def train(args):
if os.path.exists("./model/vocab.bin"):
src_vocab = Vocabulary.load("./model/vocab.bin")
else:
src_vocab = Vocabulary.new(gens.word_list(args.source), args.n_vocab)
src_vocab.save('./model/vocab.bin')
if os.path.exists("./model/tag.bin"):
trg_tag = Vocabulary.load("./model/tag.bin")
else:
trg_tag = Vocabulary.new(gens.word_list(args.target), args.n_tag)
trg_tag.save('./model/tag.bin')
print("vocab_len:{}".format(src_vocab.__len__))
print("tag_len:{}".format(trg_tag.__len__))
encdec = BiEncDecLSTM(args.n_vocab, args.layer, args.embed, args.hidden,
args.n_tag)
optimizer = optimizers.Adam()
optimizer.setup(encdec)
for e_i in range(args.epoch):
tt_list = [[src_vocab.stoi(char) for char in char_arr]
for char_arr in gens.word_list(args.source_tr)]
tag_list = [
trg_tag.stoi(tag[0]) for tag in gens.word_list(args.target_tr)
]
print("{}:{}".format(len(tt_list), len(tag_list)))
assert len(tt_list) == len(tag_list)
ind_arr = [ri for ri in range(len(tt_list))]
random.shuffle(ind_arr)
tt_now = (tt_list[ri] for ri in ind_arr)
tag_now = (tag_list[ri] for ri in ind_arr)
tt_gen = gens.batch(tt_now, args.batchsize)
tag_gen = gens.batch(tag_now, args.batchsize)
for tt, tag in zip(tt_gen, tag_gen):
y_ws = encdec(tt)
teac_arr = [src_vocab.itos(t) for t in tt[0]]
pred_arr = [trg_tag.itos(y_each.data.argmax(0)) for y_each in y_ws]
print("teach:{}:{}:{}".format(teac_arr, trg_tag.itos(tag[0]),
pred_arr[0]))
tag = xp.array(tag, dtype=xp.int32)
loss = F.softmax_cross_entropy(y_ws, tag)
encdec.cleargrads()
loss.backward()
optimizer.update()
# loss.backward()
# optimizer.target.cleargrads()
# loss.backward()
# loss.unchain_backward()
# optimizer.update()
serializers.save_npz('./model/attn_tag_model_{}.npz'.format(e_i),
encdec)
def getBatchGen(self, args):
tt_now_list = [[self.vocab.stoi(char) for char in char_arr]
for char_arr in gens.word_list(args.source)]
cat_now_list = [[self.categ_vocab.stoi(cat[0])]
for cat in gens.word_list(args.category)]
ind_arr = list(range(len(tt_now_list)))
random.shuffle(ind_arr)
tt_now = (tt_now_list[ind] for ind in ind_arr)
cat_now = (cat_now_list[ind] for ind in ind_arr)
tt_gen = gens.batch(tt_now, args.batchsize)
cat_gen = gens.batch(cat_now, args.batchsize)
for tt, cat in zip(tt_gen, cat_gen):
yield (tt, cat)
def test_model(self, model_name):
trace('loading model ...')
model = self.load(model_name)
trace('generating translation ...')
generated = 0
with open(self.test_target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(self.test_source), self.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = model.predict(src_batch, self.generation_limit)
source_cuont = 0
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
# BLEUの結果を計算するため.
print("".join(src_batch[source_cuont]).replace("</s>", ""))
print(' '.join(hyp))
print(' '.join(hyp), file=fp)
source_cuont = source_cuont + 1
generated += K
trace('finished.')
def train_model(self):
trace('making vocaburaries ...')
src_vocab = Vocabulary.new(gens.word_list(self.source), self.vocab)
trg_vocab = Vocabulary.new(gens.word_list(self.target), self.vocab)
trace('making model ...')
model = self.new(src_vocab, trg_vocab, self.embed, self.hidden, self.parameter_dict)
random_number = random.randint(0, self.minibatch)
for i_epoch in range(self.epoch):
trace('epoch %d/%d: ' % (i_epoch + 1, self.epoch))
trained = 0
gen1 = gens.word_list(self.source)
gen2 = gens.word_list(self.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * self.minibatch), self.minibatch)
model.init_optimizer()
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch = model.train(src_batch, trg_batch)
if trained == 0:
self.print_out(random_number, i_epoch, trained, src_batch, trg_batch, hyp_batch)
trained += K
trace('saving model ...')
model.save("ChainerMachineTranslation" + '.%03d' % (self.epoch + 1))
trace('finished.')
