def main(args):
#
jsonNs = JsonArgParser(ns.PARAMETERS).parse(
args.negative_sampling_json_base)
jsonNs["lr"] = args.lr_ns
jsonNs["noise_rate"] = args.noise_rate
jsonNs["num_epochs"] = args.num_epochs
jsonNs["t"] = args.t
jsonNs["power"] = args.power
jsonNs["min_count"] = args.min_count
jsonNs["save_model"] = args.save_model
jsonWnn = JsonArgParser(wt.WNN_PARAMETERS).parse(
args.wnn_with_target_json_base)
jsonWnn["lr"] = args.lr_wnn
jsonWnn["target_module"] = args.save_model
log = logging.getLogger(__name__)
log.info("Starting unsupervised training")
ns.mainWnnNegativeSampling(dict2obj(jsonNs))
log.info("Starting supervised training")
wt.mainWnn(dict2obj(jsonWnn))
def main():
nArgs = len(sys.argv)
if nArgs != 2 and nArgs != 4:
sys.stderr.write('Syntax error! Expected arguments: <params_file> [<num_folds> <params_dist>]\n')
sys.stderr.write('\t<params_file>: JSON-formatted file containing parameter values\n')
sys.stderr.write('\t<num_folds>: number of cross-validation folds used in random search\n')
sys.stderr.write('\t<params_dist>: JSON-formatted file containing the parameter distributions used in random '
'search\n')
sys.exit(1)
full_path = os.path.realpath(__file__)
path = os.path.split(full_path)[0]
logging.config.fileConfig(os.path.join(path, 'logging.conf'))
parameters = dict2obj(JsonArgParser(WNN_PARAMETERS).parse(sys.argv[1]))
mainWnnNer(parameters)
callbacks = []
# log.info("Usando lambda fixo: " + str(_lambda.get_value()))
log.info("Usando lambda variado. alpha=" + str(kwargs["alpha"]) +
" height=" + str(kwargs["height"]))
callbacks.append(
ChangeLambda(_lambda, kwargs["alpha"], numEpochs, kwargs["height"]))
if kwargs["additional_dev"]:
callbacks.append(
AdditionalDevDataset(model, kwargs["additional_dev"],
kwargs["token_label_separator"],
inputGenerators, outputGeneratorTag))
# Training Model
model.train([trainSupervisedBatch, trainUnsupervisedDatasetBatch],
numEpochs,
devReader,
callbacks=callbacks)
if __name__ == '__main__':
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'))
parameters = JsonArgParser(UNSUPERVISED_BACKPROPAGATION_PARAMETERS).parse(
sys.argv[1])
main(**parameters)
def main():
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'), defaults={})
log = logging.getLogger(__name__)
if len(sys.argv) != 2:
log.error("Missing argument: <JSON config file>")
exit(1)
argsDict = JsonArgParser(PARAMETERS).parse(sys.argv[1])
args = dict2obj(argsDict, 'ShortDocArguments')
logging.getLogger(__name__).info(argsDict)
if args.seed:
random.seed(args.seed)
np.random.seed(args.seed)
lr = args.lr
startSymbol = args.start_symbol
endSymbol = args.end_symbol
numEpochs = args.num_epochs
shuffle = args.shuffle
normalizeMethod = args.normalization
wordWindowSize = args.word_window_size
hiddenLayerSize = args.hidden_size
convSize = args.conv_size
# Load classes for filters.
filters = []
for filterName in args.filters:
moduleName, className = filterName.rsplit('.', 1)
log.info("Filtro: " + moduleName + " " + className)
module_ = importlib.import_module(moduleName)
filters.append(getattr(module_, className)())
W1 = None
b1 = None
W2 = None
b2 = None
wordEmbedding = None
if args.word_embedding:
log.info("Reading W2v File")
(wordLexicon, wordEmbedding) = Embedding.fromWord2Vec(args.word_embedding, unknownSymbol="__UNKNOWN__")
wordLexicon.stopAdd()
elif args.word_lexicon and args.word_emb_size:
wordLexicon = Lexicon.fromTextFile(args.word_lexicon, hasUnknowSymbol=False)
wordEmbedding = Embedding(wordLexicon, embeddingSize=args.word_emb_size)
wordLexicon.stopAdd()
else:
log.error("You must provide argument word_embedding or word_lexicon and word_emb_size")
# Create the lexicon of labels.
labelLexicon = None
if args.labels is not None:
if args.label_lexicon is not None:
log.error("Only one of the parameters label_lexicon and labels can be provided!")
