# Copyright (c) 2016 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. """ Before this new package paddle.v2.layer, users would need to use functions in paddle.trainer_config_helpers.layers to configure networks. The Old Way: ========= This old way requires that the creation of a network be defined in a Python function, say network_config, and that this Python function being passed to paddle.trainer_config_helpers.parse_network_config for the creation of protobuf message description of this network. ```python def network_config(): img = paddle.trainer_config_helpers.data_layer(name="pixel", size=784) inference = paddle.trainer_config_helpers.fc_layer( input=img, size=10, act=paddle.trainer_config_helpers.SoftmaxActivation()) cost = paddle.trainer_config_helpers.classification_cost( input=inference, label=paddle.trainer_config_helpers.data_layer(name="label", size=10)) proto_desc = parse_network_config(network_config) ``` When parse_network_config executes network_config, those layer definition functions like data_layer and fc_layer would change some Python global variables, so that after the execution, parse_network_config could collect information from these global variables and generates the protobuf message. The New Way: ========= In this PR, we define a function in paddle.v2.layer which creates a Python class for each layer creation function in paddle.trainer_config_helpers.layers. Users can use create a network as follows: ```python img = paddle.v2.layer.data(name="pixel", size=784) inference = paddle.v2.layer.fc(input=img, size=10, act=paddle.v2.layer.Softmax()) cost = paddle.v2.layer.classification( input=inference, label=paddle.v2.layer.data(name="label", size=10)) parameters = paddle.v2.parameters.create(cost) ``` This new way doesn't require those invocations to layer definition functions to be in a Python function but could be anywhere. Also, the creation of a protobuf message is hidden in the invocation of paddle.v2.parameters.create, no longer exposed to users. """ import collections import inspect from config_base import Layer, __convert_to_v2__ import paddle.trainer_config_helpers as conf_helps from paddle.trainer_config_helpers.config_parser_utils import \ parse_network_config as __parse__ from paddle.trainer_config_helpers.default_decorators import wrap_act_default from paddle.trainer_config_helpers.default_decorators import \ wrap_bias_attr_default from paddle.trainer_config_helpers.default_decorators import wrap_name_default from paddle.trainer_config_helpers.layers import layer_support from paddle.trainer.config_parser import \ RecurrentLayerGroupWithoutOutLinksBegin, RecurrentLayerGroupSetOutLink, \ RecurrentLayerGroupEnd, model_type import activation import data_type __all__ = ['parse_network', 'data'] __projection_names__ = filter(lambda x: x.endswith('_projection'), dir(conf_helps)) __all__ += __projection_names__ __operator_names__ = filter(lambda x: x.endswith('_operator'), dir(conf_helps)) __all__ += __operator_names__ def parse_network(*outputs): """ parse all output layers and then generate a model config proto. :param outputs: :return: """ def __real_func__(): context = dict() real_output = [each.to_proto(context=context) for each in outputs] conf_helps.outputs(real_output) return __parse__(__real_func__) """ Some layer may need some special config, and can not use __convert_to_v2__ to convert. So we also need to implement some special LayerV2. """ class DataLayerV2(Layer): def __init__(self, name, type, **kwargs): assert isinstance(type, data_type.InputType) self.type = type self.__method_name__ = 'data_layer' self.__kwargs__ = kwargs super(DataLayerV2, self).__init__(name=name, parent_layers=dict()) def to_proto_impl(self, **kwargs): args = dict() args['size'] = self.type.dim for each in kwargs: args[each] = kwargs[each] for each in self.__kwargs__: args[each] = self.__kwargs__[each] return getattr(conf_helps, self.__method_name__)(name=self.name, **args) class WithExtraParent(Layer): def extra_parent(self): return self.__extra_parent__ def __init__(self, name=None, parent_layers=None): self.__extra_parent__ = [] super(WithExtraParent, self).__init__( name=name, parent_layers=parent_layers) def append_extra_parent(self, parent): self.__extra_parent__.append(parent) def to_proto(self, context): """ function to set proto attribute """ kwargs = dict() for p in self.__extra_parent__: p.to_proto(context=context) for layer_name in self.__parent_layers__: if not isinstance(self.