def build_CNN_input(vocab_size=VocabSize, usewords=USEWORDS,
skiptop=skipTop, devsplit=DEVSPLIT, verbose=True, **kwargs):
"""
:param vocab_size:
:param usewords:
:param skiptop:
:param devsplit:
:param verbose:
:param kwargs:
:return:
"""
if verbose:
print('Building CNN Inputs')
((X_train, y_train), (X_dev, y_dev), (X_val, y_val)) = build_design_matrix(vocab_size=vocab_size,
use_words=usewords,
skip_top=skiptop,
dev_split=devsplit,
**kwargs
)
if verbose:
print('X_train shape: {}'.format(X_train.shape))
print('X_dev shape: {}'.format(X_dev.shape))
print('X_test shape: {}'.format(X_val.shape))
print('y_train shape: {}'.format(y_train.shape))
print('y_dev shape: {}'.format(y_dev.shape))
print('y_test shape: {}'.format(y_val.shape))
return {'training': (X_train, y_train), "dev": (X_dev, y_dev), "val": (X_val, y_val)}
def build_CNN_input(usewords=USEWORDS, skiptop=skipTop, devsplit=DEVSPLIT, verbose=True):
print('Building input')
((X_train, y_train), (X_dev, y_dev), (X_test, y_test)) = build_design_matrix(VocabSize,
use_words=usewords,
skip_top=skiptop,
dev_split=devsplit)
if verbose:
print(len(X_train), 'train sequences')
print(len(X_dev), 'test sequences')
print('X_train shape:', X_train.shape)
print('X_dev shape:', X_dev.shape)
print('y_train shape:', y_train.shape)
print('y_dev shape:', y_dev.shape)
return X_train, y_train, X_dev, y_dev, X_test, y_test
def build_CNN_input(usewords=USEWORDS,
skiptop=skipTop,
devsplit=DEVSPLIT,
verbose=True):
print('Building input')
((X_train, y_train), (X_dev, y_dev),
(X_test, y_test)) = build_design_matrix(VocabSize,
use_words=usewords,
skip_top=skiptop,
dev_split=devsplit)
if verbose:
print(len(X_train), 'train sequences')
print(len(X_dev), 'test sequences')
print('X_train shape:', X_train.shape)
print('X_dev shape:', X_dev.shape)
print('y_train shape:', y_train.shape)
print('y_dev shape:', y_dev.shape)
return X_train, y_train, X_dev, y_dev, X_test, y_test
pool_len2 = 2
num_filters3 = 300
filter_length3 = 4
stride_len3 = 1
pool_len3 = 2
embedding_dims = 200
hidden_dims = 100
num_epochs = 5
print( 'Loading data...' )
((X_train, y_train), (X_test, y_test)) = build_design_matrix( VocabSize,
use_words = USEWORDS,
skip_top = skipTop,
dev_split = DEVSPLIT )
print( len( X_train ), 'train sequences' )
print( len( X_test ), 'test sequences' )
print( 'X_train shape:', X_train.shape )
print( 'X_test shape:', X_test.shape )
print( 'Build model...' )
model = Sequential( )
# we start off with an efficient embedding layer which maps
# our vocab indices into embedding_dims dimensions
model.add( Embedding( VocabSize, embedding_dims, input_length = maxReviewLen ) )
def train_siamese_model(): print( 'Loading data...' ) ((X_train, y_train), (X_test, y_test)) = build_design_matrix( VocabSize, use_words = USEWORDS, skip_top = skipTop, dev_split = DEVSPLIT ) print( len( X_train ), 'train sequences' ) print( len( X_test ), 'test sequences' ) print( 'X_train shape:', X_train.shape ) print( 'X_test shape:', X_test.shape ) print( 'Build model...' ) #LEFT and RIGHT branches of siamese network modelL = Sequential( ) modelR = Sequential( ) # we start off with an efficient embedding layer which maps # our vocab indices into embedding_dims dimensions modelL.add( Embedding( VocabSize, embedding_dims, input_length = maxReviewLen ) ) modelR.add( Embedding( VocabSize, embedding_dims, input_length = maxReviewLen ) ) modelL.add( Dropout( 0.10 ) ) modelR.add( Dropout( 0.10 ) ) ###init changed from uniform to glorot_norm modelL.add( Convolution1D( nb_filter = num_filters1, filter_length = filter_length1, border_mode = 'valid', activation = 'relu', subsample_length = stride_len1, init = 'uniform' ) ) modelR.add( Convolution1D( nb_filter = num_filters1, filter_length = filter_length1, border_mode = 'valid', activation = 'relu', subsample_length = stride_len1, init = 'uniform' ) ) # input_length=maxReviewLen, input_dim=VocabSize modelL.add( Dropout( 0.25 ) ) modelR.add( Dropout( 0.25 ) ) modelL.add( MaxPooling1D( pool_length = 2 ) ) modelR.add( MaxPooling1D( pool_length = 2 ) ) modelL.add( Convolution1D( nb_filter = num_filters2, filter_length = filter_length2, border_mode = 'valid', activation = 'relu', subsample_length = stride_len2, init = 'uniform' ) ) modelR.add( Convolution1D( nb_filter = num_filters2, filter_length = filter_length2, border_mode = 'valid', activation = 'relu', subsample_length = stride_len2, init = 'uniform' ) ) modelL.add( Dropout( 0.40 ) ) modelR.add( Dropout( 0.40 ) ) # we use standard max pooling (halving the output of the previous layer): modelL.add( MaxPooling1D( pool_length = 2 ) ) modelR.add( MaxPooling1D( pool_length = 2 ) ) modelL.add( Convolution1D( nb_filter = num_filters3, filter_length = filter_length3, border_mode = 'valid', activation = 'relu', subsample_length = stride_len3, init = 'uniform' ) ) modelR.add( Convolution1D( nb_filter = num_filters3, filter_length = filter_length3, border_mode = 'valid', activation = 'relu', subsample_length = stride_len3, init = 'uniform' ) ) modelL.add( Dropout( 0.30 ) ) modelR.add( Dropout( 0.30 ) ) modelL.add( MaxPooling1D( pool_length = 2 ) ) modelR.add( MaxPooling1D( pool_length = 2 ) ) modelL.add( Convolution1D( nb_filter = num_filters4, filter_length = filter_length4, border_mode = 'valid', activation = 'relu', subsample_length = stride_len4, init = 'uniform' ) ) modelR.add( Convolution1D( nb_filter = num_filters4, filter_length = filter_length4, border_mode = 'valid', activation = 'relu', subsample_length = stride_len4, init = 'uniform' ) ) modelL.add( Dropout( 0.25 ) ) modelR.add( Dropout( 0.25 ) ) modelL.add( MaxPooling1D( pool_length = pool_len4 ) ) modelR.add( MaxPooling1D( pool_length = pool_len4 ) ) # We flatten the output of the conv layer, # so that we can add a vanilla dense layer: modelL.add( Flatten( ) ) modelR.add( Flatten( ) ) # We add a vanilla hidden layer: modelL.add( Dense( hidden_dims ) ) modelL.add( Activation( 'relu' ) ) modelR.add( Dense( hidden_dims ) ) modelR.add( Activation( 'relu' ) ) merged_vector = merge([modelL, modelR], mode='concat', concat_axis=-1) Gw=Dense(1, activation=DistanceMetric)(merged_vector) model.compile( loss = contrastiveLoss, optimizer = 'rmsprop', )
def build_CNN_model():
print( 'Loading data...' )
((X_train, y_train), (X_test, y_test)) = build_design_matrix( VocabSize,
use_words = USEWORDS,
skip_top = skipTop,
dev_split = DEVSPLIT )
print( len( X_train ), 'train sequences' )
print( len( X_test ), 'test sequences' )
print( 'X_train shape:', X_train.shape )
print( 'X_test shape:', X_test.shape )
print( 'Build model...' )
model = Sequential( )
# we start off with an efficient embedding layer which maps
# our vocab indices into embedding_dims dimensions
model.add( Embedding( VocabSize, embedding_dims, input_length = maxReviewLen ) )