exit(1)
labelLexicon = Lexicon.fromList(args.labels, hasUnknowSymbol=False)
elif args.label_lexicon is not None:
labelLexicon = Lexicon.fromTextFile(args.label_lexicon, hasUnknowSymbol=False)
else:
log.error("One of the parameters label_lexicon or labels must be provided!")
exit(1)
#
# Build the network model (Theano graph).
#
# TODO: debug
# theano.config.compute_test_value = 'warn'
# ex = trainIterator.next()
# inWords.tag.test_value = ex[0][0]
# outLabel.tag.test_value = ex[1][0]
# Matriz de entrada. Cada linha representa um token da oferta. Cada token é
# representado por uma janela de tokens (token central e alguns tokens
# próximos). Cada valor desta matriz corresponde a um índice que representa
# um token no embedding.
inWords = tensor.lmatrix("inWords")
# Categoria correta de uma oferta.
outLabel = tensor.lscalar("outLabel")
# List of input tensors. One for each input layer.
inputTensors = [inWords]
# Whether the word embedding will be updated during training.
embLayerTrainable = not args.fix_word_embedding
if not embLayerTrainable:
log.info("Not updating the word embedding!")
# Lookup table for word features.
embeddingLayer = EmbeddingLayer(inWords, wordEmbedding.getEmbeddingMatrix(), trainable=embLayerTrainable)
# if not args.train and args.load_wordEmbedding:
# attrs = np.load(args.load_wordEmbedding)
# embeddingLayer.load(attrs)
# log.info("Loaded word embedding (shape %s) from file %s" % (
# str(attrs[0].shape), args.load_wordEmbedding))
# A saída da lookup table possui 3 dimensões (numTokens, szWindow, szEmbedding).
# Esta camada dá um flat nas duas últimas dimensões, produzindo uma saída
# com a forma (numTokens, szWindow * szEmbedding).
flattenInput = FlattenLayer(embeddingLayer)
# Random weight initialization procedure.
weightInit = GlorotUniform()
# Convolution layer. Convolução no texto de uma oferta.
convW = None
convb = None
if not args.train and args.load_conv:
convNPY = np.load(args.load_conv)
convW = convNPY[0]
convb = convNPY[1]
log.info("Loaded convolutional layer (shape %s) from file %s" % (str(convW.shape), args.load_conv))
convLinear = LinearLayer(flattenInput,
wordWindowSize * wordEmbedding.getEmbeddingSize(),
convSize, W=convW, b=convb,
weightInitialization=weightInit)
if args.conv_act:
convOut = ActivationLayer(convLinear, tanh)
else:
convOut = convLinear
# Max pooling layer.
maxPooling = MaxPoolingLayer(convOut)
# Hidden layer.
if not args.train and args.load_hiddenLayer:
hiddenNPY = np.load(args.load_hiddenLayer)
W1 = hiddenNPY[0]
b1 = hiddenNPY[1]
log.info("Loaded hidden layer (shape %s) from file %s" % (str(W1.shape), args.load_hiddenLayer))
hiddenLinear = LinearLayer(maxPooling,
convSize,
hiddenLayerSize,
W=W1, b=b1,
weightInitialization=weightInit)
hiddenAct = ActivationLayer(hiddenLinear, tanh)
# Entrada linear da camada softmax.
if not args.train and args.load_softmax:
hiddenNPY = np.load(args.load_softmax)
W2 = hiddenNPY[0]
b2 = hiddenNPY[1]
log.info("Loaded softmax layer (shape %s) from file %s" % (str(W2.shape), args.load_softmax))
sotmaxLinearInput = LinearLayer(hiddenAct,
hiddenLayerSize,
labelLexicon.getLen(),
W=W2, b=b2,
weightInitialization=ZeroWeightGenerator())
