def decode(args):
num_models = len(args.model)
models = [None for i in range(num_models)]
for i in range(num_models):
option, params = loadmodel(args.model[i])
model = rnnsearch(**option)
set_variables(model.parameter, params)
models[i] = model
# use the first model
svocabs, tvocabs = models[0].option["vocabulary"]
unk_symbol = models[0].option["unk"]
eos_symbol = models[0].option["eos"]
count = 0
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
option = {}
option["maxlen"] = args.maxlen
option["minlen"] = args.minlen
option["beamsize"] = args.beamsize
option["normalize"] = args.normalize
option["arithmetic"] = args.arithmetic
while True:
line = sys.stdin.readline()
if line == "":
break
data = [line]
seq, mask = processdata(data, svocab, unk_symbol, eos_symbol)
t1 = time.time()
tlist = beamsearch(models, seq, **option)
t2 = time.time()
if len(tlist) == 0:
sys.stdout.write("\n")
score = -10000.0
else:
best, score = tlist[0]
sys.stdout.write(" ".join(best[:-1]))
sys.stdout.write("\n")
count = count + 1
sys.stderr.write(str(count) + " ")
sys.stderr.write(str(score) + " " + str(t2 - t1) + "\n")
def decode(args):
option, params = load_model(args.model)
model = rnnsearch(**option)
set_variables(ops.trainable_variables(), params)
# use the first model
svocabs, tvocabs = model.option["vocabulary"]
unk_sym = model.option["unk"]
eos_sym = model.option["eos"]
count = 0
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
option = {}
option["maxlen"] = args.maxlen
option["minlen"] = args.minlen
option["beamsize"] = args.beamsize
option["normalize"] = args.normalize
while True:
line = sys.stdin.readline()
if line == "":
break
data = [line]
seq, mask = convert_data(data, svocab, unk_sym, eos_sym)
t1 = time.time()
tlist = beamsearch(model, seq, **option)
t2 = time.time()
if len(tlist) == 0:
translation = ""
score = -10000.0
else:
best, score = tlist[0]
translation = " ".join(best[:-1])
sys.stdout.write(translation)
sys.stdout.write("\n")
count = count + 1
sys.stderr.write(str(count) + " ")
sys.stderr.write(str(score) + " " + str(t2 - t1) + "\n")
def sample(args):
option, values = loadmodel(args.model)
model = rnnsearch(**option)
set_variables(model.parameter, values)
svocabs, tvocabs = model.option["vocabulary"]
unk_symbol = model.option["unk"]
eos_symbol = model.option["eos"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
count = 0
batch = args.batch
while True:
line = sys.stdin.readline()
if line == "":
break
data = [line]
seq, mask = processdata(data, svocab, unk_symbol, eos_symbol)
t1 = time.time()
seq = numpy.repeat(seq, batch, 1)
mask = numpy.repeat(mask, batch, 1)
tlist = batchsample(model, seq, mask, maxlen=args.maxlen)
t2 = time.time()
count = count + 1
if len(tlist) == 0:
sys.stdout.write("\n")
else:
for i in range(min(args.batch, len(tlist))):
example = tlist[i]
sys.stdout.write(" ".join(example))
sys.stdout.write("\n")
sys.stderr.write(str(count) + " " + str(t2 - t1) + "\n")
def evaluate(args):
option, params = load_model(args.model)
model = rnnsearch(**option)
var_list = ops.trainable_variables()
set_variables(var_list, params)
# use the first model
svocabs, tvocabs = model.option["vocabulary"]
unk_symbol = model.option["unk"]
eos_symbol = model.option["eos"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
if args.align:
inputs = [args.source, args.target, args.align]
else:
inputs = [args.source, args.target]
reader = textreader(inputs, False)
stream = textiterator(reader, [args.batch, args.batch])
for data in stream:
xdata, xmask = convert_data(data[0], svocab, unk_symbol, eos_symbol)
ydata, ymask = convert_data(data[1], tvocab, unk_symbol, eos_symbol)
if not args.align:
align = None
else:
align = convert_align(data[0], data[1], data[2])
