How to create multilayer autoencoder?
dr.rom...@gmail.com
I would like to create multi-layer autoencoder for vector analysis. I'm using tutorial https://blog.keras.io/building-autoencoders-in-keras.html I have a long binary vector (contains 0 or 1 for each dimension) representing an object with length e.g. 200. Main idea is to train NN and to reduce a vector size to 10. Then I would like to compare these reduced vectors. An autoencoder with a single fully-connected neural layer from this tutorial compiled and seems to work. But when I try to use multiple hidden layers like in tutorial, I can't even compile source code. My code is below. As you can see I tried all possible combinations but can't define decoder. I get error:
--------------------------------------------------------------------------- AssertionError Traceback (most recent call last) <ipython-input-24-10401e276f94> in <module>() 63 #decoder_3 = Model(input = encoded_input_3, output = decoder_layer_3(encoded_input_3)) 64 ---> 65 decoder_1 = Model(input=encoded_input_1, output=decoder_layer_3(decoder_layer_2(decoder_layer_1(encoded_input_1)))) 66 67 ''' /usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in __call__(self, x, mask) 567 if inbound_layers: 568 # This will call layer.build() if necessary. --> 569 self.add_inbound_node(inbound_layers, node_indices, tensor_indices) 570 # Outputs were already computed when calling self.add_inbound_node. 571 outputs = self.inbound_nodes[-1].output_tensors /usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in add_inbound_node(self, inbound_layers, node_indices, tensor_indices) 630 # creating the node automatically updates self.inbound_nodes 631 # as well as outbound_nodes on inbound layers. --> 632 Node.create_node(self, inbound_layers, node_indices, tensor_indices) 633 634 def get_output_shape_for(self, input_shape): /usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in create_node(cls, outbound_layer, inbound_layers, node_indices, tensor_indices) 166 # TODO: try to auto-infer shape 167 # if exception is raised by get_output_shape_for. --> 168 output_shapes = to_list(outbound_layer.get_output_shape_for(input_shapes[0])) 169 else: 170 output_tensors = to_list(outbound_layer.call(input_tensors, mask=input_masks)) /usr/local/lib/python2.7/dist-packages/keras/layers/core.pyc in get_output_shape_for(self, input_shape) 773 def get_output_shape_for(self, input_shape): 774 assert input_shape and len(input_shape) >= 2 --> 775 assert input_shape[-1] and input_shape[-1] == self.input_dim 776 output_shape = list(input_shape) 777 output_shape[-1] = self.output_dim AssertionError:
How should I correctly define decoder? And how can I access trained 10 number vectors? And map it to original objects?
I found some similar codings http://stackoverflow.com/questions/37758496/python-keras-theano-wrong-dimensions-for-deep-autoencoder or http://stackoverflow.com/questions/40118965/keras-deep-autoencoder-prediction-is-inaccurate. But it doesn't help.
Thank you in advance, I would appreciate any help.