# Softmax.
# softmaxAct = ReshapeLayer(ActivationLayer(sotmaxLinearInput, softmax), (1, -1))
softmaxAct = ActivationLayer(sotmaxLinearInput, softmax)
# Prediction layer (argmax).
prediction = ArgmaxPrediction(None).predict(softmaxAct.getOutput())
# Loss function.
if args.label_weights is not None and len(args.label_weights) != labelLexicon.getLen():
log.error("Number of label weights (%d) is different from number of labels (%d)!" % (
len(args.label_weights), labelLexicon.getLen()))
nlloe = NegativeLogLikelihoodOneExample(weights=args.label_weights)
loss = nlloe.calculateError(softmaxAct.getOutput()[0], prediction, outLabel)
# Input generators: word window.
inputGenerators = [WordWindowGenerator(wordWindowSize, wordLexicon, filters, startSymbol, endSymbol)]
# Output generator: generate one label per offer.
outputGenerators = [TextLabelGenerator(labelLexicon)]
# outputGenerators = [lambda label: labelLexicon.put(label)]
evalPerIteration = None
if args.train:
trainDatasetReader = ShortDocReader(args.train)
if args.load_method == "sync":
log.info("Reading training examples...")
trainIterator = SyncBatchIterator(trainDatasetReader,
inputGenerators,
outputGenerators,
- 1,
shuffle=shuffle)
wordLexicon.stopAdd()
elif args.load_method == "async":
log.info("Examples will be asynchronously loaded.")
trainIterator = AsyncBatchIterator(trainDatasetReader,
inputGenerators,
outputGenerators,
- 1,
shuffle=shuffle,
maxqSize=1000)
else:
log.error("The argument 'load_method' has an invalid value: %s." % args.load_method)
sys.exit(1)
labelLexicon.stopAdd()
# Get dev inputs and output
dev = args.dev
evalPerIteration = args.eval_per_iteration
if not dev and evalPerIteration > 0:
log.error("Argument eval_per_iteration cannot be used without a dev argument.")
sys.exit(1)
if dev:
log.info("Reading development examples")
devReader = ShortDocReader(args.dev)
devIterator = SyncBatchIterator(devReader,
inputGenerators,
outputGenerators,
- 1,
shuffle=False)
else:
devIterator = None
else:
trainIterator = None
devIterator = None
if normalizeMethod == "minmax":
log.info("Normalization: minmax")
wordEmbedding.minMaxNormalization()
elif normalizeMethod == "mean":
log.info("Normalization: mean normalization")
wordEmbedding.meanNormalization()
elif normalizeMethod == "zscore":
log.info("Normalization: zscore normalization")
wordEmbedding.zscoreNormalization()
elif normalizeMethod:
log.error("Normalization: unknown value %s" % normalizeMethod)
sys.exit(1)
# Decaimento da taxa de aprendizado.
decay = None
if args.decay == "none":
decay = 0.0
elif args.decay == "linear":
decay = 1.0
else:
log.error("Unknown decay parameter %s." % args.decay)
exit(1)
# Algoritmo de aprendizado.
if args.alg == "adagrad":
log.info("Using Adagrad")
opt = Adagrad(lr=lr, decay=decay)
elif args.alg == "sgd":
log.info("Using SGD")
opt = SGD(lr=lr, decay=decay)
else:
log.error("Unknown algorithm: %s." % args.alg)
sys.exit(1)