def test(args):
trace('loading model ...')
src_vocab = Vocabulary.load(args.model + '.srcvocab')
trg_vocab = Vocabulary.load(args.model + '.trgvocab')
attmt = AttentionMT.load_spec(args.model + '.spec')
if args.use_gpu:
attmt.to_gpu()
serializers.load_hdf5(args.model + '.weights', attmt)
trace('generating translation ...')
generated = 0
with open(args.target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(args.source), args.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = forward(src_batch, None, src_vocab, trg_vocab, attmt, False, args.generation_limit)
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print(' '.join(hyp), file=fp)
generated += K
trace('finished.')
def train_model(args):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(args.source), args.vocab)
trg_vocab = Vocabulary.new(gens.word_list(args.target), args.vocab)
trace('making model ...')
model = EncoderDecoderModel.new(src_vocab, trg_vocab, args.embed, args.hidden)
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
trained = 0
gen1 = gens.word_list(args.source)
gen2 = gens.word_list(args.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * args.minibatch), args.minibatch)
model.init_optimizer()
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch = model.train(src_batch, trg_batch)
for k in range(K):
trace('epoch %3d/%3d, sample %8d' % (epoch + 1, args.epoch, trained + k + 1))
trace(' src = ' + ' '.join([x if x != '</s>' else '*' for x in src_batch[k]]))
trace(' trg = ' + ' '.join([x if x != '</s>' else '*' for x in trg_batch[k]]))
trace(' hyp = ' + ' '.join([x if x != '</s>' else '*' for x in hyp_batch[k]]))
trained += K
trace('saving model ...')
model.save(args.model + '.%03d' % (epoch + 1))
trace('finished.')
def test_model(self, model_name):
trace('loading model ...')
model = self.load(model_name)
trace('generating translation ...')
generated = 0
with open(self.test_target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(self.test_source),
self.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = model.predict(src_batch, self.generation_limit)
source_cuont = 0
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
# BLEUの結果を計算するため.
print("".join(src_batch[source_cuont]).replace("</s>", ""))
print(' '.join(hyp))
print(' '.join(hyp), file=fp)
source_cuont = source_cuont + 1
generated += K
trace('finished.')
def train_model(args):
trace('making vocaburaries ...')
src_vocab = Vocabulary.new(gens.word_list(args.source), args.vocab)
trg_vocab = Vocabulary.new(gens.word_list(args.target), args.vocab)
trace('making model ...')
model = EncoderDecoderModel.new(src_vocab, trg_vocab, args.embed, args.hidden)
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
trained = 0
gen1 = gens.word_list(args.source)
gen2 = gens.word_list(args.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * args.minibatch), args.minibatch)
model.init_optimizer()
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch = model.train(src_batch, trg_batch)
for k in range(K):
trace('epoch %3d/%3d, sample %8d' % (epoch + 1, args.epoch, trained + k + 1))
trace(' src = ' + ' '.join([x if x != '</s>' else '*' for x in src_batch[k]]))
trace(' trg = ' + ' '.join([x if x != '</s>' else '*' for x in trg_batch[k]]))
trace(' hyp = ' + ' '.join([x if x != '</s>' else '*' for x in hyp_batch[k]]))
trained += K
trace('saving model ...')
model.save(args.model + '.%03d' % (epoch + 1))
trace('finished.')
def train(args):
trace('loading corpus ...')
with open(args.source) as fp:
trees = [make_tree(l) for l in fp]
trace('extracting leaf nodes ...')
word_lists = [extract_words(t) for t in trees]
trace('extracting gold operations ...')
op_lists = [make_operations(t) for t in trees]
trace('making vocabulary ...')
word_vocab = Vocabulary.new(word_lists, args.vocab)
phrase_set = set()
semi_set = set()
for tree in trees:
phrase_set |= set(extract_phrase_labels(tree))
semi_set |= set(extract_semi_labels(tree))
phrase_vocab = Vocabulary.new([list(phrase_set)], len(phrase_set), add_special_tokens=False)
semi_vocab = Vocabulary.new([list(semi_set)], len(semi_set), add_special_tokens=False)
trace('converting data ...')
word_lists = [convert_word_list(x, word_vocab) for x in word_lists]
op_lists = [convert_op_list(x, phrase_vocab, semi_vocab) for x in op_lists]
trace('start training ...')