model.add( Dropout( 0.10 ) )
###init changed from uniform to glorot_norm
model.add( Convolution1D( nb_filter = num_filters1,
filter_length = filter_length1,
border_mode = 'valid',
activation = 'relu',
subsample_length = stride_len1,
init = 'uniform'
) )
# input_length=maxReviewLen, input_dim=VocabSize
model.add( Dropout( 0.25 ) )
model.add( MaxPooling1D( pool_length = 2 ) )
model.add( Convolution1D( nb_filter = num_filters2,
filter_length = filter_length2,
border_mode = 'valid',
activation = 'relu',
subsample_length = stride_len2,
init = 'uniform'
) )
model.add( Dropout( 0.40 ) )
# we use standard max pooling (halving the output of the previous layer):
model.add( MaxPooling1D( pool_length = 2 ) )
model.add( Convolution1D( nb_filter = num_filters3,
filter_length = filter_length3,
border_mode = 'valid',
activation = 'relu',
subsample_length = stride_len3,
init = 'uniform'
) )
model.add( Dropout( 0.30 ) )
model.add( MaxPooling1D( pool_length = 2 ) )
model.add( Convolution1D( nb_filter = num_filters4,
filter_length = filter_length4,
border_mode = 'valid',
activation = 'relu',
subsample_length = stride_len4,
init = 'uniform'
) )
model.add( Dropout( 0.25 ) )
model.add( MaxPooling1D( pool_length = pool_len4 ) )
# We flatten the output of the conv layer,
# so that we can add a vanilla dense layer:
model.add( Flatten( ) )
# We add a vanilla hidden layer:
model.add( Dense( hidden_dims ) )
model.add( Activation( 'relu' ) )
# We project onto a single unit output layer, and squash it with a sigmoid:
model.add( Dense( 1 ) )
model.add( Activation( 'sigmoid' ) )
model.compile( loss = 'binary_crossentropy',
optimizer = 'rmsprop',
)
weightPath = './model_data/saved_weights/'+filename
checkpoint = ModelCheckpoint( weightPath + '_W.{epoch:02d}-{val_loss:.2f}.hdf5',
verbose = 1, )
earlyStop = EarlyStopping(patience = 1,verbose = 1)
call_backs = [checkpoint,earlyStop]
return (model,X_train,y_train,X_test,y_test)
num_filters3 = 300
filter_length3 = 4
stride_len3 = 1
pool_len3 = 2
embedding_dims = 200
hidden_dims = 100
num_epochs = 5
print('Loading data...')
((X_train, y_train), (X_test,
y_test)) = build_design_matrix(VocabSize,
use_words=USEWORDS,
skip_top=skipTop,
dev_split=DEVSPLIT)
print(len(X_train), 'train sequences')
print(len(X_test), 'test sequences')
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
print('Build model...')
model = Sequential()
# we start off with an efficient embedding layer which maps
# our vocab indices into embedding_dims dimensions
model.add(Embedding(VocabSize, embedding_dims, input_length=maxReviewLen))
filter_length2 = 2
stride_len2 = 2
pool_len2 = 2
embedding_dims = 400
hidden_dims = 100
num_epochs = 3
DEVSPLIT = 14
USEWORDS = True
print('Loading data...')
((X_train, y_train),
(X_test, y_test)) = build_design_matrix(VocabSize,
use_words=USEWORDS,
skip_top=modelParameters.skip_top,
dev_split=DEVSPLIT)
print(len(X_train), 'train sequences')
print(len(X_test), 'test sequences')
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
print('Build model...')
model = Sequential()
# we start off with an efficient embedding layer which maps
# our vocab indices into embedding_dims dimensions
model.add(Embedding(VocabSize, embedding_dims, input_length=maxReviewLen))
model.add(Dropout(0.05))
def train_siamese_model():
print('Loading data...')