# TODO: debug
# opt.lr.tag.test_value = 0.05
# Printing embedding information.
dictionarySize = wordEmbedding.getNumberOfVectors()
embeddingSize = wordEmbedding.getEmbeddingSize()
log.info("Dictionary size: %d" % dictionarySize)
log.info("Embedding size: %d" % embeddingSize)
log.info("Number of categories: %d" % labelLexicon.getLen())
# Train metrics.
trainMetrics = None
if trainIterator:
trainMetrics = [
LossMetric("TrainLoss", loss),
AccuracyMetric("TrainAccuracy", outLabel, prediction)
]
# Evaluation metrics.
evalMetrics = None
if devIterator:
evalMetrics = [
LossMetric("EvalLoss", loss),
AccuracyMetric("EvalAccuracy", outLabel, prediction),
FMetric("EvalFMetric", outLabel, prediction, labels=labelLexicon.getLexiconDict().values())
]
# Test metrics.
testMetrics = None
if args.test:
testMetrics = [
LossMetric("TestLoss", loss),
AccuracyMetric("TestAccuracy", outLabel, prediction),
FMetric("TestFMetric", outLabel, prediction, labels=labelLexicon.getLexiconDict().values())
]
# TODO: debug
# mode = theano.compile.debugmode.DebugMode(optimizer=None)
mode = None
model = BasicModel(x=inputTensors,
y=[outLabel],
allLayers=softmaxAct.getLayerSet(),
optimizer=opt,
prediction=prediction,
loss=loss,
trainMetrics=trainMetrics,
evalMetrics=evalMetrics,
testMetrics=testMetrics,
mode=mode)
# Training
if trainIterator:
log.info("Training")
model.train(trainIterator, numEpochs, devIterator, evalPerIteration=evalPerIteration)
# Saving model after training
if args.save_wordEmbedding:
embeddingLayer.saveAsW2V(args.save_wordEmbedding, lexicon=wordLexicon)
log.info("Saved word to vector to file: %s" % (args.save_wordEmbedding))
if args.save_conv:
convLinear.save(args.save_conv)
log.info("Saved convolution layer to file: %s" % (args.save_conv))
if args.save_hiddenLayer:
hiddenLinear.save(args.save_hiddenLayer)
log.info("Saved hidden layer to file: %s" % (args.save_hiddenLayer))
if args.save_softmax:
sotmaxLinearInput.save(args.save_softmax)
log.info("Saved softmax to file: %s" % (args.save_softmax))
# Testing
if args.test:
log.info("Reading test examples")
testReader = ShortDocReader(args.test)
testIterator = SyncBatchIterator(testReader,
inputGenerators,
outputGenerators,
- 1,
shuffle=False)
log.info("Testing")
model.test(testIterator)
# Creates the model
mdaModel = Model(input, x, act2)
sgd = SGD(lr, decay=0.0)
encoderW = linear1.getParameters()[0]
encoderOutput = act1.getOutput()
loss = MDALoss(MeanSquaredError(), True, noiseRate, encoderW,
encoderOutput, x)
log.info("Compiling the model")
mdaModel.compile(sgd, loss)
cbs = []
if saveModel:
writter = MDAModelWritter(saveModel, linear1, linear2)
cbs.append(SaveModelCallback(writter, "loss", False))
log.info("Traning model")
mdaModel.train(trainBatchGenerator, numEpochs, callbacks=cbs)
if __name__ == '__main__':
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'))
parameters = JsonArgParser(MDA_PARAMETERS).parse(sys.argv[1])
main(**parameters)
testIterator = SyncBatchIterator(testReader,
inputGenerators,
outputGenerators,
-1,
shuffle=False)
log.info("Testing")
model.test(testIterator)
def method_name(hiddenActFunction):
if hiddenActFunction == "tanh":
return tanh
elif hiddenActFunction == "sigmoid":
return sigmoid
else:
raise Exception("'hidden_activation_function' value don't valid.")