parser = Parser(
args.vocab, args.embed, args.queue, args.stack,
len(phrase_set), len(semi_set),
)
if USE_GPU:
parser.to_gpu()
opt = optimizers.AdaGrad(lr = 0.005)
opt.setup(parser)
opt.add_hook(optimizer.GradientClipping(5))
for epoch in range(args.epoch):
n = 0
for samples in batch(zip(word_lists, op_lists), args.minibatch):
parser.zerograds()
loss = my_zeros((), np.float32)
for word_list, op_list in zip(*samples):
trace('epoch %3d, sample %6d:' % (epoch + 1, n + 1))
loss += parser.forward(word_list, op_list, 0)
n += 1
loss.backward()
opt.update()
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
word_vocab.save(prefix + '.words')
phrase_vocab.save(prefix + '.phrases')
semi_vocab.save(prefix + '.semiterminals')
parser.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', parser)
trace('finished.')
def test(self):
trace('loading model ...')
src_vocab = Vocabulary.load(self.model + '.srcvocab')
trg_vocab = Vocabulary.load(self.model + '.trgvocab')
encdec = EncoderDecoder.load_spec(self.model + '.spec')
serializers.load_hdf5(self.model + '.weights', encdec)
trace('generating translation ...')
generated = 0
with open(self.target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(self.source), self.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = self.forward(src_batch, None, src_vocab, trg_vocab, encdec, False, self.generation_limit)
source_cuont = 0
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print("src : " + "".join(src_batch[source_cuont]).replace("</s>", ""))
print('hyp : ' +''.join(hyp))
print(' '.join(hyp), file=fp)
source_cuont = source_cuont + 1
generated += K
trace('finished.')
def test(self):
trace('loading model ...')
src_vocab = Vocabulary.load(self.model + '.srcvocab')
trg_vocab = Vocabulary.load(self.model + '.trgvocab')
encdec = EncoderDecoder.load_spec(self.model + '.spec')
serializers.load_hdf5(self.model + '.weights', encdec)
trace('generating translation ...')
generated = 0
with open(self.target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(self.source),
self.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = self.forward(src_batch, None, src_vocab, trg_vocab,
encdec, False, self.generation_limit)
source_cuont = 0
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print("src : " +
"".join(src_batch[source_cuont]).replace("</s>", ""))
print('hyp : ' + ''.join(hyp))
print(' '.join(hyp), file=fp)
source_cuont = source_cuont + 1
generated += K
trace('finished.')
def test_model(args):
trace('loading model ...')
model = EncoderDecoderModel.load(args.model)
trace('generating translation ...')
generated = 0
src_vectors = read_src_vectors(args.source)
src_size = len(src_vectors[0])
with open(args.target, 'w') as fp:
for src_batch in gens.batch(src_vectors, args.minibatch):
#src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = model.predict(src_batch, args.generation_limit)
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
six.print_(' '.join(hyp), file=fp)
generated += K
trace('finished.')
def train(args):
trace('loading corpus ...')
with open(args.source) as fp:
trees = [make_tree(l) for l in fp]
trace('extracting leaf nodes ...')
word_lists = [extract_words(t) for t in trees]
trace('extracting gold operations ...')
op_lists = [make_operations(t) for t in trees]
trace('making vocabulary ...')
word_vocab = Vocabulary.new(word_lists, args.vocab)
phrase_set = set()
semi_set = set()
for tree in trees:
phrase_set |= set(extract_phrase_labels(tree))
semi_set |= set(extract_semi_labels(tree))
phrase_vocab = Vocabulary.new([list(phrase_set)], len(phrase_set), add_special_tokens=False)
semi_vocab = Vocabulary.new([list(semi_set)], len(semi_set), add_special_tokens=False)
trace('converting data ...')
word_lists = [convert_word_list(x, word_vocab) for x in word_lists]
op_lists = [convert_op_list(x, phrase_vocab, semi_vocab) for x in op_lists]
trace('start training ...')
parser = Parser(
args.vocab, args.embed, args.queue, args.stack,
len(phrase_set), len(semi_set),
)
if USE_GPU:
parser.to_gpu()
opt = optimizers.AdaGrad(lr = 0.005)
opt.setup(parser)
opt.add_hook(optimizer.GradientClipping(5))
for epoch in range(args.epoch):
n = 0
for samples in batch(zip(word_lists, op_lists), args.minibatch):
parser.zerograds()
loss = my_zeros((), np.float32)
for word_list, op_list in zip(*samples):
trace('epoch %3d, sample %6d:' % (epoch + 1, n + 1))
loss += parser.forward(word_list, op_list, 0)
n += 1
loss.backward()
opt.update()
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
word_vocab.save(prefix + '.words')
phrase_vocab.save(prefix + '.phrases')
semi_vocab.save(prefix + '.semiterminals')
parser.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', parser)
trace('finished.')
