def get_model(encoding):
return CSM(
layers=[
DictionaryEncoding(vocabulary=encoding),
WordEmbedding(
dimension={{embedding_dimension}},
vocabulary_size=len(encoding),
padding=encoding['PADDING']),
{% for layer in word_layers %}
{% set layer_index = loop.index0 %}
SentenceConvolution(
n_feature_maps={{layer.n_feature_maps}},
kernel_width={{layer.kernel_width}},
n_channels={{layer.n_channels}},
n_input_dimensions=1),
Bias(
n_input_dims=1,
n_feature_maps={{layer.n_feature_maps}}),
KMaxPooling(k={{layer.k_pooling}}, k_dynamic={{layer.k_dynamic}} if {{layer.k_dynamic}} > 0 else None),
{{layer.nonlinearity}}(),
{% endfor %}
ReshapeForDocuments(),
{% for layer in sentence_layers %}
{% set layer_index = loop.index0 %}
SentenceConvolution(
n_feature_maps={{layer.n_feature_maps}},
kernel_width={{layer.kernel_width}},
n_channels={{layer.n_channels}},
n_input_dimensions=1),
Bias(
n_input_dims=1,
n_feature_maps={{layer.n_feature_maps}}),
KMaxPooling(k={{layer.k_pooling}}, k_dynamic={{layer.k_dynamic}} if {{layer.k_dynamic}} > 0 else None),
{{layer.nonlinearity}}(),
{% endfor %}
{% if dropout %}
Dropout(('b', 'd', 'f', 'w'), 0.5),
{% endif %}
Softmax(
n_classes={{n_classes}},
n_input_dimensions={{softmax_input_dimensions}}),
])
def _sentence_convolution(conv_layer, ratio):
new_conv = SentenceConvolution(
n_feature_maps=conv_layer.n_feature_maps,
kernel_width=conv_layer.kernel_width,
n_channels=conv_layer.n_channels,
n_input_dimensions=conv_layer.n_input_dimensions)
new_conv.W = conv_layer.W.copy() * (1-ratio)
return new_conv
def _sentence_convolution(conv_layer, ratio):
new_conv = SentenceConvolution(
n_feature_maps=conv_layer.n_feature_maps,
kernel_width=conv_layer.kernel_width,
n_channels=conv_layer.n_channels,
n_input_dimensions=conv_layer.n_input_dimensions)
new_conv.W = conv_layer.W.copy() * (1 - ratio)
return new_conv
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),
])
train_data_provider = LabelledSequenceMinibatchProvider(X=X[:-500],
Y=Y[:-500],
batch_size=100)
print train_data_provider.batches_per_epoch
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,
train_data_provider = PaddedSequenceMinibatchProvider(
X=data, padding=alphabet['PADDING'], batch_size=100)
embedding_dimension = 8
vocabulary_size = len(alphabet)
n_feature_maps = 8
kernel_width = 5
pooling_size = 2
n_epochs = 1
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
# Softmax(
# n_classes=2,
# n_input_dimensions=480),
# ]
# )
tweet_model = CSM(layers=[
# cpu.model.encoding.
DictionaryEncoding(vocabulary=alphabet),
# cpu.model.embedding.
WordEmbedding(dimension=28, vocabulary_size=len(alphabet)),
# HostToDevice(),
SentenceConvolution(n_feature_maps=6,
kernel_width=7,
n_channels=1,
n_input_dimensions=28),
Bias(n_input_dims=28, n_feature_maps=6),
SumFolding(),
KMaxPooling(k=4, k_dynamic=0.5),
Tanh(),
SentenceConvolution(n_feature_maps=14,
kernel_width=5,
n_channels=6,
n_input_dimensions=14),
Bias(n_input_dims=14, n_feature_maps=14),
SumFolding(),
KMaxPooling(k=4),
Tanh(),
Softmax(n_classes=2, n_input_dimensions=392),
])
validation_data_provider = LabelledDocumentMinibatchProvider(
X=X[-n_validation:],
Y=Y[-n_validation:],
batch_size=batch_size,
padding='PADDING',
fixed_n_sentences=15,
fixed_n_words=50)
model = CSM(layers=[
DictionaryEncoding(vocabulary=encoding),
WordEmbedding(dimension=20,
vocabulary_size=len(encoding),
padding=encoding['PADDING']),
Dropout(('b', 'w', 'f'), 0.2),
SentenceConvolution(n_feature_maps=10,
kernel_width=15,
n_channels=20,
n_input_dimensions=1),
Bias(n_input_dims=1, n_feature_maps=10),
KMaxPooling(k=7, k_dynamic=0.5),
Tanh(),
SentenceConvolution(n_feature_maps=30,
kernel_width=9,
n_channels=10,
n_input_dimensions=1),
Bias(n_input_dims=1, n_feature_maps=30),
KMaxPooling(k=5),
Tanh(),
ReshapeForDocuments(),
SentenceConvolution(n_feature_maps=20,
kernel_width=11,
n_channels=30 * 5,