def evaluate(args):
option, params = load_model(args.model)
model = build_model(**option)
var_list = ops.trainable_variables()
set_variables(var_list, params)
if args.sntcost:
for costs in evaluate_snt_cost(model, option, args.source, args.target,
args.batch, args.normalize):
for cost in costs:
sys.stdout.write("{}\n".format(cost))
else:
# 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[0], args.align]
else:
inputs = [args.source, args.target[0]]
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 evaluate_snt_cost(model, option, src, refs, batch, normalize):
svocabs, tvocabs = option["vocabulary"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
unk_sym = option["unk"]
eos_sym = option["eos"]
reader = textreader([src] + refs)
stream = textiterator(reader, [batch, batch])
get_cost = model.get_snt_cost
for data in stream:
xdata, xmask = convert_data(data[0], svocab, unk_sym, eos_sym)
for i, y in enumerate(data[1:]):
ydata, ymask = convert_data(y, tvocab, unk_sym, eos_sym)
snt_cost = get_cost(xdata, xmask, ydata, ymask)
if normalize:
# per word cost
lens = numpy.array([len(item) for item in y])
snt_cost /= lens
yield snt_cost
def train(args):
option = default_option()
# predefined model names
pathname, basename = os.path.split(args.model)
modelname = get_filename(basename)
autoname_format = os.path.join(pathname,
modelname + ".iter{epoch}-{batch}.pkl")
bestname = os.path.join(pathname, modelname + ".best.pkl")
# load models
if os.path.exists(args.model):
opt, params = load_model(args.model)
override(option, opt)
init = False
else:
init = True
if args.initialize:
pretrain_params = load_model(args.initialize)
pretrain_params = pretrain_params[1]
pretrain = True
else:
pretrain = False
override(option, args_to_dict(args))
# check external validation script
ext_val_script = option['ext_val_script']
if not os.path.exists(ext_val_script):
raise ValueError('File doesn\'t exist: %s' % ext_val_script)
elif not os.access(ext_val_script, os.X_OK):
raise ValueError('File is not executable: %s' % ext_val_script)
# check references format
ref_stem = None
if option['validation'] and option['references']:
ref_stem = misc.infer_ref_stem([option['validation']],
option['references'])
ref_stem = ref_stem[0]
# .yaml for ultimate options
yaml_name = "%s.settings.yaml" % modelname
if init or not os.path.exists(yaml_name):
with open(yaml_name, "w") as w:
_opt = args.__dict__.copy()
for k, v in _opt.iteritems():
if k in option:
_opt[k] = option[k]
yaml.dump(_opt, w, default_flow_style=False)
del _opt
print_option(option)
# reader
batch = option["batch"]
sortk = option["sort"]
shuffle = option["shuffle"]
reader = textreader(option["corpus"], shuffle)
processor = [data_length, data_length]
stream = textiterator(reader, [batch, batch * sortk], processor,
option["limit"], option["sort"])
# progress
# initialize before building model
progress = Progress(option["delay_val"], stream, option["seed"])
# 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)
option["scope"] = "rnnsearch"
model = build_model(initializer=initializer,
regularizer=regularizer,
**option)
variables = None
if pretrain:
print "using pretrain"
_pp1 = {}
for name, val in pretrain_params:
names = name.split('/')[1:]
if "embedding" in names[0]:
_pp1['/'.join(names)] = val
else:
_pp1['/'.join(names[1:])] = val
matched = []
not_matched = []
for var in ops.trainable_variables():
names = var.name.split('/')[1:]
if "decoder2" in var.name:
not_matched.append((var.name, var.get_value().size))
continue
if "embedding" in names[0]:
match_name = '/'.join(names)
var.set_value(_pp1[match_name])
else:
match_name = '/'.join(names[1:])
var.set_value(_pp1[match_name])
matched.append((var.name, var.get_value().size))
print "------------------- matched -------------------"
for name, size in matched:
print name, size
print "------------------- not matched -------------------"
for name, size in not_matched:
print name, size
print "------------------- end -------------------\n"
if not init:
set_variables(ops.trainable_variables(), params)
