@ -1284,6 +1297,21 @@ class Program(object): @contextlib.contextmanager def optimized_guard(self, var): + """ + A with guard to set :code:`Optimization` :code:`OpRole` and + :code:`OpRoleVar` automatically. + + Notes: This is a very low level API. Users should not use it directly. + + Args: + var(Variable|str): The variable (name) to be optimized. + + Examples: + + >>> p, g = backward(...) + >>> with program.optimized_guard(p): + >>> p = p - 0.001 * g + """ OpRole = core.op_proto_and_checker_maker.OpRole self._current_role = OpRole.Optimize self._op_role_var = [var.name if isinstance(var, Variable) else var] @@ -1292,18 +1320,35 @@ class Program(object): self._current_role = OpRole.Forward def __str__(self): + """ + Get the protobuf debug string of this Program. + + Returns: + (str): The protobuf debug string. + + Raises: + ValueError: If any of required fields is not set. + """ return self.to_string(True) def to_string(self, throw_on_error, with_details=False): """ To debug string. + Args: - throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True - with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True + throw_on_error(bool): raise Value error when any of required fields + is not set. - Returns(str): The debug string. + with_details(bool): True if more details about variables and + parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need + to print. + + Returns + (str): The debug string. + + Raises: + ValueError: If any of required fields is not set and throw_on_error is + True. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, @@ -1319,25 +1364,89 @@ class Program(object): return res_str def get_desc(self): + """ + Get the C++ side of `ProgramDesc` object pointer. The C++ object is + exposed by :code:`pybind`. + + Notes: This is a very low level API. Users should not use this API + directly. + """ return self.desc def clone(self, for_test=False): - """Clone the Program object - Args: - for_test(bool): indicate whether clone for test. + """ + Create a new, duplicated program. + - Set for_test to False when we want to clone the program for training. - Set for_test to True when we want to clone the program for testing. + Some operators, e.g., :code:`batch_norm`, behave differently between + training and testing. They have an attribute, :code:`is_test`, to + control this behaviour. This method will change the :code:`is_test` + attribute of them to :code:`True` when :code:`for_test=True`. + + * Set for_test to False when we want to clone the program for training. + * Set for_test to True when we want to clone the program for testing. + + Notes: This API DOES NOT prune any operator. Use + :code:`clone(for_test=True)` before backward and optimization please. Args: - for_test(bool): Some operators, such as batch_norm and drop_out ops, - behave differently in training and testing. If for_test is True, - the is_test attributes in these operators will be set to True for - testing purposes, otherwise, they remain unchanged. + for_test(bool): True if change the :code:`is_test` attribute of + operators to :code:`True`. Returns: - Program: The cloned Program object. - + Program: The new, duplicated Program object. + + Examples: + + 1. To clone a test program, the sample code is: + + >>> import paddle.fluid as fluid + >>> train_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(train_program, startup_program): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> hidden = fluid.layers.fc(input=img, size=200, act='relu') + >>> hidden = fluid.layers.dropout(hidden, dropout_prob=0.5) + >>> loss = fluid.layers.cross_entropy( + >>> input=fluid.layers.fc(hidden, size=10, act='softmax'), + >>> label=fluid.layers.data(name='label', shape=[1], dtype='int64')) + >>> + >>> test_program = train_program.clone(for_test=True) + >>> + >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3) + >>> with fluid.program_guard(train_program, startup_program): + >>> sgd.minimize(loss) + + 2. The :code:`clone` method can be avoid if you create program for + training and program for testing individually. + + >>> import paddle.fluid as fluid + >>> + >>> def network(is_test): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> hidden = fluid.layers.fc(input=img, size=200, act='relu') + >>> hidden = fluid.layers.dropout(hidden, dropout_prob=0.5, is_test=is_test) + >>> loss = fluid.layers.cross_entropy( + >>> input=fluid.layers.fc(hidden, size=10, act='softmax'), + >>> label=fluid.layers.data(name='label', shape=[1], dtype='int64')) + >>> return loss + >>> + >>> train_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> test_program = fluid.Program() + >>> + >>> with fluid.program_guard(train_program, startup_program): + >>> with