def testloadData(self):
nExamples = 15
train_data = '/home/bhanu/Downloads/relationClassification/dataCamera/allDataTrain.mat'
pre_trained_weights = '/home/bhanu/Downloads/relationClassification/dataCamera/pretrainedWeights.mat'
rnnData = RNNDataCorpus()
rnnData.load_data(load_file=train_data, nExamples=nExamples)
params = Params(data=rnnData, wordSize=50, rankWo=3)
# n = params.wordSize; fanIn = params.fanIn; nWords = params.nWords; nLabels = params.categories; rank=params.rankWo
# theta = np.random.normal(loc=0.0, scale=np.math.pow(10,-5)**2, size = n*(2*n+1) + n*2*n + nLabels*fanIn + n*nWords + (2*n*rank+n)*nWords)
#
# #init respective parameters with prior values
# W, WO, Wcat, Wv, Wo = unroll_theta(theta, params)
# Wo[:n,:] = np.ones((n,Wo.shape[1])) #init word matrices with Identity matrices + epsilon
#load pre-trained weights here
mats = sio.loadmat(pre_trained_weights)
Wv = mats.get('Wv')
W = mats.get('W')
WO = mats.get('WO')
Wo = np.ndarray(
shape=(2 * params.wordSize * params.rankWo + params.wordSize) *
params.nWords,
dtype='float64')
sentencesIdx = np.arange(nExamples)
[allSNum_batch, allSNN_batch, Wv_batch, Wo_batch,
allWordInds] = getRelevantWords(rnnData, sentencesIdx, Wv, Wo, params)
print
def get_subset(self, rnnData, workerIndex):
''' data : complete train or test set for which a subset of docs is to be calculated for this worker '''
rnnData_mini = RNNDataCorpus()
dataSize = rnnData.ndoc()
sizePerNode = dataSize/(self.totalNodes-1) #exclude master node
startPos = (workerIndex-1) * sizePerNode
if(workerIndex != self.totalNodes-1): #last node on a machine gets all the remaining data
endPos = startPos+sizePerNode
else:
endPos = dataSize
rnnData.copy_into_minibatch(rnnData_mini, range(startPos, endPos))
return rnnData_mini
def initMVRNN(self):
print "Node: ",self.index," Loading training and dev sets.."
self.rnnData_train = RNNDataCorpus()
self.rnnData_train.load_data_srl(config.train_data_srl, nExamples=self.nuse)
self.rnnData_dev = RNNDataCorpus()
self.rnnData_dev.load_data_srl(config.dev_data_srl, nExamples=self.nuse)
modelfilename = config.saved_params_file+'SGD_SRLiter305'
print "Node: ", self.index," Loading model: ", modelfilename
# mats = sio.loadmat(config.saved_params_file+'iter120.mat')
# Wv = mats.get('Wv') #L, as in paper
# W = mats.get('W') #W, as in paper
# WO = mats.get('WO') #Wm, as in paper
# Wo = mats.get('Wo')
# Wcat = mats.get('Wcat')
# n = Wv.shape[0]
# r = (Wo.shape[0] - n)/(2*n)
with open(modelfilename, 'r') as loadfile:
self.rnn = cPickle.load(loadfile)#MVRNN(W, WO, Wcat, Wv, Wo)
n = self.rnn.Wv.shape[0]
r = (self.rnn.Wo.shape[0] - n)/(2*n)
print "Node: ",self.index, "initializing params.."
self.params = Params(data=self.rnnData_train, wordSize=n, rankWo=r)
# #to be removed
# Wcat = 0.005*np.random.randn(self.params.categories, self.params.fanIn)
# self.rnn.Wcat = Wcat
if(self.index == 0):
print "Master## Total trees in training set: ", self.rnnData_train.ndoc()
print "Master## nFetch: ", self.nFetch
if(self.index!=0):
self.rnnData_train = self.get_subset(self.rnnData_train, self.index)
self.rnnData_dev = None # workers don't need test set and trainTest, so to free memory release them
[_, _, self.all_train_idx] = getRelevantWords(self.rnnData_train, self.rnn.Wv,self.rnn.Wo,self.params) #sets nWords_reduced, returns new arrays
def test_write_srl_verbIndices(self):
infilenames = [
config.train_data_srl, config.test_data_srl, config.dev_data_srl
]
outfilenames = [
config.corpus_path + 'srl_vids.train',
config.corpus_path + 'srl_vids.test',
config.corpus_path + 'srl_vids.dev'
]
for infilename, outfilename in zip(infilenames, outfilenames):
rnndata = RNNDataCorpus()
rnndata.load_data_srl(infilename, nExamples=100)
allVerbIndices = rnndata.verbIndices
with open(outfilename, 'w') as wf:
for i in range(len(allVerbIndices)):
verbids = allVerbIndices[i].flatten()
verbidsStr = [str(x) for x in verbids]
wf.write("|".join(verbidsStr) + "\n")
wf.flush()
def test():
# load testing data
print "loading test data.."