print "parameters: %d\n" % 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
scopes = [".*"]
trainer = optimizer(model.inputs, model.outputs, model.cost, scopes,
**tune_opt)
# vocabulary and special symbol
svocabs, tvocabs = option["vocabulary"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
unk_sym = option["unk"]
eos_sym = option["eos"]
alpha = option["alpha"]
maxepoch = option["maxepoch"]
# restore right before training to avoid randomness changing when trying to resume progress
if not args.reset:
if "#progress" in option:
print 'Restore progress >>'
progress = (option["#progress"])
stream = progress.iterator
stream.set_processor(processor)
for ttt in progress.task_manager.tasks:
ttt.status = 4
ttt.result = 0.0
else:
print 'New progress >>'
else:
print 'Discard progress >>'
# setup progress
progress.oldname = args.model
progress.serializer = serialize
stream = progress.iterator
overwrite = not args.no_overwrite
if progress.task_manager:
print progress.task_manager
try:
while progress.epoch < maxepoch:
epc = progress.epoch
for data in stream:
progress.tic()
if progress.failed():
raise RuntimeError("progress failure")
xdata, xmask = convert_data(data[0], svocab, unk_sym, eos_sym)
ydata, ymask = convert_data(data[1], tvocab, unk_sym, eos_sym)
bydata, _ = convert_data(data[1], tvocab, unk_sym, eos_sym,
True)
t1 = time.time()
tot_cost, soft_cost, true_cost, norm = trainer.optimize(
xdata, xmask, ydata, ymask, bydata)
trainer.update(alpha=alpha)
t2 = time.time()
# per word cost
w_cost = true_cost * ymask.shape[1] / ymask.sum()
progress.batch_count += 1
progress.batch_total += 1
progress.loss_hist.append(w_cost)
count = progress.batch_count
if not args.pfreq or count % args.pfreq == 0:
print epc + 1, progress.batch_count, w_cost, tot_cost, soft_cost, true_cost, norm, t2 - t1
if count % option["vfreq"] == 0 and not should_skip_val(
args.skip_val, option["vfreq"], epc,
progress.batch_total):
if option["validation"] and option["references"]:
progress.add_valid(option['scope'],
option['validation'], ref_stem,
ext_val_script, __file__, option,
modelname, bestname, serialize)
# save after validation
progress.toc()
if count % option["freq"] == 0:
progress.save(option, autoname_format, overwrite)
progress.tic()
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])
progress.toc()
print "--------------------------------------------------"
progress.tic()
if option["validation"] and option["references"]:
progress.add_valid(option['scope'], option['validation'],
ref_stem, ext_val_script, __file__, option,
modelname, bestname, serialize)
print "--------------------------------------------------"
progress.toc()
# early stopping
if epc + 1 >= option["stop"]:
alpha = alpha * option["decay"]
stream.reset()
progress.epoch += 1
progress.batch_count = 0
# update autosave
option["alpha"] = alpha
progress.save(option, autoname_format, overwrite)
stream.close()
progress.tic()
print "syncing ..."
progress.barrier() # hangup and wait
progress.toc()
best_valid = max(progress.valid_hist, key=lambda item: item[1])
(epc, count), score = best_valid
print "best bleu {}-{}: {:.4f}".format(epc + 1, count, score)
if progress.delay_val:
task_elapse = sum(
[task.elapse for task in progress.task_manager.tasks])
print "training finished in {}({})".format(
datetime.timedelta(seconds=int(progress.elapse)),
datetime.timedelta(seconds=int(progress.elapse + task_elapse)))
else:
print "training finished in {}".format(
datetime.timedelta(seconds=int(progress.elapse)))
progress.save(option, autoname_format, overwrite)
except KeyboardInterrupt:
traceback.print_exc()
progress.terminate()
sys.exit(1)
except Exception:
traceback.print_exc()
progress.terminate()
sys.exit(1)
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 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:
if args.heuristic > 0:
raise ValueError("heuristic > 0, but no dictionary available")
heuristic = 0
for i in range(num_models):
option, params = load_model(args.model[i])
scope = "rnnsearch_%d" % i
option["scope"] = scope
