As redes de codificadores automáticos parecem ser muito mais complicadas do que as redes MLP classificadoras normais. Depois de várias tentativas usando Lasagne, tudo o que recebo na saída reconstruída é algo que se assemelha, na melhor das hipóteses, a uma média borrada de todas as imagens do banco de dados MNIST, sem distinção do que realmente é o dígito de entrada.
A estrutura de redes que escolhi são as seguintes camadas em cascata:
- camada de entrada (28x28)
- Camada convolucional 2D, tamanho do filtro 7x7
- Camada de pool máximo, tamanho 3x3, passada 2x2
- Camada achatada densa (totalmente conectada), 10 unidades (este é o gargalo)
- Camada densa (totalmente conectada), 121 unidades
- Remodelando a camada para 11x11
- Camada convolucional 2D, tamanho do filtro 3x3
- Fator 2 da camada de upscaling 2D
- Camada convolucional 2D, tamanho do filtro 3x3
- Fator 2 da camada de upscaling 2D
- Camada convolucional 2D, tamanho do filtro 5x5
- Pool máximo de recursos (de 31x28x28 a 28x28)
Todas as camadas convolucionais 2D têm os vieses desatados, ativações sigmóides e 31 filtros.
Todas as camadas totalmente conectadas possuem ativações sigmóides.
A função de perda usada é um erro ao quadrado , a função de atualização é adagrad
. O comprimento do pedaço para o aprendizado é de 100 amostras, multiplicadas por 1000 épocas.
A seguir, é apresentada uma ilustração do problema: na linha superior estão algumas amostras definidas como entradas da rede, a linha inferior é a reconstrução:
Apenas para completar, a seguir está o código que usei:
import theano.tensor as T
import theano
import sys
sys.path.insert(0,'./Lasagne') # local checkout of Lasagne
import lasagne
from theano import pp
from theano import function
import gzip
import numpy as np
from sklearn.preprocessing import OneHotEncoder
import matplotlib.pyplot as plt
def load_mnist():
def load_mnist_images(filename):
with gzip.open(filename, 'rb') as f:
data = np.frombuffer(f.read(), np.uint8, offset=16)
# The inputs are vectors now, we reshape them to monochrome 2D images,
# following the shape convention: (examples, channels, rows, columns)
data = data.reshape(-1, 1, 28, 28)
# The inputs come as bytes, we convert them to float32 in range [0,1].
# (Actually to range [0, 255/256], for compatibility to the version
# provided at http://deeplearning.net/data/mnist/mnist.pkl.gz.)
return data / np.float32(256)
def load_mnist_labels(filename):
# Read the labels in Yann LeCun's binary format.
with gzip.open(filename, 'rb') as f:
data = np.frombuffer(f.read(), np.uint8, offset=8)
# The labels are vectors of integers now, that's exactly what we want.
return data
X_train = load_mnist_images('train-images-idx3-ubyte.gz')
y_train = load_mnist_labels('train-labels-idx1-ubyte.gz')
X_test = load_mnist_images('t10k-images-idx3-ubyte.gz')
y_test = load_mnist_labels('t10k-labels-idx1-ubyte.gz')
return X_train, y_train, X_test, y_test
def plot_filters(conv_layer):
W = conv_layer.get_params()[0]
W_fn = theano.function([],W)
params = W_fn()
ks = np.squeeze(params)
kstack = np.vstack(ks)
plt.imshow(kstack,interpolation='none')
plt.show()
def main():
#theano.config.exception_verbosity="high"
#theano.config.optimizer='None'
X_train, y_train, X_test, y_test = load_mnist()
ohe = OneHotEncoder()
y_train = ohe.fit_transform(np.expand_dims(y_train,1)).toarray()
chunk_len = 100
visamount = 10
num_epochs = 1000
num_filters=31
dropout_p=.0
print "X_train.shape",X_train.shape,"y_train.shape",y_train.shape
input_var = T.tensor4('X')
output_var = T.tensor4('X')
conv_nonlinearity = lasagne.nonlinearities.sigmoid
net = lasagne.layers.InputLayer((chunk_len,1,28,28), input_var)
conv1 = net = lasagne.layers.Conv2DLayer(net,num_filters,(7,7),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(2,2))
net = lasagne.layers.DropoutLayer(net,p=dropout_p)
