def __init__(self, fold):
self.bestEpoch = None
self.maxF1 = 0
self.persistence = ModelPersistence()
self.patternRegular = Config.Instance().resultsDir + '/fold_' + str(
fold) + '/model_'
self.patternEpoch = Config.Instance().resultsDir + '/fold_' + str(
fold) + '/regular_'
self.keepBestEpoch = 5
self.keepFixedPoint = 3
self.savedByBestEpoch = []
self.savedForFixedPoint = []
def readConfig(file='/srl-config.json'):
config = Config.Instance()
config.prepare(Utils.getWorkingDirectory())
modelConfig = ModelConfig.Instance()
modelConfig.prepare(config.srlConfig + file)
return config, modelConfig
def prepareEmbeddings(useWiki=False):
config = Config.Instance()
modelConfig = ModelConfig.Instance()
tokens = extractAllTokens(config.convertedCorpusDir+'/propbank_full.csv')
if useWiki:
tokens.update(extractAllTokens(config.convertedCorpusDir+'/wiki.csv'))
print '{} tokens found'.format(len(tokens))
predicates = extractAllTokens(config.convertedCorpusDir+'/propbank_full.csv', 'predicate')
w2vFiles = {
"npzFile":config.embeddingsDir+"/wordEmbeddings.npy",
"npzModel":config.embeddingsDir+"/wordEmbeddings",
"vecFile":__getVecFile(config.embeddingsDir, modelConfig.embeddingSize),
"w2idxFile":config.embeddingsDir+"/vocabulary.json"
}
w2v = W2VModel()
w2v.setResources(w2vFiles)
loader = EmbeddingLoader(w2v)
word2idx, idx2word, weights = loader.process()
if modelConfig.embeddingType == 'w2v':
return loader, loader
sentHybridFiles = {
"npzFile":config.embeddingsDir+"/sent_hybrid.npy",
"npzModel":config.embeddingsDir+"/sent_hybrid",
"w2idxFile":config.embeddingsDir+"/sent_hybrid.json"
}
sentHybrid = HybridModel()
sentHybrid.setResources(sentHybridFiles)
print 'creating sentence corpus'
sentHybrid.generateCorpus(tokens, weights, word2idx)
Hloader = EmbeddingLoader(sentHybrid)
Hword2idx, Hidx2word, Hweights = Hloader.process()
predHybridFiles = {
"npzFile":config.embeddingsDir+"/pred_hybrid.npy",
"npzModel":config.embeddingsDir+"/pred_hybrid",
"w2idxFile":config.embeddingsDir+"/pred_hybrid.json"
}
predHybrid = HybridModel()
predHybrid.setResources(predHybridFiles)
print 'creating predicate corpus'
predHybrid.generateCorpus(predicates, weights, word2idx)
Ploader = EmbeddingLoader(predHybrid)
Pword2idx, Pidx2word, Pweights = Ploader.process()
return Hloader, Ploader
def deleteTrainingResources(k_folds=20):
__delete(Config.Instance().resourceDir + '/feature_file.npy')
for i in xrange(1, k_folds + 1):
__delete(Config.Instance().resourceDir + '/feature_file_' + str(i) +
'.npy')
__delete(Config.Instance().resourceDir + '/embeddings/vocabulary.json')
__delete(Config.Instance().resourceDir + '/embeddings/pred_hybrid.json')
__delete(Config.Instance().resourceDir + '/embeddings/pred_hybrid.npy')
__delete(Config.Instance().resourceDir + '/embeddings/sent_hybrid.json')
__delete(Config.Instance().resourceDir + '/embeddings/sent_hybrid.npy')
__delete(Config.Instance().resourceDir + '/embeddings/wordEmbeddings.npy')
def getEmbeddings():
config = Config.Instance()
modelConfig = ModelConfig.Instance()
if modelConfig.embeddingType == 'w2v':
w2vFiles = {
"npzFile":config.embeddingsDir+"/wordEmbeddings.npy",
"npzModel":config.embeddingsDir+"/wordEmbeddings",
"vecFile":__getVecFile(config.embeddingsDir, modelConfig.embeddingSize),
"w2idxFile":config.embeddingsDir+"/vocabulary.json"
}
w2v = W2VModel()
w2v.setResources(w2vFiles)
loader = EmbeddingLoader(w2v)
word2idx, idx2word, weights = loader.process()
return loader, loader
else:
sentHybridFiles = {
"npzFile":config.embeddingsDir+"/sent_hybrid.npy",