if __name__ == '__main__':
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'), defaults={})
argsDict = JsonArgParser(PARAMETERS).parse(sys.argv[1])
args = dict2obj(argsDict, 'OfertaArguments')
logging.getLogger(__name__).info(argsDict)
main(args)
jsonNs["lr"] = args.lr_ns
jsonNs["noise_rate"] = args.noise_rate
jsonNs["num_epochs"] = args.num_epochs
jsonNs["t"] = args.t
jsonNs["power"] = args.power
jsonNs["min_count"] = args.min_count
jsonNs["save_model"] = args.save_model
jsonWnn = JsonArgParser(wt.WNN_PARAMETERS).parse(
args.wnn_with_target_json_base)
jsonWnn["lr"] = args.lr_wnn
jsonWnn["target_module"] = args.save_model
log = logging.getLogger(__name__)
log.info("Starting unsupervised training")
ns.mainWnnNegativeSampling(dict2obj(jsonNs))
log.info("Starting supervised training")
wt.mainWnn(dict2obj(jsonWnn))
if __name__ == '__main__':
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'))
parameters = dict2obj(JsonArgParser(PARAMETERS).parse(sys.argv[1]))
main(parameters)
def main():
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'), defaults={})
log = logging.getLogger(__name__)
if len(sys.argv) != 3:
log.error("Missing argument: <JSON config file> or/and <Input file>")
exit(1)
argsDict = JsonArgParser(PARAMETERS).parse(sys.argv[1])
args = dict2obj(argsDict, 'ShortDocArguments')
logging.getLogger(__name__).info(argsDict)
if args.seed:
random.seed(args.seed)
np.random.seed(args.seed)
lr = args.lr
startSymbol = args.start_symbol
endSymbol = args.end_symbol
numEpochs = args.num_epochs
shuffle = args.shuffle
normalizeMethod = args.normalization
wordWindowSize = args.word_window_size
hiddenLayerSize = args.hidden_size
convSize = args.conv_size
# Load classes for filters.
filters = []
for filterName in args.filters:
moduleName, className = filterName.rsplit('.', 1)
log.info("Filtro: " + moduleName + " " + className)
module_ = importlib.import_module(moduleName)
filters.append(getattr(module_, className)())
W1 = None
b1 = None
W2 = None
b2 = None
wordEmbedding = None
if args.word_embedding:
log.info("Reading W2v File")
(wordLexicon,
wordEmbedding) = Embedding.fromWord2Vec(args.word_embedding,
unknownSymbol="__UNKNOWN__")
wordLexicon.stopAdd()
elif args.word_lexicon and args.word_emb_size:
wordLexicon = Lexicon.fromTextFile(args.word_lexicon,
hasUnknowSymbol=False)
wordEmbedding = Embedding(wordLexicon,
embeddingSize=args.word_emb_size)
wordLexicon.stopAdd()
else:
log.error(
"You must provide argument word_embedding or word_lexicon and word_emb_size"
)
# Create the lexicon of labels.
labelLexicon = None
if args.labels is not None:
if args.label_lexicon is not None:
log.error(
"Only one of the parameters label_lexicon and labels can be provided!"
)
exit(1)
labelLexicon = Lexicon.fromList(args.labels, hasUnknowSymbol=False)
elif args.label_lexicon is not None:
labelLexicon = Lexicon.fromTextFile(args.label_lexicon,
hasUnknowSymbol=False)
else:
log.error(
"One of the parameters label_lexicon or labels must be provided!")
exit(1)
#
# Build the network model (Theano graph).
#
# TODO: debug
# theano.config.compute_test_value = 'warn'
# ex = trainIterator.next()
# inWords.tag.test_value = ex[0][0]
# outLabel.tag.test_value = ex[1][0]
# Matriz de entrada. Cada linha representa um token da oferta. Cada token é
# representado por uma janela de tokens (token central e alguns tokens
# próximos). Cada valor desta matriz corresponde a um índice que representa
# um token no embedding.
inWords = tensor.lmatrix("inWords")
# Categoria correta de uma oferta.
outLabel = tensor.lscalar("outLabel")
# List of input tensors. One for each input layer.
inputTensors = [inWords]
# Whether the word embedding will be updated during training.
embLayerTrainable = not args.fix_word_embedding
if not embLayerTrainable:
log.info("Not updating the word embedding!")
