def test_lookup_table():
lt = LookupTable(5, 3)
lt.allocate()
lt.W.set_value(numpy.arange(15).reshape(5, 3).astype(theano.config.floatX))
x = tensor.lmatrix("x")
y = lt.apply(x)
f = theano.function([x], [y])
x_val = [[1, 2], [0, 3]]
desired = numpy.array([[[3, 4, 5], [6, 7, 8]], [[0, 1, 2], [9, 10, 11]]],
dtype=theano.config.floatX)
assert_equal(f(x_val)[0], desired)
# Test get_dim
assert_equal(lt.get_dim(lt.apply.inputs[0]), 0)
assert_equal(lt.get_dim(lt.apply.outputs[0]), lt.dim)
assert_raises(ValueError, lt.get_dim, 'random_name')
# Test feedforward interface
assert lt.input_dim == 0
assert lt.output_dim == 3
lt.output_dim = 4
assert lt.output_dim == 4
def assign_input_dim():
lt.input_dim = 11
assert_raises(ValueError, assign_input_dim)
lt.input_dim = 0
def build_model(self, x, config):
logger.info('building %s model for: %s ', self.nn_model, self.name)
vocabsize = self.get_vocab_size()
logger.info('%s vocab size is: %d', self.name, vocabsize)
self.embeddings, self.dim_emb = self.get_embeddings()
if self.tune_tune:
logger.info('%s lookuptable with size (%d, %d) will be tuned.', self.name, vocabsize, self.dim_emb)
lookup = LookupTable(length=vocabsize, dim=self.dim_emb)
lookup.allocate()
# add_role(lookup.W, WEIGHT)
lookup.W.name = 'lt.W'
else:
logger.info('%s lookuptable with size (%d, %d) will NOT be tuned.', self.name, vocabsize, self.dim_emb)
lookup = MyLookupTable(length=vocabsize, dim=self.dim_emb)
lookup.allocate()
lookup.name = self.name + 'lookuptable'
lookup.W.set_value(self.embeddings)
xemb = lookup.apply(x)
xemb = debug_print(xemb, 'xemb', False)
if 'cnn' in self.nn_model:
logger.info('CNN')
feature_vec, feature_vec_len = create_cnn_general(xemb, self.dim_emb, self.max_len, config, self.name)
elif self.nn_model == 'lstm':
feature_vec, feature_vec_len = create_lstm(xemb, self.dim_emb, False, config, self.name)
elif self.nn_model == 'bilstm':
feature_vec, feature_vec_len = create_lstm(xemb, self.dim_emb, True, config, self.name)
elif self.nn_model == 'rnn':
feature_vec, feature_vec_len = create_rnn(xemb, self.dim_emb, config, self.name)
elif self.nn_model == 'ff':
feature_vec, feature_vec_len = create_ff(xemb, self.dim_emb, self.max_len, config)
elif self.nn_model == 'mean':
feature_vec, feature_vec_len = create_mean(xemb, self.dim_emb, self.max_len, config)
return feature_vec, feature_vec_len
def test_lookup_table():
lt = LookupTable(5, 3)
lt.allocate()
lt.W.set_value(numpy.arange(15).reshape(5, 3).astype(theano.config.floatX))
x = tensor.lmatrix("x")
y = lt.apply(x)
f = theano.function([x], [y])
x_val = [[1, 2], [0, 3]]
desired = numpy.array([[[3, 4, 5], [6, 7, 8]], [[0, 1, 2], [9, 10, 11]]],
dtype=theano.config.floatX)
assert_equal(f(x_val)[0], desired)
# Test get_dim
assert_equal(lt.get_dim(lt.apply.inputs[0]), 0)
assert_equal(lt.get_dim(lt.apply.outputs[0]), lt.dim)
assert_raises(ValueError, lt.get_dim, 'random_name')
# Test feedforward interface
assert lt.input_dim == 0
assert lt.output_dim == 3
lt.output_dim = 4
assert lt.output_dim == 4
def assign_input_dim():
lt.input_dim = 11
assert_raises(ValueError, assign_input_dim)
lt.input_dim = 0
def test_lookup_table():
lt = LookupTable(5, 3)
lt.allocate()
lt.W.set_value(numpy.arange(15).reshape(5, 3).astype(theano.config.floatX))
x = tensor.lmatrix("x")
y = lt.apply(x)
f = theano.function([x], [y])
x_val = [[1, 2], [0, 3]]
desired = numpy.array([[[3, 4, 5], [6, 7, 8]], [[0, 1, 2], [9, 10, 11]]],