def train(args):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(args.source), args.vocab)
trg_vocab = Vocabulary.new(gens.word_list(args.target), args.vocab)
trace('making model ...')
attmt = AttentionMT(args.vocab, args.embed, args.hidden)
if args.use_gpu:
attmt.to_gpu()
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
trained = 0
gen1 = gens.word_list(args.source)
gen2 = gens.word_list(args.target)
gen3 = gens.batch(
gens.sorted_parallel(gen1, gen2, 100 * args.minibatch),
args.minibatch)
opt = optimizers.AdaGrad(lr=0.01)
opt.setup(attmt)
opt.add_hook(optimizer.GradientClipping(5))
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = forward(src_batch, trg_batch, src_vocab,
trg_vocab, attmt, True, 0)
loss.backward()
opt.update()
for k in range(K):
trace('epoch %3d/%3d, sample %8d' %
(epoch + 1, args.epoch, trained + k + 1))
trace(
' src = ' +
' '.join([x if x != '</s>' else '*'
for x in src_batch[k]]))
trace(
' trg = ' +
' '.join([x if x != '</s>' else '*'
for x in trg_batch[k]]))
trace(
' hyp = ' +
' '.join([x if x != '</s>' else '*'
for x in hyp_batch[k]]))
trained += K
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
src_vocab.save(prefix + '.srcvocab')
trg_vocab.save(prefix + '.trgvocab')
attmt.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', attmt)
trace('finished.')
def getBatchGen(self, args):
tt_now_list = [[self.vocab.stoi(char) for char in char_arr]
for char_arr in gens.word_list(args.source)]
ind_arr = list(range(len(tt_now_list)))
random.shuffle(ind_arr)
tt_now = (tt_now_list[ind] for ind in ind_arr)
tt_gen = gens.batch(tt_now, args.batchsize)
for tt in tt_gen:
yield tt
def test(args, epoch):
model_name = "./model/attn_tag_model_{}.npz".format(epoch)
encdec = BiEncDecLSTM(args.n_vocab, args.layer, args.embed, args.hidden,
args.n_tag)
serializers.load_npz(model_name, encdec)
src_vocab = Vocabulary.load("./model/vocab.bin")
trg_tag = Vocabulary.load("./model/tag.bin")
tt_now = ([src_vocab.stoi(char) for char in char_arr]
for char_arr in gens.word_list(args.source_te))
tag_now = (trg_tag.stoi(tag[0]) for tag in gens.word_list(args.target_te))
tt_gen = gens.batch(tt_now, args.batchsize)
tag_gen = gens.batch(tag_now, args.batchsize)
correct_num = 0
wrong_num = 0
fw = codecs.open("./output/result_attn_tw{}.csv".format(epoch),
"w",
encoding="utf-8")
fw.write("台詞,教師キャラ,予測キャラ,予測値,›単語\n")
for tt, tag in zip(tt_gen, tag_gen):
y, att_w = encdec.callAndAtt(tt)
max_y = [
max(
F.softmax(F.reshape(y_each.data,
(1, len(y_each.data)))).data[0])
for y_each in y
]
y = [y_each.data.argmax(0) for y_each in y]
for tt_e, y_e, tag_e, max_y_e, att_w_e in zip(tt, y, tag, max_y,
att_w):
txt = ",".join([src_vocab.itos(id) for id in tt_e])
tag_e = trg_tag.itos(tag_e)
y_e = trg_tag.itos(y_e)
att_ind = att_w_e.data.argmax()
most_word = src_vocab.itos(tt_e[att_ind])
fw.write("{}:{}:{}:{}:{}\n".format(txt, tag_e, y_e, max_y_e,
most_word))
correct_num += len([1 for y_e, tag_e in zip(y, tag) if y_e == tag_e])
wrong_num += len([1 for y_e, tag_e in zip(y, tag) if y_e != tag_e])
print("epoch:{}".format(epoch))
print(" correct:{}".format(correct_num))
print(" wrong:{}".format(wrong_num))
fw.write("correct{}\n".format(correct_num))
fw.write("wrong:{}\n".format(wrong_num))
fw.close()
def train(self):
"""
Train method
If you use the word2vec model, you possible to use the copy weight
Optimizer method use the Adagrad
"""
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(self.source), self.vocab)
trg_vocab = Vocabulary.new(gens.word_list(self.target), self.vocab)
trace('making model ...')