# Lookup table for word features.
embeddingLayer = EmbeddingLayer(inWords,
wordEmbedding.getEmbeddingMatrix(),
trainable=embLayerTrainable)
# if not args.train and args.load_wordEmbedding:
# attrs = np.load(args.load_wordEmbedding)
# embeddingLayer.load(attrs)
# log.info("Loaded word embedding (shape %s) from file %s" % (
# str(attrs[0].shape), args.load_wordEmbedding))
# A saída da lookup table possui 3 dimensões (numTokens, szWindow, szEmbedding).
# Esta camada dá um flat nas duas últimas dimensões, produzindo uma saída
# com a forma (numTokens, szWindow * szEmbedding).
flattenInput = FlattenLayer(embeddingLayer)
# Random weight initialization procedure.
weightInit = GlorotUniform()
# Convolution layer. Convolução no texto de uma oferta.
convW = None
convb = None
if not args.train and args.load_conv:
convNPY = np.load(args.load_conv)
convW = convNPY[0]
convb = convNPY[1]
log.info("Loaded convolutional layer (shape %s) from file %s" %
(str(convW.shape), args.load_conv))
convLinear = LinearLayer(flattenInput,
wordWindowSize * wordEmbedding.getEmbeddingSize(),
convSize,
W=convW,
b=convb,
weightInitialization=weightInit)
# Max pooling layer.
maxPooling = MaxPoolingLayer(convLinear)
# Hidden layer.
if not args.train and args.load_hiddenLayer:
hiddenNPY = np.load(args.load_hiddenLayer)
W1 = hiddenNPY[0]
b1 = hiddenNPY[1]
log.info("Loaded hidden layer (shape %s) from file %s" %
(str(W1.shape), args.load_hiddenLayer))
hiddenLinear = LinearLayer(maxPooling,
convSize,
hiddenLayerSize,
W=W1,
b=b1,
weightInitialization=weightInit)
hiddenAct = ActivationLayer(hiddenLinear, tanh)
# Entrada linear da camada softmax.
if not args.train and args.load_softmax:
hiddenNPY = np.load(args.load_softmax)
W2 = hiddenNPY[0]
b2 = hiddenNPY[1]
log.info("Loaded softmax layer (shape %s) from file %s" %
(str(W2.shape), args.load_softmax))
sotmaxLinearInput = LinearLayer(hiddenAct,
hiddenLayerSize,
labelLexicon.getLen(),
W=W2,
b=b2,
weightInitialization=ZeroWeightGenerator())
# Softmax.
# softmaxAct = ReshapeLayer(ActivationLayer(sotmaxLinearInput, softmax), (1, -1))
softmaxAct = ActivationLayer(sotmaxLinearInput, softmax)
# Prediction layer (argmax).
prediction = ArgmaxPrediction(None).predict(softmaxAct.getOutput())
# Loss function.
if args.label_weights is not None and len(
args.label_weights) != labelLexicon.getLen():
log.error(
"Number of label weights (%d) is different from number of labels (%d)!"
% (len(args.label_weights), labelLexicon.getLen()))
nlloe = NegativeLogLikelihoodOneExample(weights=args.label_weights)
loss = nlloe.calculateError(softmaxAct.getOutput()[0], prediction,
outLabel)
# Input generators: word window.
inputGenerators = [
WordWindowGenerator(wordWindowSize, wordLexicon, filters, startSymbol,
endSymbol)
]