"npzModel":config.embeddingsDir+"/sent_hybrid",
"w2idxFile":config.embeddingsDir+"/sent_hybrid.json"
}
sentHybrid = HybridModel()
sentHybrid.setResources(sentHybridFiles)
Hloader = EmbeddingLoader(sentHybrid)
Hword2idx, Hidx2word, Hweights = Hloader.process()
predHybridFiles = {
"npzFile":config.embeddingsDir+"/pred_hybrid.npy",
"npzModel":config.embeddingsDir+"/pred_hybrid",
"w2idxFile":config.embeddingsDir+"/pred_hybrid.json"
}
predHybrid = HybridModel()
predHybrid.setResources(predHybridFiles)
Ploader = EmbeddingLoader(predHybrid)
Pword2idx, Pidx2word, Pweights = Ploader.process()
return Hloader, Ploader
from model.configuration import Config
from model.configuration.model_config import ModelConfig
from utils.function_utils import Utils
from utils import extractFeaturesFromSentence, toNNFormat
from embeddings import getEmbeddings
import pandas as pd
print 'loading configuration'
config = Config.Instance()
config.prepare(Utils.getWorkingDirectory())
modelConfig = ModelConfig.Instance()
modelConfig.prepare(config.srlConfig + '/srl-config.json')
print 'configuration loaded'
sentenceLoader, predicateLoader = getEmbeddings(config,
modelConfig.embeddingType)
wikiFile = pd.read_csv(config.convertedCorpusDir + '/wiki.csv')
for i in xrange(0, len(wikiFile)):
predicate = wikiFile['predicate'][i]
sentence = wikiFile['sentence'][i]
convertedSentence, convertedPredicate, allCaps, firstCaps, noCaps, context, distance = extractFeaturesFromSentence(
sentence, predicate, sentenceLoader.word2idx, predicateLoader.word2idx)
inputSentence, inputPredicate, inputAux = toNNFormat(
convertedSentence, convertedPredicate, allCaps, firstCaps, noCaps,
context, distance)
print inputSentence.shape, inputPredicate.shape, inputAux.shape
break
clr = CyclicLearningRate(base_lr=0.00020,
max_lr=0.0012,
step_size=(204. * 3),
mode='exp_range',
gamma=0.99996)
msaver = ModelEvaluation()
print 'prepared'
print 'creating neural network model'
file = str(sys.argv[1])
startingEpoch = int(sys.argv[2])
lrDefault = float(sys.argv[3])
print 'model to load : {} - epoch : {} - lr : {}'.format(
file, startingEpoch, lrDefault)
model = LSTMModel(ModelConfig.Instance())
nn = model.load(Config.Instance().resultsDir + '/' + file + '.json',
Config.Instance().resultsDir + '/' + file + '.h5py')
nn.summary()
lrReducer.setNetwork(nn)
es = EarlyStopper()
lrReducer.setLearningRate(lrDefault)
print 'model loaded'
print 'start training'
number_of_epochs = ModelConfig.Instance().trainingEpochs
for epoch in xrange(startingEpoch, number_of_epochs):
print "--------- Epoch %d -----------" % (epoch)
start_time = time.time()
numIterations = len(container)
batcher = Batcher()
batcher.addAll(trainingData[0], trainingData[1], trainingData[2],
trainingData[3])
container = batcher.getBatches()
inference = SRLInference(tagMap, tagList)
evaluator = Evaluator(testData, inference, nnUtils,
config.resultsDir + '/finalResult.json')
lrReducer = RateBasedLrReducer(modelConfig.trainingEpochs)
msaver = ModelEvaluation()
print 'prepared'
print 'creating neural network model'
mp = ModelPersistence()
nn = mp.load(
Config.Instance().resultsDir + '/model_50.json',
Config.Instance().resultsDir + '/model_50.h5py',
)
nn.compile(optimizer=modelConfig.optimizer,
loss=modelConfig.lossFunction,
metrics=['accuracy'])
nn.summary()
lrReducer.setNetwork(nn)
print 'model loaded'
print 'start training'
number_of_epochs = ModelConfig.Instance().trainingEpochs
for epoch in xrange(50, 50 + number_of_epochs):
print "--------- Epoch %d -----------" % (epoch + 1)
start_time = time.time()