model = build_model(**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_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 = convert_data(data[0], svocab, unk_symbol, eos_symbol)
ydata, ymask = convert_data(data[1], tvocab, unk_symbol, eos_symbol)
for i in range(num_models):
# compute attention score
alignments[i] = models[i].align(xdata, xmask, ydata, ymask)
# ensemble, alignment: tgt_len * src_len * 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()
if "indices" in vals:
option["indices"] = vals["indices"]
return option, params
def set_variables(params, values):
for p, v in zip(params, values):
p.set_value(v)
def loadreferences(names, case=True):
references = []
reader = textreader(names)
stream = textiterator(reader, size=[1, 1])
for data in stream:
newdata = []
for batch in data:
line = batch[0]
words = line.strip().split()
if not case:
lower = [word.lower() for word in words]
newdata.append(lower)
else:
newdata.append(words)
references.append(newdata)
stream.close()
def train(args):
option = default_option()
# predefined model names
pathname, basename = os.path.split(args.model)
modelname = get_filename(basename)
autoname_format = os.path.join(pathname, modelname + ".iter{epoch}-{batch}.pkl")
bestname = os.path.join(pathname, modelname + ".best.pkl")
# load models
if os.path.exists(args.model):
opt, params = load_model(args.model)
override(option, opt)
init = False
else:
init = True
if args.initialize:
print "initialize:", args.initialize
pretrain_params = load_model(args.initialize)
pretrain_params = pretrain_params[1]
pretrain = True
else:
pretrain = False
override(option, args_to_dict(args))
# check external validation script
ext_val_script = option['ext_val_script']
if not os.path.exists(ext_val_script):
raise ValueError('File doesn\'t exist: %s' % ext_val_script)
elif not os.access(ext_val_script, os.X_OK):
raise ValueError('File is not executable: %s' % ext_val_script)
# check references format
ref_stem = option['references']
if option['validation'] and option['references']:
ref_stem = misc.infer_ref_stem([option['validation']], option['references'])
ref_stem = ref_stem[0]
# .yaml for ultimate options
yaml_name = "%s.settings.yaml" % modelname
if init or not os.path.exists(yaml_name):
with open(yaml_name, "w") as w:
_opt = args.__dict__.copy()
for k, v in _opt.iteritems():
if k in option:
_opt[k] = option[k]
yaml.dump(_opt, w,
default_flow_style=False)
del _opt
print_option(option)
# reader
batch = option["batch"]
sortk = option["sort"]
shuffle = option["shuffle"]
reader = textreader(option["corpus"][:3], shuffle)
processor = [data_length, data_length, data_length]
stream = textiterator(reader, [batch, batch * sortk], processor,
option["limit"], option["sort"])
reader = textreader(option["corpus"][3:], shuffle)
processor = [data_length, data_length, data_length]
dstream = textiterator(reader, [batch, batch * sortk], processor,
None, option["sort"])
# progress
# initialize before building model
progress = Progress(option["delay_val"], stream, option["seed"])
# 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)
option["scope"] = "rnnsearch"
model = build_model(initializer=initializer, regularizer=regularizer,
**option)
variables = None
if pretrain:
matched, not_matched = match_variables(ops.trainable_variables(),
pretrain_params)
if args.finetune:
variables = not_matched
if not variables:
raise RuntimeError("no variables to finetune")
if pretrain:
restore_variables(matched, not_matched)
if not init:
set_variables(ops.trainable_variables(), params)
print "parameters: %d\n" % 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
scopes = ["((?!Shared).)*$"]
trainer = optimizer(model.inputs, model.outputs, model.cost, scopes, **tune_opt)
clascopes = [".*(Shared).*"]
clatrainer = optimizer(model.inputs_cla, model.outputs_cla, model.cost_cla, clascopes, **tune_opt)
#scopes = [".*(DSAenc).*"]
#domain_trainer = optimizer(model.inputs, model.toutputs, model.domaincost, scopes, **tune_opt)
# vocabulary and special symbol
svocabs, tvocabs = option["vocabulary"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