cost = evaluate_model(model, xdata, xmask, ydata, ymask, align,
verbose=args.verbose)
for i in range(len(cost)):
if args.verbose:
sys.stdout.write("src: %s\n" % data[0][i])
sys.stdout.write("tgt: %s\n" % data[1][i])
sys.stdout.write("cost: %f\n" % cost[i])
stream.close()
def replace(args):
num_models = len(args.model)
models = [None for i in range(num_models)]
alignments = [None for i in range(num_models)]
if args.dictionary:
mapping = load_dictionary(args.dictionary)
heuristic = args.heuristic
else:
heuristic = 0
for i in range(num_models):
option, params = loadmodel(args.model[i])
model = rnnsearch(**option)
set_variables(model.parameter, params)
models[i] = model
# use the first model
svocabs, tvocabs = models[0].option["vocabulary"]
unk_symbol = models[0].option["unk"]
eos_symbol = models[0].option["eos"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
reader = textreader(args.text, False)
stream = textiterator(reader, [args.batch, args.batch])
for data in stream:
xdata, xmask = processdata(data[0], svocab, unk_symbol, eos_symbol)
ydata, ymask = processdata(data[1], tvocab, unk_symbol, eos_symbol)
for i in range(num_models):
# compute attention score
alignments[i] = models[i].attention(xdata, xmask, ydata, ymask)
# ensemble, alignment: yseq * xseq * batch
if args.arithmetic:
alignment = sum(alignments) / num_models
else:
alignments = map(numpy.log, alignments)
alignment = numpy.exp(sum(alignments) / num_models)
# find source word to which each target word was most aligned
indices = numpy.argmax(alignment, 1)
# write to output
for i in range(len(data[1])):
source_words = data[0][i].strip().split()
target_words = data[1][i].strip().split()
translation = []
for j in range(len(target_words)):
source_length = len(source_words)
word = target_words[j]
# found unk symbol
if word == unk_symbol:
source_index = indices[j, i]
if source_index >= source_length:
translation.append(word)
continue
source_word = source_words[source_index]
if heuristic and source_word in mapping:
if heuristic == 1:
translation.append(mapping[source_word])
else:
# source word begin with lower case letter
if source_word.decode('utf-8')[0].islower():
translation.append(mapping[source_word])
else:
translation.append(source_word)
else:
translation.append(source_word)
else:
translation.append(word)
sys.stdout.write(" ".join(translation))
sys.stdout.write("\n")
stream.close()
def train(args):
option = getoption()
if os.path.exists(args.model):
option, params = loadmodel(args.model)
init = False
else:
init = True
override(option, args)
print_option(option)
# set seed
numpy.random.seed(option["seed"])
if option["ref"]:
references = loadreferences(option["ref"])
else:
references = None
svocabs, tvocabs = option["vocabulary"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
pathname, basename = os.path.split(args.model)
modelname = getfilename(basename)
autoname = os.path.join(pathname, modelname + ".autosave.pkl")
bestname = os.path.join(pathname, modelname + ".best.pkl")
batch = option["batch"]
sortk = option["sort"] or 1
shuffle = option["seed"] if option["shuffle"] else None
reader = textreader(option["corpus"], shuffle)
processor = [getlen, getlen]
stream = textiterator(reader, [batch, batch * sortk], processor,
option["limit"], option["sort"])
if shuffle and option["indices"] is not None:
reader.set_indices(option["indices"])
if args.reset:
option["count"] = [0, 0]
option["epoch"] = 0
option["cost"] = 0.0
skipstream(reader, option["count"][1])
epoch = option["epoch"]
maxepoch = option["maxepoch"]
model = rnnsearch(**option)
if init:
uniform(model.parameter, -0.08, 0.08)
else:
set_variables(model.parameter, params)
print "parameters:", parameters(model.parameter)