dtype=theano.config.floatX)
assert_equal(f(x_val)[0], desired)
def build_model(self, x, config):
logger.info('building %s model for: %s ', self.nn_model, self.name)
vocabsize = self.get_vocab_size()
logger.info('%s vocab size is: %d', self.name, vocabsize)
self.embeddings, self.dim_emb = self.get_embeddings()
if self.tune_tune:
logger.info('%s lookuptable with size (%d, %d) will be tuned.',
self.name, vocabsize, self.dim_emb)
lookup = LookupTable(length=vocabsize, dim=self.dim_emb)
lookup.allocate()
# add_role(lookup.W, WEIGHT)
lookup.W.name = 'lt.W'
else:
logger.info('%s lookuptable with size (%d, %d) will NOT be tuned.',
self.name, vocabsize, self.dim_emb)
lookup = MyLookupTable(length=vocabsize, dim=self.dim_emb)
lookup.allocate()
lookup.name = self.name + 'lookuptable'
lookup.W.set_value(self.embeddings)
xemb = lookup.apply(x)
xemb = debug_print(xemb, 'xemb', False)
if 'cnn' in self.nn_model:
logger.info('CNN')
feature_vec, feature_vec_len = create_cnn_general(
xemb, self.dim_emb, self.max_len, config, self.name)
elif self.nn_model == 'lstm':
feature_vec, feature_vec_len = create_lstm(xemb, self.dim_emb,
False, config,
self.name)
elif self.nn_model == 'bilstm':
feature_vec, feature_vec_len = create_lstm(xemb, self.dim_emb,
True, config, self.name)
elif self.nn_model == 'rnn':
feature_vec, feature_vec_len = create_rnn(xemb, self.dim_emb,
config, self.name)
elif self.nn_model == 'ff':
feature_vec, feature_vec_len = create_ff(xemb, self.dim_emb,
self.max_len, config)
elif self.nn_model == 'mean':
feature_vec, feature_vec_len = create_mean(xemb, self.dim_emb,
self.max_len, config)
return feature_vec, feature_vec_len
activation=Tanh(),
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
), )
### Will need to reshape the rnn outputs to produce suitable input here...
gather = Linear(name='hidden_to_output',
input_dim=hidden_dim * 2,
output_dim=labels_size,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0))
p_labels = Softmax()
## Let's initialize the variables
lookup.allocate()
#print("lookup.parameters=", lookup.parameters) # ('lookup.parameters=', [W])
#lookup.weights_init = FUNCTION
#lookup.initialize()
#lookup.params[0].set_value( np.random.normal( scale = 0.1, size=(vocab_size, embedding_dim) ).astype(np.float32) )
#lookup.params[0].set_value( embedding )
# See : https://github.com/mila-udem/blocks/blob/master/tests/bricks/test_lookup.py
#lookup.W.set_value(numpy.arange(15).reshape(5, 3).astype(theano.config.floatX))
lookup.W.set_value(embedding.astype(theano.config.floatX))
rnn.initialize()
gather.initialize()
),
)
### Will need to reshape the rnn outputs to produce suitable input here...
gather = Linear(name='hidden_to_output',
input_dim=hidden_dim*2, output_dim=labels_size,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0)
)
p_labels = Softmax()
## Let's initialize the variables
lookup.allocate()
#print("lookup.parameters=", lookup.parameters) # ('lookup.parameters=', [W])
#lookup.weights_init = FUNCTION
#lookup.initialize()
#lookup.params[0].set_value( np.random.normal( scale = 0.1, size=(vocab_size, embedding_dim) ).astype(np.float32) )
#lookup.params[0].set_value( embedding )
# See : https://github.com/mila-udem/blocks/blob/master/tests/bricks/test_lookup.py
#lookup.W.set_value(numpy.arange(15).reshape(5, 3).astype(theano.config.floatX))
lookup.W.set_value( embedding.astype(theano.config.floatX) )
rnn.initialize()
gather.initialize()