unk_sym = option["unk"]
eos_sym = option["eos"]
alpha = option["alpha"]
maxepoch = option["maxepoch"]
# restore right before training to avoid randomness changing when trying to resume progress
if not args.reset:
if "#progress" in option:
print 'Restore progress >>'
progress = (option["#progress"])
stream = progress.iterator
stream.set_processor(processor)
else:
print 'New progress >>'
else:
print 'Discard progress >>'
if args.drop_tasks:
print 'drop tasks'
progress.drop_tasks()
# setup progress
progress.oldname = args.model
progress.serializer = serialize
stream = progress.iterator
overwrite = not args.no_overwrite
if progress.task_manager:
print progress.task_manager
register_killer()
tagvocab = {}
for idx, d in enumerate(option["dvocab"]):
tagvocab[d] = idx
if len(tagvocab) != option["dnum"]:
raise ValueError('length of domain vocab %f not equal to domain num %f!' % (len(tagvocab), option["dnum"]))
try:
while progress.epoch < maxepoch:
epc = progress.epoch
for data in stream:
progress.tic()
if progress.failed():
raise RuntimeError("progress failure")
# data = _stream.next()
xdata, xmask = convert_data(data[0], svocab, unk_sym, eos_sym)
ydata, ymask = convert_data(data[1], tvocab, unk_sym, eos_sym)
tag = convert_tag(data[2], tagvocab)
t1 = time.time()
cost, dcost, scost, tdcost, norm = trainer.optimize(xdata, xmask, ydata, ymask, tag)
clacost, _ = clatrainer.optimize(xdata, xmask, tag)
trainer.update(alpha=alpha)
clatrainer.update(alpha=alpha)
t2 = time.time()
# per word cost
w_cost = cost * ymask.shape[1] / ymask.sum()
progress.batch_count += 1
progress.batch_total += 1
progress.loss_hist.append(w_cost)
if not args.pfreq or count % args.pfreq == 0:
print epc + 1, progress.batch_count, w_cost, dcost, tdcost, scost, clacost, norm, t2 - t1
count = progress.batch_count
if count % option["sfreq"] == 0:
dright = 0.0
sright = 0.0
tdright = 0.0
total = 0.0
for ddata in dstream:
txdata, txmask = convert_data(ddata[0], svocab, unk_sym, eos_sym)
tydata, tymask = convert_data(ddata[1], tvocab, unk_sym, eos_sym)
txtag = convert_tag(ddata[2], tagvocab)
dtag_pred, stag_pred = model.tag_predict(txdata, txmask)
txtag = txtag[0]
dpretag = []
for i in dtag_pred:
dpretag.append(int(i))
spretag = []
for i in stag_pred:
spretag.append(int(i))
tdtag_pred = model.tgt_tag_predict(txdata, txmask, tydata, tymask)
tdpretag = []
for i in tdtag_pred[0]:
tdpretag.append(int(i))
dright = dright + sum([m == n for m, n in zip(txtag, dpretag)])
sright = sright + sum([m == n for m, n in zip(txtag, spretag)])
tdright = tdright + sum([m == n for m, n in zip(txtag, tdpretag)])
total = total + len(dpretag)
dstream.reset()
dacc = dright * 1.0 / total
sacc = sright * 1.0 / total
tdacc = tdright * 1.0 / total
print "dacc:", dright, dacc
print "sacc", sright, sacc
print "tdacc", tdright, tdacc
if count % option["vfreq"] == 0 and not should_skip_val(args.skip_val, option["vfreq"], epc,
progress.batch_total):
if option["validation"] and option["references"]:
progress.add_valid(option['scope'], option['validation'], ref_stem, ext_val_script, __file__,
option, modelname, bestname, serialize)
# save after validation
progress.toc()
if count % option["freq"] == 0:
progress.save(option, autoname_format, overwrite)
progress.tic()
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])
progress.toc()
print "--------------------------------------------------"
progress.tic()
if option["validation"] and option["references"]:
progress.add_valid(option['scope'], option['validation'], ref_stem, ext_val_script, __file__, option,
modelname, bestname, serialize)
print "--------------------------------------------------"
progress.toc()
print "epoch cost {}".format(numpy.mean(progress.loss_hist))
progress.loss_hist = []
# early stopping
if epc + 1 >= option["stop"]:
alpha = alpha * option["decay"]
stream.reset()
progress.epoch += 1
progress.batch_count = 0
# update autosave
option["alpha"] = alpha
progress.save(option, autoname_format, overwrite)
stream.close()
progress.tic()
print "syncing ..."