__parent_layers__[layer_name], collections.Sequence): v1_layer = self.__parent_layers__[layer_name].to_proto( context=context) else: v1_layer = map(lambda x: x.to_proto(context=context), self.__parent_layers__[layer_name]) kwargs[layer_name] = v1_layer if self.context_name() is None: return self.to_proto_impl(context=context, **kwargs) elif self.context_name() not in context: context[self.context_name()] = self.to_proto_impl( context=context, **kwargs) if self.use_context_name(): return context[self.context_name()] else: return context[self.name] class MemoryV2(WithExtraParent): def __init__(self, name, **kwargs): self.name = name super(MemoryV2, self).__init__(name=name, parent_layers=dict()) self.__kwargs__ = kwargs self.__boot_layer_name__ = None if 'boot_layer' in kwargs: begin_of_current_rnn = [] # TODO(yuyang18): Fix inspect, it could be wrong when user invoke a # function inside step. st = inspect.stack() for i in xrange(len(st)): locs = inspect.stack()[i][0].f_locals keys = locs.keys() for key in keys: val = locs[key] if isinstance(val, RecurrentLayerInput): begin_of_current_rnn.append(val) elif isinstance(val, collections.Sequence): for v in val: if isinstance(v, RecurrentLayerInput): begin_of_current_rnn.append(v) if begin_of_current_rnn: break assert begin_of_current_rnn is not None for extra in begin_of_current_rnn: self.append_extra_parent(extra) assert isinstance(extra, WithExtraParent) extra.append_extra_parent(kwargs['boot_layer']) self.__boot_layer_name__ = kwargs['boot_layer'].name def to_proto_impl(self, context, **kwargs): args = dict() for each in kwargs: args[each] = kwargs[each] for each in self.__kwargs__: args[each] = self.__kwargs__[each] if self.__boot_layer_name__ is not None: args['boot_layer'] = context[self.__boot_layer_name__] size = args.get('size', None) if size is not None: if callable(size): real_size = size() else: real_size = size print(real_size) args['size'] = real_size return conf_helps.memory(name=self.name, **args) def context_name(self): return self.name + "#memory" def use_context_name(self): """ memory layer will have the same name with some layer :return: """ return True class LayerOutputV2(Layer): """ LayerOutputV2 is used to store the result of LayerOutput in v1 api. It will not store it's parents because layer_output has been parsed already. """ def __init__(self, layer_output): assert isinstance(layer_output, conf_helps.LayerOutput) self.layer_output = layer_output super(LayerOutputV2, self).__init__( name=layer_output.name, parent_layers=dict()) def to_proto_impl(self): return self.layer_output class StaticInputV2(object): def __init__(self, input, is_seq=False, size=None): assert isinstance(input, LayerV2) self.name = input.name self.input = input self.is_seq = is_seq self.size = size # TODO(qiaolongfei): add size # assert input.size is not None or size is not None class MixedLayerV2(Layer): """ This class is use to support `with` grammar. If not, the following code could convert mixed_layer simply. mixed = __convert_to_v2__( 'mixed_layer', name_prefix='mixed', parent_names=['input']) """ class AddToSealedMixedLayerExceptionV2(Exception): pass def __init__(self, size=0, input=None, name=None, act=None, bias_attr=None, layer_attr=None): self.__method_name__ = 'mixed_layer' self.finalized = False self.__inputs__ = [] if input is not None: self.__inputs__ = input other_kwargs = dict() other_kwargs['name'] = name other_kwargs['size'] = size other_kwargs['act'] = act other_kwargs['bias_attr'] = bias_attr other_kwargs['layer_attr'] = layer_attr parent_layers = {"input": self.__inputs__} super(MixedLayerV2, self).__init__(name, parent_layers) self.__other_kwargs__ = other_kwargs def __iadd__(self, other): if not self.finalized: self.__inputs__.append(other) return self else: raise MixedLayerV2.AddToSealedMixedLayerExceptionV2() def __enter__(self): assert len(self.__inputs__) == 0 return self def __exit__(self, *args, **kwargs): self.finalized = True def to_proto_impl(self, **kwargs): args = dict() for each in kwargs: args[each] = kwargs[each] for each in self.__other_kwargs__: args[each] = self.__other_kwargs__[each] size = args.get('size', None) if size is not None: if callable(size): real_size = size() else: real_size = size args['size'] = real_size return getattr(conf_helps, self.