def add_productions(self, grammar: GraphGrammar):
grammar.add("PreprocessingModule", "RecurrentModule")
grammar.add("RecurrentModule", Path("RecurrentCell", "RecurrentModule"))
grammar.add("RecurrentModule", "RecurrentCell")
grammar.add("RecurrentCell", Seq2SeqLSTM)
grammar.add("RecurrentCell", Seq2SeqBiLSTM)
def generate_grammar(*modules):
top_level = []
for mod in modules:
mod.make_top_level(top_level)
grammar = GraphGrammar(start=Path(*top_level), initializer=CfgInitializer())
for mod in modules:
mod.add_productions(grammar)
return grammar
def _build_grammar(self):
from autogoal.grammar._graph import Epsilon
return GraphGrammar(Epsilon())
def build_grammar(
preprocessing=False,
preprocessing_recurrent=True,
preprocessing_conv=False,
reduction=False,
reduction_recurrent=True,
reduction_conv=False,
features=False,
features_time_distributed=False,
):
modules = []
if preprocessing:
modules.append("PreprocessingModule")
if reduction:
modules.append("ReductionModule")
if features:
modules.append("FeaturesModule")
if features_time_distributed:
modules.append("FeaturesTimeDistributedModule")
if preprocessing_recurrent and preprocessing_conv:
raise ValueError(
"Cannot combine recurrent with convolutional preprocessing.")
if reduction_recurrent and reduction_conv:
raise ValueError(
"Cannot combine recurrent with convolutional reduction modules.")
if (reduction or features) and features_time_distributed:
raise ValueError(
"Cannot combine time-distributed modules with flat modules.")
if not modules:
raise ValueError("At least one module must be activated.")
grammar = GraphGrammar(start=Path(*modules), initializer=CfgInitializer())
if preprocessing_recurrent:
grammar.add("PreprocessingModule", "PreprocessingModuleR")
grammar.add("PreprocessingModuleR",
Path("Recurrent", "PreprocessingModuleR"))
grammar.add("PreprocessingModuleR", Epsilon())
grammar.add("Recurrent", Seq2SeqLSTM)
grammar.add("Recurrent", Seq2SeqBiLSTM)
if preprocessing_conv:
grammar.add("PreprocessingModule",
Path("Convolutional", "PreprocessingModuleC"))
grammar.add("PreprocessingModuleC",
Path("Convolutional", "PreprocessingModuleC"))
grammar.add("PreprocessingModuleC", Epsilon())
grammar.add("Convolutional", Path(Conv2D, MaxPooling2D))
grammar.add("Convolutional", Path(Conv2D, MaxPooling2D, Dropout))
if reduction_recurrent:
grammar.add("ReductionModule", Seq2VecLSTM)
grammar.add("ReductionModule", Seq2VecBiLSTM)
if reduction_conv:
grammar.add("ReductionModule", Flatten)
if features:
grammar.add("FeaturesModule", Path("DenseLayer", "FeaturesModule"))
grammar.add("FeaturesModule", Epsilon())
grammar.add("DenseLayer", Block(Dense, "DenseLayer"))
grammar.add("DenseLayer", Path(Dense, "DenseLayer"))
grammar.add("DenseLayer", Path(Dense, Dropout, "DenseLayer"))
grammar.add("DenseLayer", Epsilon())
if features_time_distributed:
grammar.add(
"FeaturesTimeDistributedModule",
Path("TDLayer", "FeaturesTimeDistributedModule"),
)
grammar.add("FeaturesTimeDistributedModule", Epsilon())
grammar.add("TDLayer", Block(TimeDistributed, "TDLayer"))
grammar.add("TDLayer", Path(TimeDistributed, "TDLayer"))
grammar.add("TDLayer", Epsilon())
return grammar
def add_productions(self, grammar: GraphGrammar):
grammar.add("PreprocessingModule",
Path("Conv2DModule", Flatten))
grammar.add("Conv2DModule", Path("Conv2DBlock",
"Conv2DModule"))
grammar.add("Conv2DModule", "Conv2DBlock")
grammar.add("Conv2DBlock", Path("Conv2DCells", MaxPooling2D))
grammar.add("Conv2DBlock",
Path("Conv2DCells", MaxPooling2D, Dropout))
grammar.add("Conv2DCells", Path("Conv2DCell", "Conv2DCells"))
grammar.add("Conv2DCells", Path("Conv2DCell"))
grammar.add("Conv2DCell",
Path(Conv2D, Activation, BatchNormalization))
grammar.add("Conv2DCell", Path(Conv2D, Activation))