progress.barrier() # hangup and wait
progress.toc()
best_valid = max(progress.valid_hist, key=lambda item: item[1])
(epc, count), score = best_valid
print "best bleu {}-{}: {:.4f}".format(epc + 1, count, score)
if progress.delay_val:
task_elapse = sum([task.elapse for task in progress.task_manager.tasks])
print "training finished in {}({})".format(datetime.timedelta(seconds=int(progress.elapse)),
datetime.timedelta(seconds=int(progress.elapse + task_elapse)))
else:
print "training finished in {}".format(datetime.timedelta(seconds=int(progress.elapse)))
progress.save(option, autoname_format, overwrite)
except KeyboardInterrupt:
traceback.print_exc()
progress.terminate()
sys.exit(1)
except Exception:
traceback.print_exc()
progress.terminate()
sys.exit(1)
def rescore(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
option['scope'] = scope
model = build_model(**option)
var_list = get_variables_with_prefix(scope)
set_variables(var_list, params)
models[i] = model
with open(args.source) as source_file:
lines = source_file.readlines()
if args.n_best_list:
with open(args.n_best_list) as nbest_file:
nbest_lines = nbest_file.readlines()
else:
nbest_lines = sys.stdin.readlines()
scores = []
for model in models:
with tempfile.NamedTemporaryFile() as src, tempfile.NamedTemporaryFile() as tgt:
for line in nbest_lines:
linesplit = line.split(' ||| ')
idx = int(linesplit[0]) ##index from the source file. Starting from 0.
src.write(lines[idx])
tgt.write(linesplit[1] + '\n')
src.seek(0)
tgt.seek(0)
option = model.option
svocabs, tvocabs = option["vocabulary"]
unk_sym = option["unk"]
eos_sym = option["eos"]
svocab, isvocab = svocabs
tvocab, itvocab = tvocabs
reader = textreader([src.name, tgt.name])
stream = textiterator(reader, [args.batch, args.batch])
scores_i = []
for x, y in stream:
xdata, xmask = convert_data(x, svocab, unk_sym, eos_sym)
ydata, ymask = convert_data(y, tvocab, unk_sym, eos_sym)
_scores = model.get_snt_cost(xdata, xmask, ydata, ymask)
if args.normalize:
_scores = _scores / numpy.sum(ymask, 0)
scores_i.append(_scores)
scores.append(numpy.concatenate(scores_i))
if args.output_n_best:
writer = open(args.output_n_best, 'w')
else:
writer = sys.stdout
write = writer.write
for i, line in enumerate(nbest_lines):
score_str = ' '.join(map(str, [s[i] for s in scores]))
write('{0} {1}\n'.format(line.strip(), score_str))
if args.output_n_best:
writer.close()
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):
option, params = load_model(args.model)
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)
print_option(option)
# load references
if option["references"]:
references = load_references(option["references"])
else:
references = None
# input corpus
batch = option["batch"]
sortk = option["sort"] or 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 session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
if args.gpuid >= 0:
config.gpu_options.visible_device_list = "%d" % args.gpuid
with tf.Session(config=config):
# set seed
np.random.seed(option["seed"])
tf.set_random_seed(option["seed"])
# create model
initializer = tf.random_uniform_initializer(-0.08, 0.08)
model = rnnsearch(initializer=initializer, **option)
print "parameters:", count_parameters(tf.trainable_variables())
variables = None
if restore:
matched, not_matched = match_variables(tf.trainable_variables(),
init_params)
if args.finetune:
variables = not_matched
# create optimizer
constraint = ["norm", option["norm"]]
optim = optimizer(model, algorithm=option["optimizer"], norm=True,
constraint=constraint, variables=variables)
tf.global_variables_initializer().run()
if not init:
set_variables(tf.trainable_variables(), params)
if restore:
restore_variables(matched, not_matched)
# 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"]
for i in range(epoch, maxepoch):
for data in stream:
xdata, xlen = convert_data(data[0], svocab, unk_sym, eos_sym)
ydata, ylen = convert_data(data[1], tvocab, unk_sym, eos_sym)
t1 = time.time()
cost, norm = optim.optimize(xdata, xlen, ydata, ylen)
optim.update(alpha=alpha)
t2 = time.time()
count += 1
cost = cost * len(ylen) / sum(ylen)
totcost += cost / math.log(2)
print i + 1, count, cost, norm, t2 - t1
# save model
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:
batch = len(data[0])
ind = np.random.randint(0, batch)
sdata = data[0][ind]
tdata = data[1][ind]
xdata = xdata[:, ind : ind + 1]
xlen = xlen[ind : ind + 1]
hls = beamsearch(model, xdata, xlen, **search_opt)
best, score = hls[0]
print sdata
print tdata
print "search score:", score
print "translation:", " ".join(best[:-1])
print "--------------------------------------------------"
if option["vfreq"] 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 / option["count"][0]
print "--------------------------------------------------"
# early stopping
if i >= 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()
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()
