def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words)
counter = 0
try:
for eidx in xrange(params.epochs):
kf = self.get_minibatches_idx(len(data),
params.batchsize,
shuffle=True)
uidx = 0
for _, train_index in kf:
uidx += 1
batch = [data[t] for t in train_index]
for i in batch:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x,
p2mask) = self.getpairs(batch, params)
cost = self.train_function(g1x, g2x, p1x, p2x, g1mask,
g2mask, p1mask, p2mask)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
if utils.check_if_quarter(uidx, len(kf)):
if params.save:
counter += 1
self.save_params(
params.outfile + str(counter) + '.pickle',
words)
if params.evaluate:
evaluate_all(self, words)
for i in batch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if params.save:
counter += 1
self.save_params(params.outfile + str(counter) + '.pickle',
words)
if params.evaluate:
evaluate_all(self, words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words, params)
old_v = 0
try:
for eidx in range(params.epochs):
kf = self.get_minibatches_idx(len(data), params.batchsize, shuffle=True)
lkf = len(kf)
uidx = 0
sentence_samples = []
while(len(kf) > 0):
megabatch = []
idxs = []
idx = 0
for i in range(params.mb_batchsize):
if len(kf) > 0:
arr = [data[t] for t in kf[0][1]]
curr_idxs = [i + idx for i in range(len(kf[0][1]))]
kf.pop(0)
megabatch.extend(arr)
idxs.append(curr_idxs)
idx += len(curr_idxs)
uidx += len(idxs)
for i in megabatch:
if params.wordtype == "words":
if params.scramble > 0:
n = np.random.binomial(1, params.scramble, 1)[0]
if n > 0:
i[0].populate_embeddings_scramble(words)
i[1].populate_embeddings_scramble(words)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings_ngrams(words, 3, True)
i[1].populate_embeddings_ngrams(words, 3, True)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x, p2mask),(g1_s,g2_s,p1_s,p2_s) = self.get_pairs(megabatch, params)
cost = 0
for i in idxs:
# cc1,cc2 = self.cost_each_data(g1x[i], g2x[i], p1x[i], p2x[i], g1mask[i], g2mask[i], p1mask[i], p2mask[i])
cost += self.train_function(g1x[i], g2x[i], p1x[i], p2x[i], g1mask[i], g2mask[i], p1mask[i], p2mask[i])
# for j in range(len(i)):
# try:
# sentence_samples.append({'orig':g1_s[i[j]],'para':g2_s[i[j]], 'neg_orign':p1_s[i[j]], 'neg_para':p2_s[i[j]], 'orig_cost':str(cc1[j]),'para_cost':str(cc2[j])})
# except IndexError:
# print(j,i[j])
cost = cost / len(idxs)
if np.isnan(cost) or np.isinf(cost):
print('NaN detected')
if utils.check_if_quarter(uidx-len(idxs), uidx, lkf):
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle', words)
for i in megabatch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle', words)
print('Epoch ', (eidx + 1), 'Cost ', cost)
#with open("../data/sampled_samples/%s_epoch_%d.json" % (params.model, eidx),'w') as f:
# json_result = json.dumps(sentence_samples, indent=2)
# f.write(json_result)
except KeyboardInterrupt:
print("Training interupted")
end_time = time.time()
print("total time:", (end_time - start_time))
def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words, params)
old_v = 0
try:
for eidx in xrange(params.epochs):
kf = self.get_minibatches_idx(len(data),
params.batchsize,
shuffle=True)
lkf = len(kf)
uidx = 0
while (len(kf) > 0):
megabatch = []
idxs = []
idx = 0
for i in range(params.mb_batchsize):
if len(kf) > 0:
arr = [data[t] for t in kf[0][1]]
curr_idxs = [i + idx for i in range(len(kf[0][1]))]
kf.pop(0)
megabatch.extend(arr)
idxs.append(curr_idxs)
idx += len(curr_idxs)
uidx += len(idxs)
for i in megabatch:
if params.wordtype == "words":
if params.scramble > 0:
n = np.random.binomial(1, params.scramble,
1)[0]
if n > 0:
i[0].populate_embeddings_scramble(words)
i[1].populate_embeddings_scramble(words)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings_ngrams(words, 3, True)
i[1].populate_embeddings_ngrams(words, 3, True)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x,
p2mask) = self.get_pairs(megabatch, params)
cost = 0
for i in idxs:
cost += self.train_function(g1x[i], g2x[i], p1x[i],
p2x[i], g1mask[i],
g2mask[i], p1mask[i],
p2mask[i])
cost = cost / len(idxs)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
