def onDecodingFinish(data, output):
if op == "gen":
for src_id, (inp, out) in sorted(output.items(), key=lambda x:x[0]):
print(TRG.str_rpr(out))
elif op == "cppl":
UF.trace("Corpus PPL:", PPL(output))
print(PPL(output))
def train(self, x_data, y_data, learn=True, *args, **kwargs):
accum_loss, output = self._train(x_data, y_data, not learn, *args, **kwargs)
if learn:
if not math.isnan(float(accum_loss.data)):
self._model.zerograds()
accum_loss.backward()
self._opt.update()
else:
UF.trace("Warning: LOSS is nan, ignoring!")
return accum_loss.data, output
def main():
args = parse_args()
init_program_state(args)
# Variable
batch_size = args.batch
epoch_total = args.epoch
# data
UF.trace("Loading corpus + dictionary")
train, label, X, Y = load_data()
# Setup model
UF.trace("Setting up classifier")
model = L.Classifier(load_model(args, X, Y))
if not args.use_cpu: model = model.to_gpu()
opt = optimizers.SGD()
opt.setup(model)
# Begin training
UF.trace("Begin training MLP")
for ep in range(epoch_total):
UF.trace("Epoch %d" % (ep+1))
accum_loss = 0
for i in range(0, len(train), batch_size):
x_data = Variable(train[i:i+batch_size])
y_data = Variable(label[i:i+batch_size])
model.zerograds()
loss = model(x_data, y_data)
accum_loss += loss
loss.backward()
opt.update()
print("Loss:", loss.data, file=sys.stderr)
print("Accuracy:", model.accuracy.data, file=sys.stderr)
# # Classifying training data
# for i in range(0, len(train), batch_size):
# x_data = Variable(train[i:i+batch_size])
# y_data = model.predictor(x_data)
# UF.print_classification(y_data.data, Y)
UF.trace("Saving model....")
with ModelFile(open(args.model_out, "w")) as model_out:
model.predictor.save(model_out)
def main():
args = parse_args()
init_program_state(args)
# Variable
batch_size = args.batch
# Setup model
UF.trace("Setting up classifier")
model = load_model(args)
if not args.use_cpu: model = model.to_gpu()
# data
UF.trace("Loading corpus + dictionary")
test = load_data(sys.stdin, model._feat, model._input)
for i in range(0, len(test), batch_size):
x_data = Variable(test[i:i+batch_size])
y_data = model(x_data)
UF.print_classification(y_data.data, model._trg)
def onDecodingStart():
if op == "gen":
UF.trace("Sentence generation started.")
elif op == "cppl":
UF.trace("Corpus PPL calculation started.")
elif op == "sppl":
UF.trace("Sentence PPL calculation started.")
sys.exit(1)
# Loading data
x_all, y_all = load_data(sys.argv[1])
x_train, x_test = np.split(x_all, [60000])
y_train, y_test = np.split(y_all, [60000])
train_size = len(x_train)
test_size = len(x_test)
# Init model
model = L.Classifier(MLP())
optimizer = optimizers.SGD()
optimizer.setup(model)
# Training begins here
UF.trace("Begin training")
for epoch in range(EPOCH):
UF.trace("Epoch %d" % (epoch+1))
indexes = np.random.permutation(train_size)
for i in range(0, train_size, BATCH):
x = Variable(x_train[indexes[i:i+BATCH]])
t = Variable(y_train[indexes[i:i+BATCH]])
model.zerograds()
loss = model(x,t)
loss.backward()
optimizer.update()
# Testing begins here
UF.trace("Begin Testing")
sum_loss, sum_accuracy = 0, 0
for i in range(0, test_size, BATCH):
parser.add_argument("--operation", choices=["sppl", "cppl", "gen"], help="sppl: Sentence-wise ppl\ncppl: Corpus-wise ppl\ngen: Read input, start generating random words.", default="sppl")
parser.add_argument("--use_cpu", action="store_true")
parser.add_argument("--gpu", type=int, default=-1, help="Which GPU to use (Negative for cpu).")
parser.add_argument("--verbose", action="store_true")
parser.add_argument("--gen_limit", type=positive, default=50)
parser.add_argument("--eos_disc", type=float, default=0.0, help="Give fraction positive discount to output longer sentence.")
args = parser.parse_args()
op = args.operation
if op == "sppl" and args.batch != 1:
raise ValueError("Currently sentence based perplexity not supports multi batching.")
if args.use_cpu:
args.gpu = -1
# Loading model
UF.trace("Setting up classifier")
model = LanguageModel(args, use_gpu=args.gpu, collect_output=True)
VOC, _ = model.get_vocabularies()
decoding_options = {"gen_limit": args.gen_limit, "eos_disc": args.eos_disc}
# Testing callbacks
def PPL(loss):
try:
return math.exp(loss.data)
except:
return math.exp(loss)
def onDecodingStart():
if op == "gen":
UF.trace("Sentence generation started.")
elif op == "cppl":
def onDecodingStart():
UF.trace("Decoding started.")