fluid.unique_name.guard(): + >>> loss = network(is_test=False) + >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3) + >>> sgd.minimize(loss) + >>> + >>> # the test startup program is not used. + >>> with fluid.program_guard(test_program, fluid.Program()): + >>> with fluid.unique_name.guard(): + >>> loss = network(is_test=True) + + The two code snippets above will generate same programs. """ if for_test: p = self.inference_optimize() @@ -1352,6 +1461,21 @@ class Program(object): return p def prune(self, targets): + """ + Prune operators and variables which are not needed to generate + :code:`targets`. + + Notes: This is a very low level API. Users should not use this API + directly. This API is in flux and not stable. + + Args: + targets(list|Variable|Operator): A list of variables or operators + need to be pruned + + Returns: + Program: A new, pruned program. + + """ if not isinstance(targets, list): targets = [targets] targets_idx = [] @@ -1386,6 +1510,17 @@ class Program(object): return res def inference_optimize(self): + """ + This method will create a new program and change the :code:`is_test` + attribute of operators to :code:`True`. All the :code:`Parameter` + information will be lost. + + Notes: This API is a very low level API. Use + :code:`Program.clone(for_test=True)` instead. + + Returns: + Program: The new program. + """ # this is an alternative implement before # core.inference_optimize being fixed. res = Program() @@ -1402,6 +1537,18 @@ class Program(object): @staticmethod def parse_from_string(binary_str): + """ + Deserialize a program desc from protobuf binary string. + + Notes: All information about parameters will be lost after serialization + and deserialization. + + Args: + binary_str(str): The binary prootbuf string. + + Returns: + Program: A deserialized program desc. + """ p = Program() p.desc = core.ProgramDesc(binary_str) p.blocks = [Block(p, i) for i in xrange(p.desc.num_blocks())] @@ -1410,10 +1557,19 @@ class Program(object): @property def random_seed(self): + """ + The default random seed for random operators in Program. Zero means get + the random seed from random device. + + Notes: It must be set before the operators have been added. + """ return self._seed @property def num_blocks(self): + """ + The number of blocks in this program. + """ return self.desc.num_blocks() @random_seed.setter @@ -1426,15 +1582,40 @@ class Program(object): return str(self) def global_block(self): + """ + Get the first block of this program. + """ return self.blocks[0] def block(self, index): + """ + Get the :code:`index` block of this program + Args: + index(int): The index of block to get + + Returns: + Block: The :code:`index` block + """ return self.blocks[index] def current_block(self): + """ + Get the current block. The :code:`current` block is the block to append + operators. + """ return self.blocks[self.current_block_idx] def create_block(self, parent_idx=None): + """ + Create a new block with the :code:`parent_idx` and change the current block + to new block. + + Args: + parent_idx(int): The parent block index. + + Returns: + Block: The new block. + """ new_block_idx = len(self.blocks) parent = self.current_block() if parent_idx is None else self.block( parent_idx) @@ -1444,9 +1625,24 @@ class Program(object): return self.current_block() def rollback(self): + """ + Exit a code block, i.e., roll back to the parent block. + Returns: + None + """ self.current_block_idx = self.current_block().parent_idx def sync_with_cpp(self): + """ + Synchronize Python instance to its binding C++ object instance. + If the program is modified in C++ space, this method should be invoked. + + Notes: This is a very low level API. Users should not invoke it + directly. + + Returns: + None + """ for block_idx in range(len(self.blocks), self.desc.num_blocks()): self.blocks.append(Block(self, block_idx)) for block in self.blocks: @@ -1456,6 +1652,9 @@ class Program(object): """ Copy the information of parameters from other program. + Notes: This is a very low level API. Users should not invoke it + directly. + Args: other(Program): Other program @@ -1475,6 +1674,9 @@ class Program(object): """ Copy the information of data variables from other program. + Notes: This is a very low level API. Users should not invoke it + directly. + Args: other(Program): Other program @@ -1493,6 +1695,12 @@ class Program(object): self.global_block().var(var.name).is_data = True def list_vars(self): + """ + Get all variables from this Program. A iterable object is returned. + + Returns: + iterable: The generator will yield every variable in this program. + """ for each_block in self.blocks: for each_var in each_block.vars.itervalues(): yield each_var @@ -1584,8 +1792,15 @@ _startup_program_ = Program() def default_startup_program(): """ - Get default startup program. In startup program, Paddle will initialize - parameters, initialize nccl handle, etc. + Get default/global startup program. + + The layer function in :code:`fluid.layers` will create parameters, readers, + NCCL handles as global variables. The :code:`startup_program` will + initialize them by the operators in startup program. The layer function will + append these initialization operators into startup program. + + This method will return the :code:`default` or the :code:`current` startup + program. Users can use :code:`fluid.program_guard` to switch program. Returns: Program: startup program @@ -1595,7 +1810,15 @@ def default_startup_program(): def default_main_program(): """ - Get default main program. The main program is used for training or testing. + Get default/global main program. The main program is used for training or + testing. + + All layer function in :code:`fluid.layers` will append operators and + variables to the :code:`default_main_program`. + + The :code:`default_main_program` is the default program in a lot of APIs. + For example, the :code:`Executor.run()` will execute the + :code:`default_main_program` when the program is not specified. Returns: Program: main program @@ -1637,20 +1860,34 @@ def switch_startup_program(program): @contextlib.contextmanager def program_guard(main_program, startup_program=None): """ - Switch program with `with` statement + Change the global main program and startup program with `with` statement. + Layer functions in the Python `with` block will append operators and + variables to the new main programs. Examples: - >>> with program_guard(Program()): - >>> data = fluid.layers.data(...) - >>> hidden = fluid.layers.fc(...) + + >>> import paddle.fluid as fluid + >>> main_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(main_program, startup_program): + >>> data = fluid.layers.data(...) + >>> hidden = fluid.layers.fc(...) + + Notes: The temporary :code:`Program` can be used if the user does not need + to construct either of startup program or main program. + + Examples: + + >>> import paddle.fluid as fluid + >>> main_program = fluid.Program() + >>> # does not care about startup program. Just pass a temporary value. + >>> with fluid.program_guard(main_program, fluid.Program()): + >>> data = ... Args: - main_program(Program): New main program inside `with` statement + main_program(Program): New main program inside `with` statement. startup_program(Program): New startup program inside `with` statement. None means do not change startup program. - - Returns: - None """ if not isinstance(main_program, Program): raise TypeError("main_program should be Program") @@ -1667,7 +1904,8 @@ def program_guard(main_program, startup_program=None): def get_var(name, program=None): """ - Get a variable by name from the global block of a program + Get a variable by name from the global block of a program. + Args: name(str): name of the variable program(Program|None): program object. diff --git a/python/paddle/fluid/lod_tensor.py b/python/paddle/fluid/lod_tensor.py index 61be39c25912604f842ef8a9a6ec5f0d1cf70257..c417ab393fca88d476d2f1fe83d12f99271d6883 100644 --- a/python/paddle/fluid/lod_tensor.py +++ b/python/paddle/fluid/lod_tensor.py @@ -19,33 +19,41 @@ __all__ = ['create_lod_tensor', 'create_random_int_lodtensor'] def create_lod_tensor(data, lod, place): - """Create a lod tensor from a numpy array, a list, or an existing lod tensor. + """ + Create a lod tensor from a numpy array, a list, or an existing lod tensor. Create a lod tensor by doing the following: + 1. Check that the length-based input lod is valid. + 2. Convert the length-based lod to a offset-based LoD. - 3. Copy the data from a numpy array, a list or a existing lod tensor to + + 3. Copy the data from a numpy array, a list or a existing lod tensor to CPU or GPU device (based on input place). + 4. Set the level of detail (LoD) using the offset-based LoD. - Use example: - Suppose we want LoDTensor to hold data for sequences of word, where each word is - represented by an integer. If we want to create a LoDTensor to represent two - sentences, one of 2 words, and one of 3 words. + Examples: - Then 'data' can be a numpy array of integers with shape (5, 1). - 'lod' will be [[2, 3]], indicating the length(# of words) in each sentence. - This length-based input lod [[2, 3]] will be converted to offset-based lod [[0, 2, 5]] - inside the function call. + Suppose we want LoDTensor to hold data for sequences of word, where each + word is represented by an integer. If we want to create a LoDTensor to + represent two sentences, one of 2 