rnnData = RNNDataCorpus()
rnnData.load_data_srl(load_file=config.test_data_srl, nExamples=10)
# rnnData.load_data(load_file=config.test_data, nExamples=-1)
print 'loading trained model : ', sys.argv[1]
# mats = sio.loadmat(config.model_path+str(sys.argv[1]))
# Wv = mats.get('Wv') #L, as in paper
# W = mats.get('W') #W, as in paper
# WO = mats.get('WO') #Wm, as in paper
# Wo = mats.get('Wo')
# Wcat = mats.get('Wcat')
# n = Wv.shape[0]
# r = (Wo.shape[0] - n)/(2*n)
# rnn = MVRNN(W, WO, Wcat, Wv, Wo)
with open(config.model_path + sys.argv[1], 'r') as loadfile:
rnn = cPickle.load(loadfile)
n = rnn.Wv.shape[0]
r = (rnn.Wo.shape[0] - n) / (2 * n)
print "initializing params.."
params = Params(data=rnnData, wordSize=n, rankWo=r)
print "evaluating.."
predictLabels = rnn.evaluate(params, rnnData)
print "creating labels file .."
create_preds_file(predictLabels,
rnnData.categories,
rnnData.sentenceLabels,
predictions_file=config.results_path + 'preds_srl.txt',
testKeys_file='srl_test_keys.txt')
def train():
np.random.seed(131742)
#get sentences, trees and labels
nExamples = -1
print "loading data.."
rnnData = RNNDataCorpus()
rnnData.load_data(load_file=config.train_data, nExamples=nExamples)
#initialize params
print "initializing params"
params = Params(data=rnnData, wordSize=50, rankWo=2)
#define theta
#one vector for all the parameters of mvrnn model: W, Wm, Wlabel, L, Lm
n = params.wordSize
fanIn = params.fanIn
nWords = params.nWords
nLabels = params.categories
rank = params.rankWo
Wo = 0.01 * np.random.randn(n + 2 * n * rank, nWords) #Lm, as in paper
Wo[:n, :] = np.ones((n, Wo.shape[1])) #Lm, as in paper
Wcat = 0.005 * np.random.randn(nLabels, fanIn) #Wlabel, as in paper
# Wv = 0.01*np.random.randn(n, nWords)
# WO = 0.01*np.random.randn(n, 2*n)
# W = 0.01*np.random.randn(n, 2*n+1)
#load pre-trained weights here
mats = sio.loadmat(config.pre_trained_weights)
Wv = mats.get('Wv') #L, as in paper
W = mats.get('W') #W, as in paper
WO = mats.get('WO') #Wm, as in paper
sentencesIdx = np.arange(rnnData.ndoc())
np.random.shuffle(sentencesIdx)
nTrain = 4 * len(sentencesIdx) / 5
trainSentIdx = sentencesIdx[0:nTrain]
testSentIdx = sentencesIdx[nTrain:]
batchSize = 5
nBatches = len(trainSentIdx) / batchSize
evalFreq = 5 #evaluate after every 5 minibatches
nTestSentEval = 50 #number of test sentences to be evaluated
rnnData_train = RNNDataCorpus()
rnnData.copy_into_minibatch(rnnData_train, trainSentIdx)
rnnData_test = RNNDataCorpus()
if (len(testSentIdx) > nTestSentEval):
# np.random.shuffle(testSentIdx) #choose random test examples
thisTestSentIdx = testSentIdx[:nTestSentEval]
else:
thisTestSentIdx = testSentIdx
rnnData.copy_into_minibatch(rnnData_test, thisTestSentIdx)
# [Wv_test, Wo_test, _] = getRelevantWords(rnnData_test, Wv,Wo,params)
[Wv_trainTest, Wo_trainTest, all_train_idx
] = getRelevantWords(rnnData, Wv, Wo,
params) #sets nWords_reduced, returns new arrays
theta = np.concatenate((W.flatten(), WO.flatten(), Wcat.flatten(),
Wv_trainTest.flatten(), Wo_trainTest.flatten()))
#optimize
print "starting training..."