# tuning option
toption = {}
toption["algorithm"] = option["optimizer"]
toption["variant"] = option["variant"]
toption["constraint"] = ("norm", option["norm"])
toption["norm"] = True
trainer = optimizer(model, **toption)
alpha = option["alpha"]
# beamsearch option
doption = {}
doption["beamsize"] = option["beamsize"]
doption["normalize"] = option["normalize"]
doption["maxlen"] = option["maxlen"]
doption["minlen"] = option["minlen"]
best_score = option["bleu"]
unk_sym = option["unk"]
eos_sym = option["eos"]
count = option["count"][0]
totcost = option["cost"]
for i in range(epoch, maxepoch):
for data in stream:
xdata, xmask = processdata(data[0], svocab, unk_sym, eos_sym)
ydata, ymask = processdata(data[1], tvocab, unk_sym, eos_sym)
t1 = time.time()
cost, norm = trainer.optimize(xdata, xmask, ydata, ymask)
trainer.update(alpha=alpha)
t2 = time.time()
count += 1
cost = cost * ymask.shape[1] / ymask.sum()
totcost += cost / math.log(2)
print i + 1, count, cost, norm, t2 - t1
# autosave
if count % option["freq"] == 0:
model.option["indices"] = reader.get_indices()
model.option["bleu"] = best_score
model.option["cost"] = totcost
model.option["count"] = [count, reader.count]
serialize(autoname, model)
if count % option["vfreq"] == 0:
if option["validation"] and references:
trans = translate(model, option["validation"], **doption)
bleu_score = bleu(trans, references)
print "bleu: %2.4f" % bleu_score
if bleu_score > best_score:
best_score = bleu_score
model.option["indices"] = reader.get_indices()
model.option["bleu"] = best_score
model.option["cost"] = totcost
model.option["count"] = [count, reader.count]
serialize(bestname, model)
if count % option["sfreq"] == 0:
n = len(data[0])
ind = numpy.random.randint(0, n)
sdata = data[0][ind]
tdata = data[1][ind]
xdata = xdata[:, ind:ind + 1]
hls = beamsearch(model, xdata)
if len(hls) > 0:
best, score = hls[0]
print sdata
print tdata
print " ".join(best[:-1])
else:
print sdata
print tdata
print "warning: no translation"
print "--------------------------------------------------"
if option["validation"] and references:
trans = translate(model, option["validation"], **doption)
bleu_score = bleu(trans, references)
print "iter: %d, bleu: %2.4f" % (i + 1, bleu_score)
if bleu_score > best_score:
best_score = bleu_score
model.option["indices"] = reader.get_indices()
model.option["bleu"] = best_score
model.option["cost"] = totcost
model.option["count"] = [count, reader.count]
serialize(bestname, model)
print "averaged cost: ", totcost / count
print "--------------------------------------------------"
# early stopping
if i + 1 >= option["stop"]:
alpha = alpha * option["decay"]
count = 0
totcost = 0.0
stream.reset()
# update autosave
model.option["epoch"] = i + 1
model.option["alpha"] = alpha
model.option["indices"] = reader.get_indices()
model.option["bleu"] = best_score
model.option["cost"] = totcost
model.option["count"] = [0, 0]
serialize(autoname, model)
print "best(bleu): %2.4f" % best_score
stream.close()
def decode(args):
num_models = len(args.model)
models = [None for i in range(num_models)]
for i in range(num_models):
option, params = load_model(args.model[i])
scope = "rnnsearch_%d" % i
model = rnnsearch(scope=scope, **option)
var_list = get_variables_with_prefix(scope)
set_variables(var_list, params)
models[i] = model
# use the first model
svocabs, tvocabs = models[0].option["vocabulary"]
unk_sym = models[0].option["unk"]
eos_sym = models[0].option["eos"]
count = 0
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
option = {}
option["maxlen"] = args.maxlen
option["minlen"] = args.minlen
option["beamsize"] = args.beamsize
option["normalize"] = args.normalize
option["arithmetic"] = args.arithmetic
if args.oracle:
