def eval(name):
xp = cuda.cupy
cuda.get_device(0).use()
vocab = Vocabulary.load_from_file("model.vocab")
m = LetterClassifyer(3000, 200, 1000, 2)
chainer.serializers.load_hdf5("model.hdf5", m)
m.to_gpu()
x_batch = [letter_list_text(name)]
output = forward(x_batch, None, m, False, vocab, xp)
return np.argmax(output.data)
def eval(args):
if args.use_gpu:
xp = cuda.cupy
cuda.get_device(0).use()
else:
xp = np
vocab = Vocabulary.load_from_file("%s.vocab" % args.model)
m = LetterClassifyer(args.vocab, args.embed, args.hidden, args.classes)
chainer.serializers.load_hdf5("%s.hdf5" % args.model, m)
if args.use_gpu:
m.to_gpu()
x_batch = [letter_list_text(args.file)]
output = forward(x_batch, None, m, False, vocab, xp)
print(output.data)
print("hyp: %d" % np.argmax(output.data)) # label
def train(args):
if args.use_gpu:
xp = cuda.cupy
cuda.get_device(0).use()
else:
xp = np
vocab = Vocabulary(args.file)
m = LetterClassifyer(args.vocab, args.embed, args.hidden)
m.zerograds()
if args.use_gpu:
m.to_gpu()
time_t = 10
for e in range(args.epoch):
opt = chainer.optimizers.Adam(alpha=0.001)
opt.setup(m)
opt.add_hook(chainer.optimizer.GradientClipping(5.0))
print("epoch: %d" % e)
i =0
total_acc = 0
e_acc = 0.0
for x_batch, y in letter_list(args.file):
x_batch = [x_batch]
output, acc, loss = forward(x_batch, y, m, True, vocab, xp)
total_acc += acc
e_acc += acc
if i % time_t == 0:
if i != 0:
total_acc /= time_t
print("time: %d, accuracy %f loss %f" % (i, total_acc.data, loss.data))
total_acc = 0
# print("".join(x_batch[0]))
# print(",".join([str(vocab.stoi(x)) for x in x_batch[0]]))
loss.backward()
if args.unchain:
loss.unchain_backward()
opt.update()
i += 1
sys.stdout.flush()
chainer.serializers.save_hdf5(args.model + ".hdf5", m)
vocab.save_vocab(args.model + ".vocab")
e_acc /= i
print("total acc: %f" % e_acc.data)
def __eval(args):
xp = np
vocab = Vocabulary.load_from_file("%s.vocab" % args.model)
m = LetterClassifyer(args.vocab, args.embed, args.hidden, args.classes)
chainer.serializers.load_hdf5("%s.hdf5" % args.model, m)
x_batch = [letter_list_text(args.file)]
output = forward(x_batch, None, m, False, vocab, xp)
print(args)
print(output.data)
print("hyp: %d" % np.argmax(output.data)) # output.dataの最大値を返す
return {"result": np.argmax(output.data), "data": output.data}
def evaluate(args):
xp = np
vocab = Vocabulary.load_from_file("%s.vocab" % args['model'])
m = LetterClassifyer(args['vocab'], args['embed'], args['hidden'],
args['classes'])
chainer.serializers.load_hdf5("%s.hdf5" % args['model'], m)
x_batch = [letter_list_text(args['file'])]
print('m: ', m)
print('x_batch: ', x_batch)
output = forward(x_batch, None, m, False, vocab, xp)
print(args)
print('output: ', output)
print(output.data)
print("hyp: %d" % np.argmax(output.data)) # output.dataの最大値を返す
return {
"result": int(np.argmax(output.data)),
"data": output.data.tolist()
}
def train(args):
if args.use_gpu:
xp = cuda.cupy
cuda.get_device(0).use()
else:
xp = np
vocab = Vocabulary(args.file)
m = LetterClassifyer(args.vocab, args.embed, args.hidden, args.classes)
m.cleargrads()
if args.use_gpu:
m.to_gpu()
time_t = 10
for e in range(args.epoch):
opt = chainer.optimizers.Adam(alpha=0.001)
opt.setup(m)
opt.add_hook(chainer.optimizer.GradientClipping(5.0))
print("epoch: %d" % e)
i = 0
total_acc = 0
e_acc = 0.0
for x_batch, y in letter_list(args.file):
x_batch = [x_batch]
output, acc, loss = forward(x_batch, y, m, True, vocab, xp)
total_acc += acc
e_acc += acc
if i % time_t == 0:
if i != 0:
total_acc /= time_t
print("time: %d, accuracy %f loss %f" %
(i, total_acc.data, loss.data))
total_acc = 0
# print("".join(x_batch[0]))
# print(",".join([str(vocab.stoi(x)) for x in x_batch[0]]))
loss.backward()
if args.unchain:
loss.unchain_backward()
opt.update()
i += 1
sys.stdout.flush()
chainer.serializers.save_hdf5(args.model + ".hdf5", m)
vocab.save_vocab(args.model + ".vocab")
e_acc /= i
print("total acc: %f" % e_acc.data)