def main():
# Settings
device = int(sys.argv[1]) if len(sys.argv) > 1 else None
T = 5
batch_size = 128
inp_dim = 784
out_dim = n_cls = 10
n_l_train_data = 100
n_train_data = 60000
dims = [inp_dim, 250, 100, out_dim]
lambdas = to_device(np.array([1., 1.], np.float32), device)
learning_rate = 1. * 1e-3
n_epoch = 100
decay = 0.5
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = n_epoch * iter_epoch
l_type = "hard"
# Separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/mnist/train.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/mnist/l_train.npz")
u_train_path = os.path.join(home, "datasets/mnist/train.npz")
test_path = os.path.join(home, "datasets/mnist/test.npz")
# DataReader, Model, Optimizer, Losses
data_reader = MNISTDataReader(l_train_path,
u_train_path,
test_path,
batch_size=batch_size,
n_cls=n_cls)
model = ElmanOnestep(dims)
rnn = ElmanNet(model, T)
model.to_gpu(device) if device else None
#TODO: Add gradient clip as hook
optimizer = optimizers.Adam(learning_rate)
optimizer.setup(model)
#optimizer.add_hook(chainer.optimizer.GradientHardClipping(-3, 3))
rnn_labeled_losses = RNNLabeledLosses(T)
rnn_unlabeled_losses = RNNUnlabeledLosses(T, l_type)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
for i in range(n_iter):
# Get data
x_l, y_l = [
to_device(x, device) for x in data_reader.get_l_train_batch()
]
x_u, _ = [
to_device(x, device) for x in data_reader.get_u_train_batch()
]
# Forward/Backward
forward_backward_update_052(rnn, rnn_labeled_losses,
rnn_unlabeled_losses, optimizer, model,
x_l, y_l, x_u)
# Eval
if (i + 1) % iter_epoch == 0:
print("Evaluation at {}-th epoch".format(epoch))
# Get data
x_l, y_l = [
to_device(x, device) for x in data_reader.get_test_batch()
]
# Compute loss and accuracy
losses = evaluate_052(rnn, rnn_labeled_losses, model, x_l, y_l)
# Report
print("Loss:{},Accuracy:{},Time/epoch:{}[s]".format(
[float(to_device(loss.data, device)) for loss in losses],
[float(to_device(accrucray.data, device)) * 100 \
for accrucray in rnn_labeled_losses.accuracies],
time.time() - st))
epoch += 1
st = time.time()
def main():
# Settings
device = int(sys.argv[1]) if len(sys.argv) > 1 else None
T = 5
batch_size = 128
inp_dim = 784
out_dim = n_cls = 10
n_l_train_data = 100
n_train_data = 60000
dims = [inp_dim, 250, 100, out_dim]
lambdas = to_device(np.array([1., 1.], np.float32), device)
learning_rate = 1. * 1e-3
n_epoch = 100
decay = 0.5
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = n_epoch * iter_epoch
l_type = "hard"
# Separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/mnist/train.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/mnist/l_train.npz")
u_train_path = os.path.join(home, "datasets/mnist/train.npz")
test_path = os.path.join(home, "datasets/mnist/test.npz")
# DataReader, Model, Optimizer, Losses
data_reader = MNISTDataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls)
model = LSTMOnestep(dims)
rnn = LSTMNet(model, T)
model.to_gpu(device) if device else None
#TODO: Add gradient clip as hook
optimizer = optimizers.Adam(learning_rate)
optimizer.setup(model)
#optimizer.add_hook(chainer.optimizer.GradientHardClipping(-3, 3))
rnn_labeled_losses = RNNLabeledLosses(T)
rnn_unlabeled_losses = RNNUnlabeledLosses(T, l_type)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
for i in range(n_iter):
# Get data
x_l, y_l = [to_device(x, device) for x in data_reader.get_l_train_batch()]
x_u, _ = [to_device(x, device) for x in data_reader.get_u_train_batch()]
# Forward/Backward
forward_backward_update_100(
rnn, rnn_labeled_losses, rnn_unlabeled_losses,
optimizer, model,
x_l, y_l, x_u)
# Eval
if (i+1) % iter_epoch == 0:
print("Evaluation at {}-th epoch".format(epoch))
# Get data
x_l, y_l = [to_device(x, device) for x in data_reader.get_test_batch()]
# Compute loss and accuracy
losses = evaluate_100(rnn, rnn_labeled_losses, model, x_l, y_l)
# Report
print("Loss:{},Accuracy:{},Time/epoch:{}[s]".format(
[float(to_device(loss.data, device)) for loss in losses],
[float(to_device(accrucray.data, device)) * 100 \
for accrucray in rnn_labeled_losses.accuracies],
time.time() - st))
epoch +=1
st = time.time()