def main():
args = parse_args()
beam_width = args.beam_width
d = builder.initiate(args.model_dir)
m = d['model']
src_vcb, trg_vcb = d['vocabularies']
_, valid_data = d['datasets']
source, source_variables = get_agiga_example(src_vcb, valid_data,
args.index)
generator = BeamSearchGenerator(m, src_vcb, trg_vcb, beam_width=beam_width)
length = np.array([args.length], dtype=np.int32)
ret = generator.generate(source_variables, length, byte=True)
output, score, state, beams = ret
out = trg_vcb.revert(output, with_tags=False)
print('{} & {} & {} \\\\'.format(m.name, args.length, ' '.join(out)))
print(' ', len(out), len(' '.join(out)))
print('- others')
for b in beams:
sent = ' '.join(trg_vcb.revert(b[1], with_tags=False))
print(' %0.2f' % b[0], len(sent), sent, sep=' & ')
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('model_dir')
parser.add_argument("--gpu", type=int, default=-1)
parser.add_argument("--reset", action='store_true')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
if args.reset:
builder.reset(args.model_dir)
d = builder.initiate(args.model_dir)
model = d['model']
if args.gpu >= 0:
chainer.cuda.get_device(args.gpu).use()
model.to_gpu()
trainer = Trainer(model, d['optimizer'], d['vocabularies'], args.model_dir)
train_data, valid_data = d['datasets']
b = d['builder']
batch_size = int(b.config["Training"]["batch_size"])
batches_for_sort = int(b.config["Training"]["batches_for_sort"])
epochs = int(b.config["Training"]["epochs"])
trainer.train(epochs, train_data, valid_data, batch_size, batches_for_sort)