nIter = 100
rnnData_minibatch = RNNDataCorpus()
for i in range(nIter):
#train in minibatches
# ftrain = np.zeros(nBatches)
# for ibatch in range(nBatches):
# set_minibatch(rnnData, rnnData_minibatch, ibatch, nBatches, trainSentIdx)
# print 'Iteration: ', i, ' minibatch: ', ibatch
tunedTheta, fbatch_train, _ = lbfgsb.fmin_l_bfgs_b(
func=costFn,
x0=theta,
fprime=None,
args=(rnnData_train, params),
approx_grad=0,
bounds=None,
m=5,
factr=1000000000000000.0,
pgtol=1.0000000000000001e-5,
epsilon=1e-08,
iprint=3,
maxfun=1,
disp=0)
#map parameters back
W[:, :], WO[:, :], Wcat[:, :], Wv_trainTest, Wo_trainTest = unroll_theta(
tunedTheta, params)
Wv[:, all_train_idx] = Wv_trainTest
Wo[:, all_train_idx] = Wo_trainTest
# ftrain[ibatch] = fbatch_train
theta = tunedTheta #for next iteration
print "========================================"
print "XXXXXXIteration ", i,
print "Average cost: ", np.average(fbatch_train)
evaluate(Wv, Wo, W, WO, Wcat, params, rnnData_test)
print "========================================"
#save weights
save_dict = {'Wv': Wv, 'Wo': Wo, 'Wcat': Wcat, 'W': W, 'WO': WO}
sio.savemat(config.saved_params_file + '_lbfgs_iter' + str(i),
mdict=save_dict)
print "saved tuned theta. "
def writeVectors():
vecFileName = config.results_path + "vectors.out"
vecFile = open(vecFileName, 'w')
mats = sio.loadmat(config.corpus_path + 'vars.normalized.100.mat')
We_orig = mats.get('We')
params = sio.loadmat(config.corpus_path + 'params_rae.mat')
W1 = params.get('W1')
W2 = params.get('W2')
b1 = params.get('b1')
We = params.get('We')
b = params.get('b')
W = params.get('W')
hiddenSize = 100
nExamples = 5
print "loading data.."
rnnData_train = RNNDataCorpus()
rnnData_train.load_data_srl(load_file=config.train_data_srl,
nExamples=nExamples)
print 'writing vectors to: ', vecFileName
for ii in range(len(rnnData_train.allSNum)):
sNum = rnnData_train.allSNum[ii]
sStr = rnnData_train.allSStr[ii]
sTree = rnnData_train.allSTree[ii]
sKids = rnnData_train.allSKids[ii]
words_indexed = np.where(sNum >= 0)[0]
#L is only the part of the embedding matrix that is relevant for this sentence
#L is deltaWe
if We.shape[1] != 0:
L = We[:, words_indexed]
words_embedded = We_orig[:, words_indexed] + L
else:
words_embedded = We_orig[:, words_indexed]
# sl = words_embedded.shape[1]
tree = Tree()
tree.pp = all #np.zeros(((2*sl-1),1))
tree.nodeScores = np.zeros(len(sNum))
# tree.nodeNames = np.arange(1,(2*sl-1))
tree.kids = np.zeros((len(sNum), 2))
tree.nodeFeatures = np.zeros((hiddenSize, len(sNum)))
tree.nodeFeatures[:, :len(words_indexed)] = words_embedded
toMerge = np.zeros(shape=(words_indexed.shape), dtype='int32')
toMerge[:] = words_indexed[:]
while len(toMerge) > 1:
# find unpaired bottom leaf pairs (initially words) that share parent
i = -1
foundGoodPair = False
while (not foundGoodPair):
i += 1
if sTree[toMerge[i]] == sTree[toMerge[i + 1]]:
foundGoodPair = True
newParent = sTree[toMerge[i]]
kid1 = toMerge[i]
kid2 = toMerge[i + 1]
tree.kids[newParent, :] = [kid1, kid2]
# set new parent to be possible merge candidate
toMerge[i] = newParent
# delete other kid
toMerge = np.delete(toMerge, i + 1)
c1 = tree.nodeFeatures[:, kid1]
c2 = tree.nodeFeatures[:, kid2]
p = np.tanh(np.dot(W1, c1) + np.dot(W2, c2) + b1.flatten())
tree.nodeFeatures[:, newParent] = p
vec = tree.nodeFeatures[-1]
vecFile.write(" ".join([str(x) for x in vec]) + '\n')
vecFile.close()
print "finished! "
def train():
SEED = 13742
load_model = False
custom_load = False
np.random.seed(SEED)
#get sentences, trees and labels
nExamples = 5
print "loading data.."