self.attention_dialogue = AttentionDialogue(self.vocab, self.embed,
self.hidden, self.XP)
if self.word2vecFlag:
self.copy_model(self.word2vec, self.attention_dialogue.emb)
self.copy_model(self.word2vec,
self.attention_dialogue.dec,
dec_flag=True)
for epoch in range(self.epoch):
trace('epoch %d/%d: ' % (epoch + 1, self.epoch))
trained = 0
gen1 = gens.word_list(self.source)
gen2 = gens.word_list(self.target)
gen3 = gens.batch(
gens.sorted_parallel(gen1, gen2, 100 * self.minibatch),
self.minibatch)
opt = optimizers.AdaGrad(lr=0.01)
opt.setup(self.attention_dialogue)
opt.add_hook(optimizer.GradientClipping(5))
random_number = random.randint(0, self.minibatch - 1)
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = self.forward_implement(
src_batch, trg_batch, src_vocab, trg_vocab,
self.attention_dialogue, True, 0)
loss.backward()
opt.update()
self.print_out(random_number, epoch, trained, src_batch,
trg_batch, hyp_batch)
trained += K
trace('saving model ...')
prefix = self.model
model_path = APP_ROOT + "/model/" + prefix
src_vocab.save(model_path + '.srcvocab')
trg_vocab.save(model_path + '.trgvocab')
self.attention_dialogue.save_spec(model_path + '.spec')
serializers.save_hdf5(model_path + '.weights', self.attention_dialogue)
trace('finished.')
def dqn(self, s):
words = s.strip().split()
v = ([b for b in gens.batch([words], 1)])
src_batch = fill_batch(v[0])
hyp_batch = forward(src_batch, None, self.src_vocab, self.trg_vocab, self.encdec, False, self.args.generation_limit)
result = ""
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print(' '.join(hyp))
result = ' '.join(hyp)
return result
def train(self):
"""
Train method
If you use the word2vec model, you possible to use the copy weight
Optimizer method use the Adagrad
"""
trace("making vocabularies ...")
src_vocab = Vocabulary.new(gens.word_list(self.source), self.vocab)
trg_vocab = Vocabulary.new(gens.word_list(self.target), self.vocab)
trace("making model ...")
self.attention_dialogue = AttentionDialogue(self.vocab, self.embed, self.hidden, self.XP)
if self.word2vecFlag:
self.copy_model(self.word2vec, self.attention_dialogue.emb)
self.copy_model(self.word2vec, self.attention_dialogue.dec, dec_flag=True)
for epoch in range(self.epoch):
trace("epoch %d/%d: " % (epoch + 1, self.epoch))
trained = 0
gen1 = gens.word_list(self.source)
gen2 = gens.word_list(self.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * self.minibatch), self.minibatch)
opt = optimizers.AdaGrad(lr=0.01)
opt.setup(self.attention_dialogue)
opt.add_hook(optimizer.GradientClipping(5))
random_number = random.randint(0, self.minibatch - 1)
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = self.forward_implement(
src_batch, trg_batch, src_vocab, trg_vocab, self.attention_dialogue, True, 0
)
loss.backward()
opt.update()
self.print_out(random_number, epoch, trained, src_batch, trg_batch, hyp_batch)
trained += K
trace("saving model ...")
prefix = self.model
model_path = APP_ROOT + "/model/" + prefix
src_vocab.save(model_path + ".srcvocab")
trg_vocab.save(model_path + ".trgvocab")
self.attention_dialogue.save_spec(model_path + ".spec")
serializers.save_hdf5(model_path + ".weights", self.attention_dialogue)
trace("finished.")
def train(self):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(self.source), self.vocab)
trg_vocab = Vocabulary.new(gens.word_list(self.target), self.vocab)
trace('making model ...')
encdec = EncoderDecoder(self.vocab, self.embed, self.hidden)
if self.word2vecFlag:
self.copy_model(self.word2vec, encdec.enc)
self.copy_model(self.word2vec, encdec.dec, dec_flag=True)
for epoch in range(self.epoch):
trace('epoch %d/%d: ' % (epoch + 1, self.epoch))
trained = 0
gen1 = gens.word_list(self.source)
gen2 = gens.word_list(self.target)
gen3 = gens.batch(
gens.sorted_parallel(gen1, gen2, 100 * self.minibatch),
self.minibatch)
opt = optimizers.AdaGrad(lr=0.01)
opt.setup(encdec)
opt.add_hook(optimizer.GradientClipping(5))
random_number = random.randint(0, self.minibatch - 1)
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
# If you use the ipython note book you hace to use the forward function
# hyp_batch, loss = self.forward(src_batch, trg_batch, src_vocab, trg_vocab, encdec, True, 0)
hyp_batch, loss = self.forward_implement(
src_batch, trg_batch, src_vocab, trg_vocab, encdec, True,
0)
loss.backward()
opt.update()
self.print_out(random_number, epoch, trained, src_batch,
trg_batch, hyp_batch)
trained += K
trace('saving model ...')