words, and one of 3 words. - Please refer to - github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md - for more details regarding LoD. + Then :code:`data` can be a numpy array of integers with shape (5, 1). + :code:`lod` will be [[2, 3]], indicating the length(# of words) in each + sentence. This length-based input lod [[2, 3]] will be converted to + offset-based lod [[0, 2, 5]] inside the function call. + + Please reference :ref:`api_guide_low_level_lod_tensor` for more details + regarding LoD. Args: - data: a numpy array or a LoDTensor or a list holding the data to be copied. - lod: a list of lists indicating the length-based LoD info specified by the user. - place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. + data(numpy.ndarray|list|LoDTensor): a numpy array or a LoDTensor or a + list holding the data to be copied. + lod(list): a list of lists indicating the length-based LoD info + specified by the user. + place(Place): CPU or GPU place indicating where the data in the new + LoDTensor will be stored. Returns: A fluid LoDTensor object with tensor data and lod info. @@ -77,31 +85,38 @@ def create_lod_tensor(data, lod, place): def create_random_int_lodtensor(lod, base_shape, place, low, high): - """Create a LoDTensor containing random integers. + """ + Create a LoDTensor containing random integers. - This function is frequently used in the book examples. So we revised it based on - the new create_lod_tensor API and put it here in the lod_tensor module to simplify - the code. + This function is frequently used in the book examples. So we revised it + based on the new create_lod_tensor API and put it here in the lod_tensor + module to simplify the code. The function does the following: - 1. Calculate the overall shape of the LoDTensor based on the length-based 'lod' input - and the shape of the basic element in 'base_shape'. + + 1. Calculate the overall shape of the LoDTensor based on the length-based + :code:`lod` input and the shape of the basic element in + :code:`base_shape`. + 2. Create a numpy array of this shape. + 3. Create the LoDTensor using create_lod_tensor API. - Suppose we want LoDTensor to hold data for sequences of word, where each word is - represented by an integer. If we want to create a LoDTensor to represent two - sentences, one of 2 words, and one of 3 words. Then 'base_shape' is [1], input - length-based 'lod' is [[2, 3]]. Then the overall shape of the LoDTensor would be - [5, 1], holding 5 words for two sentences. + Suppose we want LoDTensor to hold data for sequences of word, where each + word is represented by an integer. If we want to create a LoDTensor to + represent two sentences, one of 2 words, and one of 3 words. Then + 'base_shape' is [1], input length-based 'lod' is [[2, 3]]. Then the overall + shape of the LoDTensor would be [5, 1], holding 5 words for two sentences. Args: - data: a numpy array or a LoDTensor holding the data to be copied. - lod: a list of lists indicating the length-based LoD info specified by the user. - base_shape: the shape of the basic element to be held by the LoDTensor. - place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. - low: the lower bound of the random integers. - high: the upper bound of the random integers. + lod(list): a list of lists indicating the length-based LoD info + specified by the user. + base_shape(list): the shape of the basic element to be held by the + LoDTensor. + place(Place): CPU or GPU place indicating where the data in the new + LoDTensor will be stored. + low(int): the lower bound of the random integers. + high(int): the upper bound of the random integers. Returns: A fluid LoDTensor object with tensor data and lod info. diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index 8d48e9abef0fb9861284c6302b30efb0e3994989..bd57772713057f12b876942de58ee43527e94834 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -36,6 +36,45 @@ def convert_reader_to_recordio_file( compressor=core.RecordIOWriter.Compressor.Snappy, max_num_records=1000, feed_order=None): + """ + Convert a Python Reader to a recordio file. + + Please see :ref:`api_guide_python_reader` and :ref:`api_guide_reader_op` for + details. + + Examples: + + >>> import paddle.fluid as fluid + >>> import paddle.dataset.mnist as mnist + >>> import paddle + >>> + >>> tmp_program = fluid.Program() + >>> with fluid.program_guard(tmp_program): + >>> img = fluid.layers.data(name='img', shape=[784]) + >>> label = fluid.layers.data(name='label', shape=[1], dtype='int64') + >>> feeder = fluid.DataFeeder(feed_list=[img, label], place=fluid.CPUPlace()) + >>> # mnist.recordio will be generated in current directory + >>> fluid.recordio_writer.convert_reader_to_recordio_file( + >>> filename="mnist.recordio", + >>> reader_creator=paddle.batch(mnist.train(), batch_size=32), + >>> feeder=feeder) + + Args: + filename(str): The recordio filename. + reader_creator(callable): The Python Reader Creator. See + :ref:`api_guide_python_reader`. + feeder(DataFeeder): The DataFeeder instance. Used to convert + :code:`reader_creator` to :code: `lod_tensor` + compressor: Must in fluid.core.RecordIOWriter.Compressor.Snappy or + fluid.core.RecordIOWriter.Compressor.NoCompress. Use :code:`Snappy` + by default. + max_num_records(int): Maximum number of records in one chuck. Each record + is each return value from reader function + feed_order(list): The order of variable names that the reader returns + + Returns: + int: the number of record that saved. + """ if feed_order is None: feed_order = feeder.feed_names counter = 0 @@ -58,6 +97,17 @@ def convert_reader_to_recordio_files( compressor=core.RecordIOWriter.Compressor.Snappy, max_num_records=1000, feed_order=None): + """ + convert a python reader to many recordio files. + + This API is basically same as :code:`convert_reader_to_recordio_file`, + instead of it will create many recordio files. Each file contains at + most :code:`batch_per_file` records. + + Please reference + :ref:`api_fluid_recordio_writer_convert_reader_to_recordio_file` for more + details. + """ if feed_order is None: feed_order = feeder.feed_names f_name, f_ext = os.path.splitext(filename) diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index efc28d899304b01a3085891f3ae9396d57c589a1..45ab889beaa1355d0e1e2922aedf0340f70809ba 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -33,23 +33,59 @@ __all__ = [ class BeginEpochEvent(object): + """ + The begin of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + """ + def __init__(self, epoch_id): self.epoch = epoch_id class EndEpochEvent(object): + """ + The end of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + """ + def __init__(self, epoch_id): self.epoch = epoch_id class BeginStepEvent(object): + """ + The begin of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + step_id(int): The current step ID. + """ + def __init__(self, epoch_id, step_id): self.epoch = epoch_id self.step = step_id self.fetch_metrics = True + """ + If fetch_metrics is true, the metrics will be fetched at the + EndStepEvent. Default is True. + """ class EndStepEvent(object): + """ + The end of a training step. + + Args: + epoch_id(int): The current epoch ID. + step_id(int): The current step ID. + metrics(list): A list of fetched tensor. The order of this list is same + as the :code:`train_func` returns. + """ + def __init__(self, epoch_id, step_id, metrics): self.epoch = epoch_id self.step = step_id @@ -57,6 +93,27 @@ class EndStepEvent(object): class CheckpointConfig(object): + """ + Parameter object for :code:`fluid.io.save_checkpoint` and + :code:`fluid.Trainer`. Used to configuration how to save checkpoint. + + Args: + checkpoint_dir(str): Directory path to save check point. Default is the + current directory. + + max_num_checkpoints(int): The max number of local check points. + epoch_interval(int): Every number of epoch to save check point. + step_interval(int): Every number of step to save check point. + + Examples: + >>> config = fluid.CheckpointConfig("./checkpoints") + >>> trainer = fluid.Trainer(train_func=train_program, + >>> place=place, + >>> optimizer_func=optimizer_func, + >>> checkpoint_config=config) + >>> trainer.train(...) + """ + def __init__(self, checkpoint_dir=None, max_num_checkpoints=3, @@ -113,11 +170,62 @@ def check_and_get_place(place): class Trainer(object): """ + A trainer wraps MultiGPU/MultiNode training loops and can be used to train a + simple neural network easily. + + This API takes a :code:`train_func`. A :code:`train_func` is a function that + return loss as it first return value. The reset value can be fetched by + EndStepEvent.metrics + + This API also takes a :code:`optimizer_func` that will return an optimizer + instance. + + For example, to train a MLP for MNIST dataset, the sample program is + + >>> import paddle.fluid as fluid + >>> + >>> def mlp(image, layer_sizes=[200, 100], activation="relu", num_classes=10): + >>> hidden = image + >>> for layer_size in layer_sizes: + >>> hidden = fluid.layers.fc(input=hidden, size=layer_size, act=activation) + >>> return fluid.layers.fc(input=hidden, size=num_classes, act="softmax") + >>> + >>> def train_mnist_mlp(): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> label = fluid.layers.data(name='label', shape=[1], dtype='int64') + >>> prediction = mlp(img) + >>> return fluid.layers.mean(fluid.layers.cross_entropy(prediction, label)) + >>> + >>> def optimizer(): + >>> return