def model_two_layer_small_embedding(alphabet):
return CSM(layers=[
DictionaryEncoding(vocabulary=alphabet),
WordEmbedding(dimension=32, vocabulary_size=len(alphabet)),
SentenceConvolution(n_feature_maps=5,
kernel_width=10,
n_channels=1,
n_input_dimensions=32),
SumFolding(),
KMaxPooling(k=7),
Bias(n_input_dims=16, n_feature_maps=5),
Tanh(),
SentenceConvolution(n_feature_maps=5,
kernel_width=5,
n_channels=5,
n_input_dimensions=16),
KMaxPooling(k=4),
Bias(n_input_dims=16, n_feature_maps=5),
Tanh(),
Softmax(n_classes=2, n_input_dimensions=320),
])
def model_one_layer_variant_2(alphabet):
return CSM(layers=[
DictionaryEncoding(vocabulary=alphabet),
WordEmbedding(dimension=42, vocabulary_size=len(alphabet)),
SentenceConvolution(n_feature_maps=5,
kernel_width=6,
n_channels=1,
n_input_dimensions=42),
SumFolding(),
KMaxPooling(k=4),
Bias(n_input_dims=21, n_feature_maps=5),
Tanh(),
Softmax(n_classes=2, n_input_dimensions=420),
])
validation_data_provider = LabelledSequenceMinibatchProvider(
X=X[-500:], Y=Y[-500:], batch_size=500)
word_embedding_model = CSM(layers=[
WordEmbedding( # really a character embedding
dimension=16,
vocabulary_size=len(alphabet)),
SentenceConvolution(n_feature_maps=10,
kernel_width=5,
n_channels=1,
n_input_dimensions=16),
SumFolding(),
KMaxPooling(k=2),
MaxFolding(),
Tanh(),
])
word_embedding = WordFromCharacterEmbedding(
embedding_model=word_embedding_model, alphabet_encoding=alphabet)
# print word_embedding.fprop(X, meta)
tweet_model = CSM(layers=[
word_embedding,
SentenceConvolution(n_feature_maps=5,
kernel_width=10,
n_channels=1,
n_input_dimensions=80),
SumFolding(),
KMaxPooling(k=7),
model = CSM(layers=[
WordEmbedding(dimension=embedding_dimension,
vocabulary_size=len(alphabet)),
SentenceConvolution(n_feature_maps=n_feature_maps,
kernel_width=kernel_width,
n_channels=1,
n_input_dimensions=embedding_dimension),
SumFolding(),
KMaxPooling(k=pooling_size),
# Bias(
# n_input_dims=embedding_dimension / 2,
# n_feature_maps=n_feature_maps),
Linear(n_input=n_feature_maps * pooling_size * embedding_dimension / 2,
n_output=64),
Tanh(),
Linear(n_output=1, n_input=64),
])
print model
cost_function = LargeMarginCost(0.1)
noise_model = RandomAlphabetCorruption(alphabet)
objective = NoiseContrastiveObjective(cost=cost_function,
data_provider=train_data_provider,
noise_model=noise_model)
update_rule = AdaGrad(gamma=0.1, model_template=model)
optimizer = SGD(model=model, objective=objective, update_rule=update_rule)