prefix = self.model
src_vocab.save(prefix + '.srcvocab')
trg_vocab.save(prefix + '.trgvocab')
encdec.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', encdec)
trace('finished.')
def train(self):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(self.source), self.vocab)
trg_vocab = Vocabulary.new(gens.word_list(self.target), self.vocab)
trace('making model ...')
encdec = EncoderDecoder(self.vocab, self.embed, self.hidden)
if self.word2vecFlag:
self.copy_model(self.word2vec, encdec.enc)
self.copy_model(self.word2vec, encdec.dec, dec_flag=True)
else:
encdec = self.encdec
for epoch in range(self.epoch):
trace('epoch %d/%d: ' % (epoch + 1, self.epoch))
trained = 0
gen1 = gens.word_list(self.source)
gen2 = gens.word_list(self.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * self.minibatch), self.minibatch)
opt = optimizers.AdaGrad(lr = 0.01)
opt.setup(encdec)
opt.add_hook(optimizer.GradientClipping(5))
random_number = random.randint(0, self.minibatch - 1)
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = self.forward(src_batch, trg_batch, src_vocab, trg_vocab, encdec, True, 0)
loss.backward()
opt.update()
if trained == 0:
self.print_out(random_number, epoch, trained, src_batch, trg_batch, hyp_batch)
trained += K
trace('saving model ...')
prefix = self.model
src_vocab.save(prefix + '.srcvocab')
trg_vocab.save(prefix + '.trgvocab')
encdec.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', encdec)
trace('finished.')
def train(args):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(args.source), args.vocab)
trg_vocab = Vocabulary.new(gens.word_list(args.target), args.vocab)
trace('making model ...')
attmt = AttentionMT(args.vocab, args.embed, args.hidden)
if args.use_gpu:
attmt.to_gpu()
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
trained = 0
gen1 = gens.word_list(args.source)
gen2 = gens.word_list(args.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * args.minibatch), args.minibatch)
opt = optimizers.AdaGrad(lr = 0.01)
opt.setup(attmt)
opt.add_hook(optimizer.GradientClipping(5))
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = forward(src_batch, trg_batch, src_vocab, trg_vocab, attmt, True, 0)
loss.backward()
opt.update()
for k in range(K):
trace('epoch %3d/%3d, sample %8d' % (epoch + 1, args.epoch, trained + k + 1))
trace(' src = ' + ' '.join([x if x != '</s>' else '*' for x in src_batch[k]]))
trace(' trg = ' + ' '.join([x if x != '</s>' else '*' for x in trg_batch[k]]))
trace(' hyp = ' + ' '.join([x if x != '</s>' else '*' for x in hyp_batch[k]]))
trained += K
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
src_vocab.save(prefix + '.srcvocab')
trg_vocab.save(prefix + '.trgvocab')
attmt.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', attmt)
trace('finished.')
def test(self):
"""
Test method
You have to parepare the train model
"""
trace('loading model ...')
prefix = self.model
model_path = APP_ROOT + "/model/" + prefix
src_vocab = Vocabulary.load(model_path + '.srcvocab')
trg_vocab = Vocabulary.load(model_path + '.trgvocab')
self.attention_dialogue = AttentionDialogue.load_spec(
model_path + '.spec', self.XP)
serializers.load_hdf5(model_path + '.weights', self.attention_dialogue)
trace('generating translation ...')
generated = 0
with open(self.test_target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(self.source),
self.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = self.forward_implement(src_batch, None, src_vocab,
trg_vocab,
self.attention_dialogue,
False,
self.generation_limit)
source_cuont = 0
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print("src : " +
"".join(src_batch[source_cuont]).replace("</s>", ""))
print('hyp : ' + ''.join(hyp))
print(' '.join(hyp), file=fp)
source_cuont = source_cuont + 1
generated += K
trace('finished.')
def test(self):
"""
Test method
You have to parepare the train model
"""
trace("loading model ...")
prefix = self.model
model_path = APP_ROOT + "/model/" + prefix
src_vocab = Vocabulary.load(model_path + ".srcvocab")
trg_vocab = Vocabulary.load(model_path + ".trgvocab")
self.attention_dialogue = AttentionDialogue.load_spec(model_path + ".spec", self.XP)
serializers.load_hdf5(model_path + ".weights", self.attention_dialogue)
trace("generating translation ...")