if utils.check_if_quarter(uidx - len(idxs), uidx, lkf):
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle',
words)
for i in megabatch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle', words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def train(self, data, ngram_words, word_words, params):
start_time = time.time()
evaluate_all(self, ngram_words, word_words, params)
old_v = 0
try:
for eidx in xrange(params.epochs):
kf = self.get_minibatches_idx(len(data),
params.batchsize,
shuffle=True)
lkf = len(kf)
uidx = 0
while (len(kf) > 0):
megabatch = []
idxs = []
idx = 0
for i in range(params.mb_batchsize):
if len(kf) > 0:
arr = [data[t] for t in kf[0][1]]
curr_idxs = [i + idx for i in range(len(kf[0][1]))]
kf.pop(0)
megabatch.extend(arr)
idxs.append(curr_idxs)
idx += len(curr_idxs)
uidx += len(idxs)
megabatch2 = []
for n, i in enumerate(megabatch):
example = (i[0], copy.deepcopy(i[0]), i[1],
copy.deepcopy(i[1]))
if params.combination_type == "ngram-word-lstm":
example = (i[0], copy.deepcopy(i[0]),
copy.deepcopy(i[0]), i[1],
copy.deepcopy(i[1]),
copy.deepcopy(i[1]))
if params.combination_type == "ngram-word" or params.combination_type == "ngram-lstm":
example[0].populate_embeddings_ngrams(
ngram_words, 3, True)
example[1].populate_embeddings(word_words, True)
example[2].populate_embeddings_ngrams(
ngram_words, 3, True)
example[3].populate_embeddings(word_words, True)
elif params.combination_type == "word-lstm":
example[0].populate_embeddings(word_words, True)
example[1].populate_embeddings(word_words, True)
example[2].populate_embeddings(word_words, True)
example[3].populate_embeddings(word_words, True)
elif params.combination_type == "ngram-word-lstm":
example[0].populate_embeddings_ngrams(
ngram_words, 3, True)
example[1].populate_embeddings(word_words, True)
example[2].populate_embeddings(word_words, True)
example[3].populate_embeddings_ngrams(
ngram_words, 3, True)
example[4].populate_embeddings(word_words, True)
example[5].populate_embeddings(word_words, True)
megabatch2.append(example)
megabatch = megabatch2
if params.combination_type != "ngram-word-lstm":
(g1nx, g1nmask, g1wx, g1wmask, g2nx, g2nmask, g2wx, g2wmask,
p1nx, p1nmask, p1wx, p1wmask, p2nx, p2nmask, p2wx, p2wmask) \
= self.get_pairs(megabatch, params)
else:
(g1nx, g1nmask, g1wx, g1wmask, g1lx, g1lmask, g2nx, g2nmask, g2wx, g2wmask, g2lx, g2lmask,
p1nx, p1nmask, p1wx, p1wmask, p1lx, p1lmask, p2nx, p2nmask, p2wx, p2wmask, p2lx, p2lmask) \
= self.get_pairs(megabatch, params)
cost = 0
for i in idxs:
if params.combination_type != "ngram-word-lstm":
cost += self.train_function(
g1nx[i], g2nx[i], p1nx[i], p2nx[i], g1nmask[i],
g2nmask[i], p1nmask[i], p2nmask[i], g1wx[i],
g2wx[i], p1wx[i], p2wx[i], g1wmask[i],
g2wmask[i], p1wmask[i], p2wmask[i])
else:
cost += self.train_function(
g1nx[i], g2nx[i], p1nx[i], p2nx[i], g1nmask[i],
g2nmask[i], p1nmask[i], p2nmask[i], g1wx[i],
g2wx[i], p1wx[i], p2wx[i], g1wmask[i],
g2wmask[i], p1wmask[i], p2wmask[i], g1lx[i],
g2lx[i], p1lx[i], p2lx[i], g1lmask[i],
g2lmask[i], p1lmask[i], p2lmask[i])
cost = cost / len(idxs)
if np.isnan(cost) or np.isinf(cost):
print('NaN detected')
if utils.check_if_quarter(uidx - len(idxs), uidx, lkf):
if params.evaluate:
v = evaluate_all(self, ngram_words, word_words,
params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle',
(ngram_words, word_words))
for i in megabatch:
i[0].representation = None
i[1].representation = None
i[2].representation = None
i[3].representation = None
if params.combination_type == "ngram-word-lstm":
i[4].representation = None
i[5].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
i[2].unpopulate_embeddings()
i[3].unpopulate_embeddings()
if params.combination_type == "ngram-word-lstm":
i[4].representation = None
i[5].representation = None
if params.evaluate:
v = evaluate_all(self, ngram_words, word_words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle',
(ngram_words, word_words))
print('Epoch ', (eidx + 1), 'Cost ', cost)
except KeyboardInterrupt:
print("Training interupted")
end_time = time.time()
print("total time:", (end_time - start_time))