__method_name__)(**args) @wrap_name_default("mixed") @wrap_act_default(act=activation.Linear()) @wrap_bias_attr_default(has_bias=False) @layer_support(conf_helps.layers.ERROR_CLIPPING, conf_helps.layers.DROPOUT) def mixed(size=0, name=None, input=None, act=None, bias_attr=False, layer_attr=None): return MixedLayerV2(size, input, name, act, bias_attr, layer_attr) class RecurrentLayerInput(WithExtraParent): def __init__(self, recurrent_name, index, parent_layers): assert len(parent_layers) == 1 self.__parents__ = parent_layers.values()[0] super(RecurrentLayerInput, self).__init__( name=self.__parents__[index].name, parent_layers=parent_layers) self.__recurrent_name__ = recurrent_name def context_name(self): return self.__recurrent_name__ + ".begin" def to_proto_impl(self, context, **kwargs): model_type('recurrent_nn') RecurrentLayerGroupWithoutOutLinksBegin( name=self.__recurrent_name__, in_links=map(lambda x: x.name, self.__parents__)) return self class RecurrentLayerOutput(Layer): def __init__(self, recurrent_name, index, parent_layers): assert len(parent_layers) == 1 self.__parents__ = parent_layers.values()[0] super(RecurrentLayerOutput, self).__init__( name=self.__parents__[index].name, parent_layers=parent_layers) self.__recurrent_name__ = recurrent_name def context_name(self): return self.__recurrent_name__ + ".end" def to_proto_impl(self, **kwargs): for l in self.__parents__: RecurrentLayerGroupSetOutLink(l.name) RecurrentLayerGroupEnd(name=self.__recurrent_name__) LayerV2 = Layer data = DataLayerV2 AggregateLevel = conf_helps.layers.AggregateLevel ExpandLevel = conf_helps.layers.ExpandLevel memory = MemoryV2 def __layer_name_mapping__(inname): if inname in ['data_layer', 'memory', 'mixed_layer', 'recurrent_group']: # Do Not handle these layers return elif inname == 'maxid_layer': return 'max_id' elif inname.endswith('memory') or inname.endswith( '_seq') or inname.endswith('_sim') or inname == 'hsigmoid': return inname elif inname in [ 'cross_entropy', 'multi_binary_label_cross_entropy', 'cross_entropy_with_selfnorm' ]: return inname + "_cost" elif inname.endswith('_cost'): return inname elif inname.endswith("_layer"): return inname[:-len("_layer")] def __layer_name_mapping_parent_names__(inname): all_args = getattr(conf_helps, inname).argspec.args return filter( lambda x: x in ['input1', 'input2', 'label', 'input', 'a', 'b', 'expand_as', 'weights', 'vectors', 'weight', 'score', 'left', 'right', 'output_mem'], all_args) def __convert_layer__(_new_name_, _old_name_, _parent_names_): global __all__ __all__.append(_new_name_) globals()[new_name] = __convert_to_v2__(_old_name_, _parent_names_) for each_layer_name in dir(conf_helps): new_name = __layer_name_mapping__(each_layer_name) if new_name is not None: parent_names = __layer_name_mapping_parent_names__(each_layer_name) assert len(parent_names) != 0, each_layer_name __convert_layer__(new_name, each_layer_name, parent_names) del parent_names del new_name del each_layer_name # convert projection for prj in __projection_names__: globals()[prj] = __convert_to_v2__( prj, parent_names=['input'], is_default_name=False) # convert operator operator_list = [ # [V1_method_name, parent_names], ['dotmul_operator', ['a', 'b']], ['conv_operator', ['img', 'filter']] ] for op in operator_list: globals()[op[0]] = __convert_to_v2__( op[0], parent_names=op[1], is_default_name=False) @wrap_name_default() def recurrent_group(step, input, name=None): if not isinstance(input, collections.Sequence): input = [input] non_static_inputs = filter(lambda x: not isinstance(x, StaticInputV2), input) actual_input = [ RecurrentLayerInput( recurrent_name=name, index=i, parent_layers={'recurrent_inputs': non_static_inputs}) for i in xrange(len(non_static_inputs)) ] def __real_step__(*args): rnn_input = list(args) static_inputs = filter(lambda x: isinstance(x, StaticInputV2), input) for static_input in static_inputs: mem_name = "__%s_memory__" % static_input.input.name print memory mem = memory( name=mem_name, is_seq=static_input.is_seq, size=static_input.input.calculate_size, boot_layer=static_input.input) with mixed( name=mem_name, size=static_input.input.calculate_size, act=activation.Identity()) as mix: mix += identity_projection(input=mem) rnn_input.insert(input.index(static_input), mix) return step(*rnn_input) actual_output = __real_step__(*actual_input) if not isinstance(actual_output, collections.Sequence): actual_output = [actual_output] retv = [ RecurrentLayerOutput( recurrent_name=name, index=i, parent_layers={'recurrent_outputs': actual_output}) for i in xrange(len(actual_output)) ] if len(retv) == 1: return retv[0] else: return retv