# Output generator: generate one label per offer.
outputGenerators = [TextLabelGenerator(labelLexicon)]
# outputGenerators = [lambda label: labelLexicon.put(label)]
evalPerIteration = None
if normalizeMethod == "minmax":
log.info("Normalization: minmax")
wordEmbedding.minMaxNormalization()
elif normalizeMethod == "mean":
log.info("Normalization: mean normalization")
wordEmbedding.meanNormalization()
elif normalizeMethod == "zscore":
log.info("Normalization: zscore normalization")
wordEmbedding.zscoreNormalization()
elif normalizeMethod:
log.error("Normalization: unknown value %s" % normalizeMethod)
sys.exit(1)
# Decaimento da taxa de aprendizado.
decay = None
if args.decay == "none":
decay = 0.0
elif args.decay == "linear":
decay = 1.0
else:
log.error("Unknown decay parameter %s." % args.decay)
exit(1)
# Algoritmo de aprendizado.
if args.alg == "adagrad":
log.info("Using Adagrad")
opt = Adagrad(lr=lr, decay=decay)
elif args.alg == "sgd":
log.info("Using SGD")
opt = SGD(lr=lr, decay=decay)
else:
log.error("Unknown algorithm: %s." % args.alg)
sys.exit(1)
# TODO: debug
# opt.lr.tag.test_value = 0.05
# Printing embedding information.
dictionarySize = wordEmbedding.getNumberOfVectors()
embeddingSize = wordEmbedding.getEmbeddingSize()
log.info("Dictionary size: %d" % dictionarySize)
log.info("Embedding size: %d" % embeddingSize)
log.info("Number of categories: %d" % labelLexicon.getLen())
# TODO: debug
# mode = theano.compile.debugmode.DebugMode(optimizer=None)
mode = None
model = BasicModel(x=inputTensors,
y=[outLabel],
allLayers=softmaxAct.getLayerSet(),
optimizer=opt,
prediction=prediction,
loss=loss,
mode=mode)
wordWindow = WordWindowGenerator(wordWindowSize, wordLexicon, filters,
startSymbol, endSymbol)
# GETS HIDDEN LAYER:
# graph = EmbeddingGraph([inWords], [hiddenAct.getOutput()], wordWindow)
# GRAPH FOR PREDICTION LAYER
graph = EmbeddingGraph(inputTensors, prediction, wordWindow, mode)
lblTxt = ["Sim", "Nao"]
tweets = []
with open(sys.argv[2]) as inputFile:
content = inputFile.readlines()
for line in content:
tweets.append(line.decode('utf-8').encode('utf-8'))
#print tweets
# graph.getResultsFor(t) retorna a predição para dado Tweet t
try:
output_file = open("Output.txt", "w")
except:
print "Falha em criar o arquivo de saida\n"
try:
for t in tweets:
output_file.write(
t.replace('\n', '').replace('\t', '') + "\t " +
lblTxt[graph.getResultsFor(t)] + "\n")
print "Resultados gerados com sucesso!\n"
except:
print "Erro na geração de resultados\n"
[outputGenerator],
sys.maxint,
shuffle=False)
lambdaChange = ChangeLambda(_lambdaShared, kwargs["lambda"],
kwargs["loss_uns_epoch"])
lossCallback = LossCallback(loss1, loss2, input1, input2, y)
# trainUnsupervisedDatasetReaderAcc = SyncBatchIterator(trainUnsupervisedDataset,
# [inputGenerator1, inputGenerator2],
# [outputGenerator], sys.maxint)
# accCallBack = AccCallBack(prediction1, prediction2, input1, input2,
# unsurpervisedDataset=trainUnsupervisedDatasetReaderAcc)
# Training Model
model.train([trainSupervisedDatasetReader, trainUnsupervisedDatasetReader],
numEpochs,
devReader,
callbacks=[lambdaChange, lossCallback])
if __name__ == '__main__':
full_path = os.path.realpath(__file__)
path, filename = os.path.split(full_path)
logging.config.fileConfig(os.path.join(path, 'logging.conf'))
parameters = JsonArgParser(CO_LEARNING_PARAMETERS).parse(sys.argv[1])
main(**parameters)