def build_model(self):
# Remark: Share weights for embedding is not supported.
# Thus here the model takes concatenated input and slice to split the input.
input = Input(name='input',
shape=(self.text1_length + self.text2_length, ))
embedding = Embedding(self.vocab_size,
self.embed_size,
weights=self.embed_weights,
trainable=self.train_embed)(input)
query_embed = embedding.slice(1, 0, self.text1_length)
doc_embed = embedding.slice(1, self.text1_length, self.text2_length)
mm = A.batch_dot(query_embed, doc_embed,
axes=[2, 2]) # Translation Matrix.
KM = []
for i in range(self.kernel_num):
mu = 1. / (self.kernel_num - 1) + (2. * i) / (self.kernel_num -
1) - 1.0
sigma = self.sigma
if mu > 1.0: # Exact match.
sigma = self.exact_sigma
mu = 1.0
mm_exp = A.exp(-0.5 * (mm - mu) * (mm - mu) / sigma / sigma)
mm_doc_sum = A.sum(mm_exp, 2)
mm_log = A.log(mm_doc_sum + 1.0)
# Remark: Keep the reduced dimension for the last sum and squeeze after stack.
# Otherwise, when batch=1, the output will become a Scalar not compatible for stack.
mm_sum = A.sum(mm_log, 1, keepDims=True)
KM.append(mm_sum)
Phi = Squeeze(2)(A.stack(KM, 1))
output = Dense(1, init="uniform", activation="sigmoid")(Phi)
model = Model(input=input, output=output)
return model
def _to_tensor(self):
input = self.model_inputs[0].zvalue
assert len(self.onnx_attr['axes']
) == 1, "we only support axes with 1 elements for now"
axes = self.onnx_attr['axes'][0]
keepdims = True if self.onnx_attr['keepdims'] == 1 else False
return autograd.sum(input, axis=int(axes), keepDims=keepdims)
def build(self):
def Kernel_layer(mu, sigma):
def kernel(x):
return A.exp(-0.5 * (x - mu) * (x - mu) / sigma / sigma)
return A.Lambda(lambda x: kernel(x)) #Activation(kernel)
input = klayers1.Input(name='input', shape=(50, ))
query = input.slice(
1, 0, 10
) #klayers1.Input(name='query', shape=(self.config['text1_maxlen'],))
#show_layer_info('Input', query)
doc = input.slice(
1, 10, 40
) #klayers1.Input(name='doc', shape=(self.config['text2_maxlen'],))
#show_layer_info('Input', doc)
embedding = klayers1.Embedding(
self.config['vocab_size'],
self.config['embed_size'],
name="embedding"
) # weights=[self.config['embed']], trainable=self.config['train_embed'],
q_embed, d_embed = self.share(embedding, query, doc)
#show_layer_info('Embedding', q_embed)
#show_layer_info('Embedding', d_embed)
mm = A.batch_dot(q_embed, d_embed, axes=[2, 2], normalize=False)
#show_layer_info('Dot', mm)
KM = []
for i in range(self.config['kernel_num']):
mu = 1. / (self.config['kernel_num'] -
1) + (2. * i) / (self.config['kernel_num'] - 1) - 1.0
sigma = self.config['sigma']
if mu > 1.0:
sigma = self.config['exact_sigma']
mu = 1.0
mm_exp = Kernel_layer(mu, sigma)(mm)
#show_layer_info('Exponent of mm:', mm_exp)
mm_doc_sum = A.Lambda(lambda x: A.sum(x, 2))(
mm_exp) #reduce_sum(x, 2)
#show_layer_info('Sum of document', mm_doc_sum)
mm_log = A.Lambda(lambda x: A.log(x + 1.0))(mm_doc_sum)
#show_layer_info('Logarithm of sum', mm_log)
mm_sum = A.Lambda(lambda x: A.sum(x, 1))(mm_log)
#show_layer_info('Sum of all exponent', mm_sum)
KM.append(mm_sum)
KMStack = A.stack(KM, 1)
Phi = A.Lambda(lambda x: x)(KMStack)
#show_layer_info('Stack', Phi)
if self.config['target_mode'] == 'classification':
out_ = Dense(2,
activation='softmax',
bias_initializer='zeros',
name="dense")(Phi)
elif self.config['target_mode'] in ['regression', 'ranking']:
out_ = Dense(1, bias_initializer='zero', name="dense")(Phi)
#show_layer_info('Dense', out_)
#timedistributed = layer1.TimeDistributed(out_, input_shape=(10, 12))
model = Model(input=input, output=[out_])
return model