x_train, x_test = np.split(x_all, [N])
N_test = x_test.shape[0]
# モデルの作成
da = DenoisingAutoencoder(784, 100, noise=SaltAndPepperNoise())
# 学習
all_loss = []
for epoch in xrange(n_epoch):
print 'epoch', epoch
indexes = np.random.permutation(N)
losses = []
sum_loss = 0
for i in xrange(0, N, batchsize):
x_batch = x_train[indexes[i:i+batchsize]]
loss = da.train(x_batch)
sum_loss += float(loss.data) * batchsize
if epoch == 0 and i == 0:
with open('../output/da/graph.dot', 'w') as o:
o.write(computational_graph.build_computational_graph((loss, )).dump())
with open('../output/da/graph.wo_split.dot', 'w') as o:
g = computational_graph.build_computational_graph((loss, ), remove_split=True)
o.write(g.dump())
print 'train mean loss={}'.format(sum_loss / N)
losses += [sum_loss / N]
# 評価
sum_loss = 0
for i in xrange(0, N_test, batchsize):
x_batch = x_test[i:i+batchsize]
loss = da.test(x_batch)
sum_loss += float(loss.data) * batchsize
# 1層目
da1_filename = '../output/sda/model_da1.pkl'
try:
da1 = pickle.load(open(da1_filename))
except IOError:
da1 = DenoisingAutoencoder(784, 100, noise=SaltAndPepperNoise())
n_epoch = 30
all_loss = []
for epoch in xrange(n_epoch):
print 'epoch', epoch
indexes = np.random.permutation(N)
losses = []
sum_loss = 0
for i in xrange(0, N, batchsize):
x_batch = x_train[indexes[i:i+batchsize]]
loss = da1.train(x_batch)
sum_loss += float(loss.data) * batchsize
print 'train mean loss={}'.format(sum_loss / N)
losses += [sum_loss / N]
# 評価
sum_loss = 0
for i in xrange(0, N_test, batchsize):
x_batch = x_test[i:i+batchsize]
loss = da1.test(x_batch)
sum_loss += float(loss.data) * batchsize
print 'test mean loss={}'.format(sum_loss / N_test)
losses += [sum_loss / N_test]
all_loss.append(losses)
# 可視化
h1 = MinMaxScaler().fit_transform(da1.model.encoder.W)
visualize(h1, '../img/sda/da1/sda_da1_{0:04d}.jpg'.format(epoch + 1), (8, 8), (10, 10))