generated = 0
with open(self.test_target, "w") as fp:
for src_batch in gens.batch(gens.word_list(self.source), self.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace("sample %8d - %8d ..." % (generated + 1, generated + K))
hyp_batch = self.forward_implement(
src_batch, None, src_vocab, trg_vocab, self.attention_dialogue, False, self.generation_limit
)
source_cuont = 0
for hyp in hyp_batch:
hyp.append("</s>")
hyp = hyp[: hyp.index("</s>")]
print("src : " + "".join(src_batch[source_cuont]).replace("</s>", ""))
print("hyp : " + "".join(hyp))
print(" ".join(hyp), file=fp)
source_cuont = source_cuont + 1
generated += K
trace("finished.")
def test_model(args):
trace('loading model ...')
model = AttentionalTranslationModel.load(args.model)
trace('generating translation ...')
generated = 0
with open(args.target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(args.source), args.minibatch):
src_batch = fill_batch2(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = model.predict(src_batch, args.generation_limit)
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
six.print_(' '.join(hyp), file=fp)
generated += K
trace('finished.')
def test_model(args):
trace('loading model ...')
model = EncoderDecoderModel.load(args.model)
trace('generating translation ...')
generated = 0
with open(args.target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(args.source), args.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = model.predict(src_batch, args.generation_limit)
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print(' '.join(hyp), file=fp)
generated += K
trace('finished.')
def train_model(self):
trace('making vocaburaries ...')
src_vocab = Vocabulary.new(gens.word_list(self.source), self.vocab)
trg_vocab = Vocabulary.new(gens.word_list(self.target), self.vocab)
trace('making model ...')
model = self.new(src_vocab, trg_vocab, self.embed, self.hidden,
self.parameter_dict)
random_number = random.randint(0, self.minibatch)
for i_epoch in range(self.epoch):
trace('epoch %d/%d: ' % (i_epoch + 1, self.epoch))
trained = 0
gen1 = gens.word_list(self.source)
gen2 = gens.word_list(self.target)
gen3 = gens.batch(
gens.sorted_parallel(gen1, gen2, 100 * self.minibatch),
self.minibatch)
model.init_optimizer()
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch = model.train(src_batch, trg_batch)
if trained == 0:
self.print_out(random_number, i_epoch, trained, src_batch,
trg_batch, hyp_batch)
trained += K
trace('saving model ...')
model.save("ChainerMachineTranslation" + '.%03d' %
(self.epoch + 1))
trace('finished.')
def train(args):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.input_word_list(), args.vocab)
trg_vocab = Vocabulary.new(gens.output_word_list(), args.vocab)
trace('making model ...')
encdec = EncoderDecoder(args.vocab, args.embed, args.hidden)
if args.load_model != "":
print("model load %s ... " % (args.load_model))
src_vocab = Vocabulary.load(args.load_model + '.srcvocab')
trg_vocab = Vocabulary.load(args.load_model + '.trgvocab')
encdec = EncoderDecoder.load_spec(args.load_model + '.spec')
serializers.load_hdf5(args.load_model + '.weights', encdec)
if args.use_gpu:
encdec.to_gpu()
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
trained = 0
gen1 = gens.input_word_list()
gen2 = gens.output_word_list()
gen3 = gens.batch(
gens.sorted_parallel(gen1, gen2, 100 * args.minibatch),
args.minibatch)
opt = optimizers.AdaGrad(lr=0.01)
opt.setup(encdec)
opt.add_hook(optimizer.GradientClipping(5))
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = forward(src_batch, trg_batch, src_vocab,
trg_vocab, encdec, True, 0)
loss.backward()
opt.update()
for k in range(K):
trace('epoch %3d/%3d, sample %8d' %
(epoch + 1, args.epoch, trained + k + 1))
trace(
' src = ' +
' '.join([x if x != '</s>' else '*'
for x in src_batch[k]]))
trace(
' trg = ' +
' '.join([x if x != '</s>' else '*'
for x in trg_batch[k]]))
trace(
' hyp = ' +
' '.join([x if x != '</s>' else '*'
for x in hyp_batch[k]]))
trained += K
if epoch % args.model_save_timing == 0:
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
src_vocab.save(prefix + '.srcvocab')
trg_vocab.save(prefix + '.trgvocab')
encdec.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', encdec)
trace('finished.')
def train(args):
trace('loading corpus ...')
