def evaluation(model, fname, show_progress=False):
model.predictor.forget_history()
model.predictor.to_cpu()
devset = data_util.load_data(fname)
dev_data_resource = data_util.data_spliter(devset, batchsize=1, n_epoch=1)
dev_loss = 0
pred = []
target = []
for dev_idx, x_batch, y_batch, epoch, percentage, eos in dev_data_resource:
if not dev_idx:
break
if show_progress:
print_stats(percentage, 0, 0, dev_loss)
x = Variable(np.asarray(x_batch))
t = Variable(np.asarray(y_batch))
loss_i = model(x, t)
target.append(y_batch[0])
pred.append(model.y.data[0].argmax())
dev_loss = (dev_loss * (dev_idx - 1) + loss_i.data) / dev_idx
if eos and opts.forget_on_new_utt:
model.predictor.forget_history()
model.predictor.forget_history()
return dev_loss, pred, target
def train_mode():
nnet_model = None
if opts.nnet_struct:
nnet_model = NNET_Model.parse_structure(opts.nnet_struct)
elif opts.fname_in_model:
nnet_model = NNET_Model.load(opts.fname_in_model)
trainset = data_util.load_data(opts.fname_train)
eos_pad = misc.get_pading(opts.eos_pad)
train_data_resource = data_util.data_spliter(trainset, batchsize=opts.batchsize,
n_epoch=opts.n_epoch, EOS=eos_pad)
optimizer, model = setup_training(nnet_model, opts)
train_nnet(model, optimizer, train_data_resource, opts)
if opts.fname_test:
print('====================TESTING=========================')
test_loss, pred, target = evaluation(model, opts.fname_test, show_progress=True)
if 'cross_entropy' in opts.loss_function:
misc.f_measure(pred, target)
print(' test loss: %.3f' % test_loss)
if opts.fname_out_model:
nnet_model.save(opts.fname_out_model)