Regards,
Roman
##############
from keras.layers import Input, Dense
from keras.models import Model
# this is the size of our encoded representations
VECTOR_SIZE = 200
input_doc = Input(shape=(VECTOR_SIZE,))
encoded = Dense(VECTOR_SIZE, activation='relu')(input_doc)
#'''
#encoded = Dense(200, activation='relu')(input_doc)
encoded = Dense(50, activation='relu')(encoded)
encoded = Dense(10, activation='relu')(encoded)
###decoded = Dense(10, activation='relu')(encoded)
decoded = Dense(50, activation='relu')(encoded)
#decoded = Dense(200, activation='relu')(decoded)
decoded = Dense(VECTOR_SIZE, activation='sigmoid')(decoded)
#'''
#decoded = Dense(VECTOR_SIZE, activation='sigmoid')(encoded)
# this model maps an input to its reconstruction
autoencoder = Model(input=input_doc, output=decoded)
# this model maps an input to its encoded representation
encoder = Model(input=input_doc, output=encoded)
#encoder_layer = autoencoder.layers[0]
#encoder_layer2 = autoencoder.layers[1]
#encoder_layer3 = autoencoder.layers[2]
#encoder10 = Model(input=input_doc, output=encoder_layer3(encoder_layer2(encoder_layer(input_doc))))
# create a placeholder for an encoded (32-dimensional) input
encoded_input_1 = Input(shape=(10,))
encoded_input_2 = Input(shape=(50,))
encoded_input_3 = Input(shape=(VECTOR_SIZE,))
# retrieve the last layer of the autoencoder model
decoder_layer_1 = autoencoder.layers[-3]
decoder_layer_2 = autoencoder.layers[-2]
decoder_layer_3 = autoencoder.layers[-1]
# create the decoder model
#decoder_1 = Model(input = encoded_input_1, output = decoder_layer_1(encoded_input_1))
#decoder_2 = Model(input = encoded_input_2, output = decoder_layer_2(encoded_input_2))
#decoder_3 = Model(input = encoded_input_3, output = decoder_layer_3(encoded_input_3))
decoder_1 = Model(input=encoded_input_1, output=decoder_layer_3(decoder_layer_2(decoder_layer_1(encoded_input_1))))
'''
# create a placeholder for an encoded (32-dimensional) input
encoded_input = Input(shape=(VECTOR_SIZE,))
# retrieve the last layer of the autoencoder model
decoder_layer = autoencoder.layers[-1]
decoder_layer3 = autoencoder.layers[-3]
decoder_layer2 = autoencoder.layers[-2]
#decoder_layer3 = autoencoder.layers[-1]
# create the decoder model
##decoder10 = Model(input=encoded_input, output=decoder_layer(decoder_layer2(decoder_layer3(encoded_input))))
#decoder10 = Model(input=encoded_input, output=decoder_layer3(decoder_layer2(decoder_layer(encoded_input))))
'''
# create the decoder model
#decoder = Model(input=encoded_input, output=decoder_layer(encoded_input))
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')
# train autoencoder
autoencoder.fit(train_vectors, train_vectors,
nb_epoch=50,
batch_size=256,
shuffle=True,
validation_data=(test_vectors, test_vectors))
# after encoder reaches stable train/test loss value
autoencoder.summary()
encoder.summary()
decoder.summary()
print 'decoders:'
#decoder10.summary()
decoder_1.summary()
decoder_2.summary()
decoder_3.summary()
Jim Sharpe
bale...@lbl.gov
#-------------------
def build_model():
''' based on https://blog.keras.io/building-autoencoders-in-keras.html
'''
# this is MNIST input placeholder
input_img = Input(shape=(784,),name='inp_img')
net = Dense(encoding_dim*4, activation='relu',name='img2zip')(input_img)
encoded = Dense(encoding_dim, activation='relu',name='zip_data')(net)
# "decoded" is the lossy reconstruction of the input
net = Dense(encoding_dim*4, activation='relu',name='zip2img')(encoded)
decoded = Dense(784, activation='sigmoid',name='out_img')(net)
#1) this model maps an input to its reconstruction
autoencoderM = Model(input_img, decoded)
#2) this model maps an input to its encoded representation
encoderM = Model(input_img, encoded)
# create a placeholder for an encoded (zip-dimensional) input
encoded_input = Input(shape=(encoding_dim,),name='zip_input')
# retrieve last 2 layers of the autoencoder model
net = autoencoderM.layers[-2](encoded_input)
net = autoencoderM.layers[-1](net)
#3) this model maps encoded representation to input image
decoderM = Model(encoded_input, net)
autoencoderM.compile(optimizer='adadelta', loss='binary_crossentropy', metrics=['accuracy'])
print('\nFull auto-encoder:',autoencoderM.count_params())
autoencoderM.summary() # will print
print('\nEncoder:',encoderM.count_params())
encoderM.summary() # will print
print('\nDecoder:',decoderM.count_params())
decoderM.summary() # will print
return autoencoderM,encoderM,decoderM
- - -
#use case:
autoM,encoM,decoM=build_model()
hir=autoM.fit(X,X,shuffle=True, verbose=args.verb,
validation_data=(X_test, X_test),
batch_size=args.batch_size, nb_epoch=args.epochs )
encoded_imgs = encoM.predict(X_test)
decoded_imgs = decoM.predict(encoded_imgs)