def __init__(self,batch_size=16, seed=1234,nhu=300,width=5,n_out=len(nerarray),activation_f="hardtanh",
embeddingfile=senna_embmtxfile,trainingfile=trainingfile,paramfile=None):
modeldir=os.path.join(nerdir,"models",'model_%i'%(len(os.listdir(nerdir+"/models"))))
os.mkdir(modeldir)
for handler in logging.root.handlers[:]:
logging.root.removeHandler(handler)
logging.basicConfig(filename=os.path.join(modeldir,'log.txt'), level=logging.INFO,
format='%(asctime)s : %(levelname)s : %(message)s')
logger.info("\n"+"\n".join(["\t%s : "%key+str(val) for key,val in locals().iteritems() if key!="self"]))
self.modeldir=modeldir
self.batch_size = batch_size
activation=None
if activation_f=="hardtanh":
activation=hardtanh
elif activation_f=="tanh":
activation=T.tanh
self.load_data(embeddingfile,trainingfile,batch_size)
#==============================================================================
# BUILD MODEL
#==============================================================================
logger.info('... building the model')
# allocate symbolic variables for the data
self.index = T.iscalar() # index to a [mini]batch
self.x = T.itensor3('x') # the data is presented as matrix of integers
self.y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
self.permutation = T.ivector('permutation')
if paramfile!=None:
params=pickle.load(open(paramfile,"rb"))
else:
params=None
self.model = SennaNER(input=self.x, embeddings=self.embeddings,features=capsfeatures,n_out=n_out, mini_batch_size=batch_size,
nhu=nhu,width=width,activation=activation,seed=seed,params=params)
self.test_model = theano.function(inputs=[self.index],
outputs=self.model.errors(self.y),
givens={
self.x: self.test_set_x[self.index * batch_size:(self.index + 1) * batch_size],
self.y: self.test_set_y[self.index * batch_size:(self.index + 1) * batch_size]},
name="test_model")
self.validation_cost = theano.function(inputs=[self.index],
outputs=self.model.negative_log_likelihood(self.y),
givens={
self.x: self.valid_set_x[self.index * batch_size:(self.index + 1) * batch_size],
self.y: self.valid_set_y[self.index * batch_size:(self.index + 1) * batch_size]},
name="validation_cost")
self.predictions = theano.function(inputs=[self.index],
outputs=self.model.predictions,
givens={
self.x: self.test_set_x[self.index * batch_size:(self.index + 1) * batch_size]},
name="predictions")
self.visualize_hidden = theano.function(inputs=[self.index],
outputs=self.model.HiddenLayer.output,
givens={
self.x: self.valid_set_x[self.index * batch_size:(self.index + 1) * batch_size]},
name="visualize_hidden")