references = load_references(args.oracle)
else:
references = None
while True:
line = sys.stdin.readline()
if line == "":
break
data = [line]
seq, mask = convert_data(data, svocab, unk_sym, eos_sym)
t1 = time.time()
tlist = beamsearch(models, seq, **option)
t2 = time.time()
if len(tlist) == 0:
translation = ""
score = -10000.0
else:
if references is None:
best, score = tlist[0]
translation = " ".join(best[:-1])
sys.stdout.write(translation)
sys.stdout.write("\n")
else:
best_ind = 0
best_score = 0
# find the best translation according to oracle
for i, (trans, score) in enumerate(tlist):
trans = trans[:-1]
bleu_score = bleu([trans], [references[count]],
smoothing=True)
if bleu_score > best_score:
best_score = bleu_score
best_ind = i
output = " ".join(tlist[0][0][:-1]) + " ||| "
output += str(tlist[0][1]) + " ||| " + str(best_ind) + " ||| "
output += " ".join(tlist[best_ind][0][:-1]) + " ||| "
output += str(tlist[best_ind][1])
sys.stdout.write(output)
sys.stdout.write("\n")
count = count + 1
sys.stderr.write(str(count) + " ")
sys.stderr.write(str(score) + " " + str(t2 - t1) + "\n")
def train(args):
option = default_option()
# predefined model names
pathname, basename = os.path.split(args.model)
modelname = get_filename(basename)
autoname = os.path.join(pathname, modelname + ".autosave.pkl")
bestname = os.path.join(pathname, modelname + ".best.pkl")
# load models
if os.path.exists(args.model):
opt, params = load_model(args.model)
option = opt
init = False
else:
init = True
if args.initialize:
init_params = load_model(args.initialize)
init_params = init_params[1]
restore = True
else:
restore = False
override(option, args_to_dict(args))
print_option(option)
# load references
if option["references"]:
references = load_references(option["references"])
else:
references = None
if args.skip_val:
references = None
criterion = option["criterion"]
if criterion == "mrt":
sys.stderr.write("warning: In MRT mode, batch is set to 1\n")
# input corpus
batch = option["batch"] if criterion == "mle" else 1
sortk = option["sort"] or 1 if criterion == "mle" else 1
shuffle = option["seed"] if option["shuffle"] else None
reader = textreader(option["corpus"], shuffle)
processor = [data_length, data_length]
stream = textiterator(reader, [batch, batch * sortk], processor,
option["limit"], option["sort"])
if shuffle and option["indices"] is not None:
reader.set_indices(option["indices"])
if args.reset:
option["count"] = [0, 0]
option["epoch"] = 0
option["cost"] = 0.0
skip_stream(reader, option["count"][1])
epoch = option["epoch"]
maxepoch = option["maxepoch"]
# create model
regularizer = []
if option["l1_scale"]:
regularizer.append(ops.l1_regularizer(option["l1_scale"]))
if option["l2_scale"]:
regularizer.append(ops.l2_regularizer(option["l2_scale"]))
scale = option["scale"]
initializer = ops.random_uniform_initializer(-scale, scale)
regularizer = ops.sum_regularizer(regularizer)
# set seed
numpy.random.seed(option["seed"])
model = rnnsearch(initializer=initializer, regularizer=regularizer,
**option)
variables = None
if restore:
matched, not_matched = match_variables(ops.trainable_variables(),
init_params)
if args.finetune:
variables = not_matched
if not variables:
raise RuntimeError("no variables to finetune")
if not init:
set_variables(ops.trainable_variables(), params)
if restore:
restore_variables(matched, not_matched)
print "parameters:", count_parameters(ops.trainable_variables())
# tuning option
tune_opt = {}
tune_opt["algorithm"] = option["optimizer"]
tune_opt["constraint"] = ("norm", option["norm"])
tune_opt["norm"] = True
tune_opt["variables"] = variables
# create optimizer
trainer = optimizer(model, **tune_opt)
# beamsearch option
search_opt = {}