#conv2_layer = lasagne.layers.Conv2DLayer(dropout_layer,num_filters,(3,3),nonlinearity=conv_nonlinearity)
#pool2_layer = lasagne.layers.MaxPool2DLayer(conv2_layer,(3,3),stride=(2,2))
net = lasagne.layers.DenseLayer(net,10,nonlinearity=lasagne.nonlinearities.sigmoid)
#augment_layer1 = lasagne.layers.DenseLayer(reduction_layer,33,nonlinearity=lasagne.nonlinearities.sigmoid)
net = lasagne.layers.DenseLayer(net,121,nonlinearity=lasagne.nonlinearities.sigmoid)
net = lasagne.layers.ReshapeLayer(net,(chunk_len,1,11,11))
net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.Upscale2DLayer(net,2)
net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
#pool_after0 = lasagne.layers.MaxPool2DLayer(conv_after1,(3,3),stride=(2,2))
net = lasagne.layers.Upscale2DLayer(net,2)
net = lasagne.layers.DropoutLayer(net,p=dropout_p)
#conv_after2 = lasagne.layers.Conv2DLayer(upscale_layer1,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
#pool_after1 = lasagne.layers.MaxPool2DLayer(conv_after2,(3,3),stride=(1,1))
#upscale_layer2 = lasagne.layers.Upscale2DLayer(pool_after1,4)
net = lasagne.layers.Conv2DLayer(net,num_filters,(5,5),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.FeaturePoolLayer(net,num_filters,pool_function=theano.tensor.max)
print "output_shape:",lasagne.layers.get_output_shape(net)
params = lasagne.layers.get_all_params(net, trainable=True)
prediction = lasagne.layers.get_output(net)
loss = lasagne.objectives.squared_error(prediction, output_var)
#loss = lasagne.objectives.binary_crossentropy(prediction, output_var)
aggregated_loss = lasagne.objectives.aggregate(loss)
updates = lasagne.updates.adagrad(aggregated_loss,params)
train_fn = theano.function([input_var, output_var], loss, updates=updates)
test_prediction = lasagne.layers.get_output(net, deterministic=True)
predict_fn = theano.function([input_var], test_prediction)
print "starting training..."
for epoch in range(num_epochs):
selected = list(set(np.random.random_integers(0,59999,chunk_len*4)))[:chunk_len]
X_train_sub = X_train[selected,:]
_loss = train_fn(X_train_sub, X_train_sub)
print("Epoch %d: Loss %g" % (epoch + 1, np.sum(_loss) / len(X_train)))
"""
chunk = X_train[0:chunk_len,:,:,:]
result = predict_fn(chunk)
vis1 = np.hstack([chunk[j,0,:,:] for j in range(visamount)])
vis2 = np.hstack([result[j,0,:,:] for j in range(visamount)])
plt.imshow(np.vstack([vis1,vis2]))
plt.show()
"""
print "done."
chunk = X_train[0:chunk_len,:,:,:]
result = predict_fn(chunk)
print "chunk.shape",chunk.shape
print "result.shape",result.shape
plot_filters(conv1)
for i in range(chunk_len/visamount):
vis1 = np.hstack([chunk[i*visamount+j,0,:,:] for j in range(visamount)])
vis2 = np.hstack([result[i*visamount+j,0,:,:] for j in range(visamount)])
plt.imshow(np.vstack([vis1,vis2]))
plt.show()
import ipdb; ipdb.set_trace()
if __name__ == "__main__":
main()
Alguma idéia de como melhorar essa rede para obter um autoencoder razoavelmente funcional?
Problema resolvido!
Com uma implementação bastante diferente, usando um retificador com vazamento em vez de uma função sigmóide nas camadas convolucionais, apenas 2 (!!) nós na camada de gargalo e convoluções com kernels 1x1 no final.
Aqui está o resultado de alguma reconstrução:
Código:
import theano.tensor as T
import theano
import sys
sys.path.insert(0,'./Lasagne') # local checkout of Lasagne
import lasagne
from theano import pp
from theano import function
import theano.tensor.nnet
import gzip
import numpy as np
from sklearn.preprocessing import OneHotEncoder
import matplotlib.pyplot as plt
def load_mnist():
def load_mnist_images(filename):
with gzip.open(filename, 'rb') as f:
data = np.frombuffer(f.read(), np.uint8, offset=16)
# The inputs are vectors now, we reshape them to monochrome 2D images,
# following the shape convention: (examples, channels, rows, columns)
data = data.reshape(-1, 1, 28, 28)