rnnData_train = RNNDataCorpus()
rnnData_train.load_data_srl(load_file=config.train_data_srl,
nExamples=nExamples)
rnnData_dev = RNNDataCorpus()
rnnData_dev.load_data_srl(load_file=config.dev_data_srl,
nExamples=nExamples)
print "Number of sentences loaded in training data: ", rnnData_train.ndoc()
#initialize params
print "initializing params"
params = Params(data=rnnData_train, wordSize=52, rankWo=2)
n = params.wordSize
fanIn = params.fanIn
nWords = params.nWords
nLabels = params.categories
rank = params.rankWo
if (load_model):
modelfile = config.saved_params_file + 'SGD_SLL300'
rnn = MVRNNSLL.load(modelfile)
print 'loaded model : ', modelfile
elif (custom_load):
modelfile = config.saved_params_file + 'SGD_SLL300'
print "loading customized model..", modelfile
# d = 2#extra features for wordvectors
# Wo = 0.01*np.random.randn(n + 2*n*rank, nWords) #Lm, as in paper
# Wo[:n,:] = np.ones((n,Wo.shape[1])) #Lm, as in paper
# Wcat = 0.005*np.random.randn(nLabels, fanIn) #Wlabel, as in paper
# Wv = 0.01*np.random.randn(n, nWords)
# WO = 0.01*np.random.randn(n, 2*n)
# W = 0.01*np.random.randn(n, 2*n+1)
#load pre-trained weights here
oldrnn = MVRNNSLL.load(modelfile)
# Wv[:-d,:] = oldrnn.Wv
categories = [x.strip() for x in rnnData_train.categories]
Tran = init_transitions(dict(zip(categories, range(len(categories)))),
'iob')
rnn = MVRNNSLL(oldrnn.W, oldrnn.WO, oldrnn.Wcat, oldrnn.Wv, oldrnn.Wo,
Tran)
else:
#define theta
#one vector for all the parameters of mvrnn model: W, Wm, Wlabel, L, Lm
# n = params.wordSize; fanIn = params.fanIn; nWords = params.nWords; nLabels = params.categories; rank=params.rankWo
# Wo = 0.01*np.random.randn(n + 2*n*rank, nWords) #Lm, as in paper
# Wo[:n,:] = np.ones((n,Wo.shape[1])) #Lm, as in paper
# Wcat = 0.005*np.random.randn(nLabels, fanIn) #Wlabel, as in paper
# #load pre-trained weights here
## mats = sio.loadmat(config.saved_params_file)
# oldrnn = MVRNNSLL.load(modelfile)
# Wv = oldrnn.Wv #L, as in paper
# W = oldrnn..get('W') #W, as in paper
# WO = mats.get('WO') #Wm, as in paper
Wo = 0.01 * np.random.randn(n + 2 * n * rank, nWords) #Lm, as in paper
Wo[:n, :] = np.ones((n, Wo.shape[1])) #Lm, as in paper
Wcat = 0.005 * np.random.randn(nLabels, fanIn) #Wlabel, as in paper
Wv = 0.01 * np.random.randn(n, nWords)
WO = 0.01 * np.random.randn(n, 2 * n)
W = 0.01 * np.random.randn(n, 2 * n + 1)
categories = [x.strip() for x in rnnData_train.categories]
Tran = init_transitions(dict(zip(categories, range(len(categories)))),
'iob')
rnn = MVRNNSLL(W, WO, Wcat, Wv, Wo, Tran)
# rnn = MVRNNSLL(W, WO, Wcat, Wv, Wo, Tran)
[_, _, all_train_idx
] = getRelevantWords(rnnData_train, rnn.Wv, rnn.Wo,
params) #sets nWords_reduced, returns new arrays
params.setNumReducedWords(len(all_train_idx))
theta = rnn.getTheta(params, all_train_idx)
#optimize
print "starting training using SGD..."