search_opt["beamsize"] = option["beamsize"]
search_opt["normalize"] = option["normalize"]
search_opt["maxlen"] = option["maxlen"]
search_opt["minlen"] = option["minlen"]
# vocabulary and special symbol
svocabs, tvocabs = option["vocabulary"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
unk_sym = option["unk"]
eos_sym = option["eos"]
# summary
count = option["count"][0]
totcost = option["cost"]
best_score = option["bleu"]
alpha = option["alpha"]
sharp = option["sharp"]
for i in range(epoch, maxepoch):
for data in stream:
xdata, xmask = convert_data(data[0], svocab, unk_sym, eos_sym)
ydata, ymask = convert_data(data[1], tvocab, unk_sym, eos_sym)
if criterion == "mrt":
refs = []
for item in data[1]:
item = item.split()
item = [unk_sym if word not in tvocab else word
for word in item]
refs.append(" ".join(item))
t1 = time.time()
# sample from model
nsample = option["sample"] - len(refs)
xdata = numpy.repeat(xdata, nsample, 1)
xmask = numpy.repeat(xmask, nsample, 1)
maxlen = int(1.5 * len(ydata))
examples = batchsample(model, xdata, xmask, maxlen)
space = build_sample_space(refs, examples)
score = numpy.zeros((len(space),), "float32")
refs = [ref.split() for ref in refs]
for j in range(len(space)):
example = space[j].split()
score[j] = 1.0 - bleu([example], [refs], smoothing=True)
ydata, ymask = convert_data(space, tvocab, unk_sym, eos_sym)
cost, norm = trainer.optimize(xdata[:, 0:1], xmask[:, 0:1],
ydata, ymask, score, sharp)
trainer.update(alpha=alpha)
t2 = time.time()
totcost += cost
count += 1
t = t2 - t1
ac = totcost / count
print i + 1, count, len(space), cost, norm, ac, t
else:
t1 = time.time()
cost, norm = trainer.optimize(xdata, xmask, ydata, ymask)
trainer.update(alpha = alpha)
t2 = time.time()
count += 1
cost = cost * ymask.shape[1] / ymask.sum()
totcost += cost / math.log(2)
print i + 1, count, cost, norm, t2 - t1
# autosave
if count % option["freq"] == 0:
option["indices"] = reader.get_indices()
option["bleu"] = best_score
option["cost"] = totcost
option["count"] = [count, reader.count]
serialize(autoname, option)
if count % option["vfreq"] == 0:
if option["validation"] and references:
trans = translate(model, option["validation"],
**search_opt)
bleu_score = bleu(trans, references)
print "bleu: %2.4f" % bleu_score
if bleu_score > best_score:
best_score = bleu_score
option["indices"] = reader.get_indices()
option["bleu"] = best_score
option["cost"] = totcost
option["count"] = [count, reader.count]
serialize(bestname, option)
if count % option["sfreq"] == 0:
n = len(data[0])
ind = numpy.random.randint(0, n)
sdata = data[0][ind]
tdata = data[1][ind]
xdata = xdata[:, ind : ind + 1]
xmask = xmask[:, ind : ind + 1]
hls = beamsearch(model, xdata, xmask)
best, score = hls[0]
print sdata
print tdata
print " ".join(best[:-1])
print "--------------------------------------------------"
if option["validation"] and references:
trans = translate(model, option["validation"], **search_opt)
bleu_score = bleu(trans, references)
print "iter: %d, bleu: %2.4f" % (i + 1, bleu_score)
if bleu_score > best_score:
best_score = bleu_score
option["indices"] = reader.get_indices()
option["bleu"] = best_score
option["cost"] = totcost
option["count"] = [count, reader.count]
serialize(bestname, option)
print "averaged cost: ", totcost / count
print "--------------------------------------------------"
# early stopping
if i + 1 >= option["stop"]:
alpha = alpha * option["decay"]
count = 0
totcost = 0.0
stream.reset()
# update autosave
option["epoch"] = i + 1
option["alpha"] = alpha
option["indices"] = reader.get_indices()
option["bleu"] = best_score
option["cost"] = totcost
option["count"] = [0, 0]
serialize(autoname, option)
print "best(bleu): %2.4f" % best_score
stream.close()