# The inputs come as bytes, we convert them to float32 in range [0,1].
# (Actually to range [0, 255/256], for compatibility to the version
# provided at http://deeplearning.net/data/mnist/mnist.pkl.gz.)
return data / np.float32(256)
def load_mnist_labels(filename):
# Read the labels in Yann LeCun's binary format.
with gzip.open(filename, 'rb') as f:
data = np.frombuffer(f.read(), np.uint8, offset=8)
# The labels are vectors of integers now, that's exactly what we want.
return data
X_train = load_mnist_images('train-images-idx3-ubyte.gz')
y_train = load_mnist_labels('train-labels-idx1-ubyte.gz')
X_test = load_mnist_images('t10k-images-idx3-ubyte.gz')
y_test = load_mnist_labels('t10k-labels-idx1-ubyte.gz')
return X_train, y_train, X_test, y_test
def main():
X_train, y_train, X_test, y_test = load_mnist()
ohe = OneHotEncoder()
y_train = ohe.fit_transform(np.expand_dims(y_train,1)).toarray()
chunk_len = 100
num_epochs = 10000
num_filters=7
input_var = T.tensor4('X')
output_var = T.tensor4('X')
#conv_nonlinearity = lasagne.nonlinearities.sigmoid
#conv_nonlinearity = lasagne.nonlinearities.rectify
conv_nonlinearity = lasagne.nonlinearities.LeakyRectify(.1)
softplus = theano.tensor.nnet.softplus
#conv_nonlinearity = theano.tensor.nnet.softplus
net = lasagne.layers.InputLayer((chunk_len,1,28,28), input_var)
conv1 = net = lasagne.layers.Conv2DLayer(net,num_filters,(7,7),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(2,2))
net = lasagne.layers.DenseLayer(net,2,nonlinearity=lasagne.nonlinearities.sigmoid)
net = lasagne.layers.DenseLayer(net,49,nonlinearity=lasagne.nonlinearities.sigmoid)
net = lasagne.layers.ReshapeLayer(net,(chunk_len,1,7,7))
net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(1,1))
net = lasagne.layers.Upscale2DLayer(net,4)
net = lasagne.layers.Conv2DLayer(net,num_filters,(3,3),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.MaxPool2DLayer(net,(3,3),stride=(1,1))
net = lasagne.layers.Upscale2DLayer(net,4)
net = lasagne.layers.Conv2DLayer(net,num_filters,(5,5),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.Conv2DLayer(net,num_filters,(1,1),nonlinearity=conv_nonlinearity,untie_biases=True)
net = lasagne.layers.FeaturePoolLayer(net,num_filters,pool_function=theano.tensor.max)
net = lasagne.layers.Conv2DLayer(net,1,(1,1),nonlinearity=conv_nonlinearity,untie_biases=True)
print "output shape:",net.output_shape
params = lasagne.layers.get_all_params(net, trainable=True)
prediction = lasagne.layers.get_output(net)
loss = lasagne.objectives.squared_error(prediction, output_var)
#loss = lasagne.objectives.binary_hinge_loss(prediction, output_var)
aggregated_loss = lasagne.objectives.aggregate(loss)
#updates = lasagne.updates.adagrad(aggregated_loss,params)
updates = lasagne.updates.nesterov_momentum(aggregated_loss,params,0.5)#.005
train_fn = theano.function([input_var, output_var], loss, updates=updates)
test_prediction = lasagne.layers.get_output(net, deterministic=True)
predict_fn = theano.function([input_var], test_prediction)
print "starting training..."
for epoch in range(num_epochs):
selected = list(set(np.random.random_integers(0,59999,chunk_len*4)))[:chunk_len]
X_train_sub = X_train[selected,:]
_loss = train_fn(X_train_sub, X_train_sub)
print("Epoch %d: Loss %g" % (epoch + 1, np.sum(_loss) / len(X_train)))
print "done."
chunk = X_train[0:chunk_len,:,:,:]
result = predict_fn(chunk)
print "chunk.shape",chunk.shape
print "result.shape",result.shape
visamount = 10
for i in range(10):
vis1 = np.hstack([chunk[i*visamount+j,0,:,:] for j in range(visamount)])
vis2 = np.hstack([result[i*visamount+j,0,:,:] for j in range(visamount)])
plt.imshow(np.vstack([vis1,vis2]))
plt.show()
import ipdb; ipdb.set_trace()
if __name__ == "__main__":
main()