nIter = 500
optimizer = StochasticGradientDescent(niter=nIter,
learning_rate=0.01,
learningrateFactor=1.0,
printAt10Iter='.',
printAt100Iter='\n+')
optimizer.minimizeBatches(rnn=rnn,
rnnData_train=rnnData_train,
allTrainSentIdx=all_train_idx,
params=params,
x0=theta,
func=costFn,
fprime=None,
rnnData_test=rnnData_dev,
initialSetSize=1,
niter=nIter,
seed=SEED,
modelFileName=config.saved_params_file +
'SGD_SLL',
printStatistics=True,
modelSaveIter=100,
nIterInPart=1,
nFetch=-1,
rnd=None,
nodeid=-1)
def test_load_srl_data(self):
rnndata = RNNDataCorpus()
rnndata.load_data_srl(config.test_data_srl, nExamples=-1)
def train():
SEED = 131742
load_model = False
custom_load = True #loads model from previously saved model except Wcat
np.random.seed(SEED)
#get sentences, trees and labels
nExamples = 5
print "loading data.."
rnnData_train = RNNDataCorpus()
rnnData_train.load_data_srl(load_file=config.train_data_srl,
nExamples=nExamples)
rnnData_dev = RNNDataCorpus()
rnnData_dev.load_data_srl(load_file=config.dev_data_srl,
nExamples=nExamples)
print "Number of sentences loaded in training data: ", rnnData_train.ndoc()
#initialize params
print "initializing params"
params = Params(data=rnnData_train, wordSize=52, rankWo=2)
n = params.wordSize
fanIn = params.fanIn
nWords = params.nWords
nLabels = params.categories
rank = params.rankWo
if (load_model):
with open(config.saved_params_file + "_45", 'r') as loadfile:
rnn = cPickle.load(loadfile)
elif (custom_load):
d = 2 #extra features for wordvectors
Wo = 0.01 * np.random.randn(n + 2 * n * rank, nWords) #Lm, as in paper
Wo[:n, :] = np.ones((n, Wo.shape[1])) #Lm, as in paper
Wcat = 0.005 * np.random.randn(nLabels, fanIn) #Wlabel, as in paper
Wv = 0.01 * np.random.randn(n, nWords)
WO = 0.01 * np.random.randn(n, 2 * n)
W = 0.01 * np.random.randn(n, 2 * n + 1)
#load pre-trained weights here
with open(config.saved_params_file + '_65', 'r') as loadfile:
oldrnn = cPickle.load(loadfile)
Wv[:-d, :] = oldrnn.Wv
# WO[:-d,:] = oldrnn.WO
rnn = MVRNN(W, WO, Wcat, Wv, Wo)
else:
Wo = 0.01 * np.random.randn(n + 2 * n * rank, nWords) #Lm, as in paper
Wo[:n, :] = np.ones((n, Wo.shape[1])) #Lm, as in paper
Wcat = 0.005 * np.random.randn(nLabels, fanIn) #Wlabel, as in paper
Wv = 0.01 * np.random.randn(n, nWords)
WO = 0.01 * np.random.randn(n, 2 * n)
W = 0.01 * np.random.randn(n, 2 * n + 1)
rnn = MVRNN(W, WO, Wcat, Wv, Wo)
[_, _, all_train_idx
] = getRelevantWords(rnnData_train, rnn.Wv, rnn.Wo,
params) #sets nWords_reduced, returns new arrays
params.setNumReducedWords(len(all_train_idx))
theta = rnn.getTheta(params, all_train_idx)
#optimize
print "starting training using SGD..."
nIter = 500
optimizer = StochasticGradientDescent(niter=nIter,
learning_rate=0.01,
learningrateFactor=1.0,
printAt10Iter='.',
printAt100Iter='\n+')
optimizer.minimizeBatches(rnn=rnn,
rnnData_train=rnnData_train,
allTrainSentIdx=all_train_idx,
params=params,
x0=theta,
func=costFn,
fprime=None,
rnnData_test=rnnData_dev,
initialSetSize=1,
niter=nIter,
seed=17,
modelFileName=config.saved_params_file +
'SGD_SRL',
printStatistics=True,
modelSaveIter=1,
nIterInPart=1,
nFetch=-1,
rnd=None,
nodeid=-1)
print "Finished training! "