fluid.optimizer.Adam() + >>> + >>> trainer = Trainer(train_func=train_mnist_mlp, + >>> optimizer_func=optimizer, + >>> place=fluid.CUDAPlace(0), + >>> parallel=True) + >>> + >>> def train_callback(event): + >>> if isinstance(event, fluid.EndStepEvent): + >>> print "Epoch ID", event.epoch, "Step ID",\ + >>> event.step, "AvgLoss", event.metrics[0] + >>> elif isinstance(event, fluid.EndEpochEvent): + >>> trainer.save_params("./model_{0}".format(event.epoch)) + >>> + >>> trainer.train(num_epochs=100, event_handler=train_callback) + + For more example, please see :ref:`api_guide_high_level_api`. + Args: - train_func(callable): A function which will return loss. The loss must be a scalar. + train_func(callable): A function which will return loss. The loss must be + a scalar tensor. optimizer_func(callable): A function that returns an Optimizer object. - place: The device place of this trainer. + place(CUDAPlace|CPUPlace): The device place of this trainer. If + :code:`parallel=True,` all CUDA Places will be used if :code:`place` + is a :code:`CUDAPlace`. + parallel(bool): True if use multiple devices. + checkpoint_config(CheckpointConfig): Configuration about how to save + checkpoints. """ def __init__(self, @@ -129,9 +237,6 @@ class Trainer(object): checkpoint_config=None): self.__stop = False self.parallel = parallel - # 1. we need to generate a framework.Program by calling - # program_func. Reference: fluid.program_guard in - # test_word2vec.py # config for checkpoint # only chief worker will save variables @@ -145,6 +250,10 @@ class Trainer(object): self.scope = core.Scope() + # 1. we need to generate a framework.Program by calling + # program_func. Reference: fluid.program_guard in + # test_word2vec.py + self.startup_program = framework.Program() self.train_program = framework.Program() @@ -277,17 +386,18 @@ class Trainer(object): def train(self, num_epochs, event_handler, reader=None, feed_order=None): """ - Train the model. + Start the train loop to train the model. Args: - num_epochs: The number of epoch. An epoch will process all data in reader - event_handler: The event handler. A function with type (ev:Event)->void - reader: - feed_order: Feeding order of reader. None will following the defining + num_epochs(int): The number of epoch. An epoch will process all data in reader + event_handler(callable): The event handler. A function with type (ev:Event)->void + reader(callable): A reader creator object. See also + :ref:`api_guide_python_reader` . + feed_order(list): Feeding order of reader. None will following the defining order in program Returns: - + None """ training_role = os.getenv("PADDLE_TRAINING_ROLE", "") if training_role == "PSERVER": @@ -307,16 +417,24 @@ class Trainer(object): Test the model on given test data Args: - reader: The reader that yields test data. - feed_order: Feeding order of reader. None will following the defining - order in program + reader(callable): The reader that yields test data. + feed_order(list): Feeding order of reader. None will following the + defining order in program """ return self._test_by_executor(reader, feed_order, self.train_func_outputs) def save_params(self, param_path): - # reference: save_persistables in io.py + """ + Save all parameters into :code:`param_path`. + + Args: + param_path(str): The path to save parameters. + + Returns: + None + """ with self._prog_and_scope_guard(): exe = executor.Executor(self.place) io.save_persistables(exe, dirname=param_path) diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 8bfb554845d9b128f000d6c90cf626416a198eef..999ef43ca0feacbddff5f9db59589ce7097fe77e 100644 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -383,6 +383,16 @@ def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0): def release_memory(input_program, skip_opt_set=None): + """ + Modify the input program and insert :code:`delete_op` to early drop not used + variables. The modification will be performed inplace. + + Notes: This is an experimental API and could be removed in next few + releases. Users should not use this API. + + Args: + input_program(Program): The program will be inserted :code:`delete_op`. + """ cfgs = _get_cfgs(input_program) for cfg in cfgs: cfg.release_memory(skip_opt_set=skip_opt_set) diff --git a/python/paddle/fluid/unique_name.py b/python/paddle/fluid/unique_name.py index 33c53113ae7e8ed9aeada31f2aed6990b6fea110..776619cd36722e338a9fdd5e13bceeaf3724de2c 100644 --- a/python/paddle/fluid/unique_name.py +++ b/python/paddle/fluid/unique_name.py @@ -16,7 +16,7 @@ import collections import contextlib import sys -__all__ = ['generate', 'switch', 'guard', 'UniqueNameGenerator'] +__all__ = ['generate', 'switch', 'guard'] class UniqueNameGenerator(object):