with open(args.source) as fp:
trees = [make_tree(l) for l in fp]
trace('extracting leaf nodes ...')
word_lists = [extract_words(t) for t in trees]
lower_lists = [[w.lower() for w in words] for words in word_lists]
trace('extracting gold operations ...')
op_lists = [make_operations(t) for t in trees]
trace('making vocabulary ...')
word_vocab = Vocabulary.new(lower_lists, args.vocab)
phrase_set = set()
semiterminal_set = set()
for tree in trees:
phrase_set |= set(extract_phrase_labels(tree))
semiterminal_set |= set(extract_semiterminals(tree))
phrase_vocab = Vocabulary.new([list(phrase_set)],
len(phrase_set),
add_special_tokens=False)
semiterminal_vocab = Vocabulary.new([list(semiterminal_set)],
len(semiterminal_set),
add_special_tokens=False)
trace('converting data ...')
word_lists = [convert_word_list(x, word_vocab) for x in word_lists]
op_lists = [
convert_op_list(x, phrase_vocab, semiterminal_vocab) for x in op_lists
]
trace('start training ...')
parser = Parser(
args.vocab,
args.embed,
args.char_embed,
args.queue,
args.stack,
args.srstate,
len(phrase_set),
len(semiterminal_set),
)
if args.use_gpu:
parser.to_gpu()
opt = optimizers.SGD(lr=0.1)
opt.setup(parser)
opt.add_hook(optimizer.GradientClipping(10))
opt.add_hook(optimizer.WeightDecay(0.0001))
batch_set = list(zip(word_lists, op_lists))
for epoch in range(args.epoch):
n = 0
random.shuffle(batch_set)
for samples in batch(batch_set, args.minibatch):
parser.zerograds()
loss = XP.fzeros(())
for word_list, op_list in zip(*samples):
trace('epoch %3d, sample %6d:' % (epoch + 1, n + 1))
loss += parser.forward_train(word_list, op_list)
n += 1
loss.backward()
opt.update()
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
word_vocab.save(prefix + '.words')
phrase_vocab.save(prefix + '.phrases')
semiterminal_vocab.save(prefix + '.semiterminals')
parser.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', parser)
opt.lr *= 0.92
trace('finished.')
def train(args):
trace('loading corpus ...')
with open(args.source) as fp:
trees = [make_tree(l) for l in fp]
trace('extracting leaf nodes ...')
word_lists = [extract_words(t) for t in trees]
lower_lists = [[w.lower() for w in words] for words in word_lists]
trace('extracting gold operations ...')
op_lists = [make_operations(t) for t in trees]
trace('making vocabulary ...')
word_vocab = Vocabulary.new(lower_lists, args.vocab)
phrase_set = set()
semiterminal_set = set()
for tree in trees:
phrase_set |= set(extract_phrase_labels(tree))
semiterminal_set |= set(extract_semiterminals(tree))
phrase_vocab = Vocabulary.new([list(phrase_set)], len(phrase_set), add_special_tokens=False)
semiterminal_vocab = Vocabulary.new([list(semiterminal_set)], len(semiterminal_set), add_special_tokens=False)
trace('converting data ...')
word_lists = [to_vram_words(convert_word_list(x, word_vocab)) for x in word_lists]
op_lists = [to_vram_ops(convert_op_list(x, phrase_vocab, semiterminal_vocab)) for x in op_lists]
trace('start training ...')
parser = Parser(
args.vocab, args.embed, args.char_embed, args.queue,
args.stack, args.srstate, len(phrase_set), len(semiterminal_set),
)
if args.use_gpu:
parser.to_gpu()
opt = optimizers.SGD(lr = 0.1)
opt.setup(parser)
opt.add_hook(optimizer.GradientClipping(10))
opt.add_hook(optimizer.WeightDecay(0.0001))
batch_set = list(zip(word_lists, op_lists))
for epoch in range(args.epoch):
n = 0
random.shuffle(batch_set)
for samples in batch(batch_set, args.minibatch):
parser.zerograds()
loss = XP.fzeros(())
embed_cache = {}
for word_list, op_list in zip(*samples):
trace('epoch %3d, sample %6d:' % (epoch + 1, n + 1))
loss += parser.forward(word_list, op_list, 0, embed_cache)
n += 1
loss.backward()
opt.update()
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
word_vocab.save(prefix + '.words')
phrase_vocab.save(prefix + '.phrases')
semiterminal_vocab.save(prefix + '.semiterminals')
parser.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', parser)
opt.lr *= 0.92
trace('finished.')