((X_train, y_train), (X_test,
y_test)) = build_design_matrix(VocabSize,
use_words=USEWORDS,
skip_top=skipTop,
dev_split=DEVSPLIT)
print(len(X_train), 'train sequences')
print(len(X_test), 'test sequences')
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
print('Build model...')
#LEFT and RIGHT branches of siamese network
modelL = Sequential()
modelR = Sequential()
# we start off with an efficient embedding layer which maps
# our vocab indices into embedding_dims dimensions
modelL.add(Embedding(VocabSize, embedding_dims, input_length=maxReviewLen))
modelR.add(Embedding(VocabSize, embedding_dims, input_length=maxReviewLen))
modelL.add(Dropout(0.10))
modelR.add(Dropout(0.10))
###init changed from uniform to glorot_norm
modelL.add(
Convolution1D(nb_filter=num_filters1,
filter_length=filter_length1,
border_mode='valid',
activation='relu',
subsample_length=stride_len1,
init='uniform'))
modelR.add(
Convolution1D(nb_filter=num_filters1,
filter_length=filter_length1,
border_mode='valid',
activation='relu',
subsample_length=stride_len1,
init='uniform'))
# input_length=maxReviewLen, input_dim=VocabSize
modelL.add(Dropout(0.25))
modelR.add(Dropout(0.25))
modelL.add(MaxPooling1D(pool_length=2))
modelR.add(MaxPooling1D(pool_length=2))
modelL.add(
Convolution1D(nb_filter=num_filters2,
filter_length=filter_length2,
border_mode='valid',
activation='relu',
subsample_length=stride_len2,
init='uniform'))
modelR.add(
Convolution1D(nb_filter=num_filters2,
filter_length=filter_length2,
border_mode='valid',
activation='relu',
subsample_length=stride_len2,
init='uniform'))
modelL.add(Dropout(0.40))
modelR.add(Dropout(0.40))
# we use standard max pooling (halving the output of the previous layer):
modelL.add(MaxPooling1D(pool_length=2))
modelR.add(MaxPooling1D(pool_length=2))
modelL.add(
Convolution1D(nb_filter=num_filters3,
filter_length=filter_length3,
border_mode='valid',
activation='relu',
subsample_length=stride_len3,
init='uniform'))
modelR.add(
Convolution1D(nb_filter=num_filters3,
filter_length=filter_length3,
border_mode='valid',
activation='relu',
subsample_length=stride_len3,
init='uniform'))
modelL.add(Dropout(0.30))
modelR.add(Dropout(0.30))
modelL.add(MaxPooling1D(pool_length=2))
modelR.add(MaxPooling1D(pool_length=2))
modelL.add(
Convolution1D(nb_filter=num_filters4,
filter_length=filter_length4,
border_mode='valid',
activation='relu',
subsample_length=stride_len4,
init='uniform'))
modelR.add(
Convolution1D(nb_filter=num_filters4,
filter_length=filter_length4,
border_mode='valid',
activation='relu',
subsample_length=stride_len4,
init='uniform'))
modelL.add(Dropout(0.25))
modelR.add(Dropout(0.25))
modelL.add(MaxPooling1D(pool_length=pool_len4))
modelR.add(MaxPooling1D(pool_length=pool_len4))
# We flatten the output of the conv layer,
# so that we can add a vanilla dense layer:
modelL.add(Flatten())
modelR.add(Flatten())
# We add a vanilla hidden layer:
modelL.add(Dense(hidden_dims))
modelL.add(Activation('relu'))
modelR.add(Dense(hidden_dims))
modelR.add(Activation('relu'))
merged_vector = merge([modelL, modelR], mode='concat', concat_axis=-1)
Gw = Dense(1, activation=DistanceMetric)(merged_vector)
model.compile(
loss=contrastiveLoss,
optimizer='rmsprop',
)