train_ds = nds.VarLenDataSet(train_embed, train_label)
dev_ds = nds.VarLenDataSet(dev_embed, dev_label)
test_ds = nds.VarLenDataSet(test_embed, test_label)
# Build Computation Graph
graph = ng.Graph(nl.LogLoss(), ns.Adam(eta=0.01))
# Word Embedding Matrix using Xavier
word_embedding = graph.param_of([len(word_dict), wv_dim])
# Weight vector
weight = graph.param_of([wv_dim, 1])
input_node = graph.input()
embed = nd.Embed(input_node, word_embedding)
average = nd.Average(embed)
dot = nd.Dot(average, weight)
sigmoid = nd.Sigmoid(dot)
graph.output(sigmoid)
epochs = 100
batch_size = 50
def train():
loss_sum = 0.0
batch_counter = 0
for batch in train_ds.batches(batch_size):
input_node.value = batch.data
graph.expect(np.float64(batch.expect).reshape(-1, 1))
loss, accuracy = graph.train()
loss_sum += loss
self.smax = smax
def compute(self):
return np.einsum("bld,bl->bd", self.embed.value, self.smax.value)
def updateGrad(self):
self.embed.grad += np.einsum("bd,bl->bld", self.grad, self.smax.value)
self.smax.grad += np.einsum("bd,bld->bl", self.grad, self.embed.value)
input_node = graph.input()
embed = nd.Embed(input_node, word_embedding)
mapped = EmbedMap(embed, attention_weight)
softmax = nd.SoftMax(mapped)
attention = Attention(embed, softmax)
dot = nd.Dot(attention, weight)
sigmoid = nd.Sigmoid(dot)
graph.output(sigmoid)
epochs = 100
batch_size = 50
def train():
loss_sum = 0.0
batch_counter = 0
for batch in train_ds.batches(batch_size):
input_node.value = batch.data
graph.expect(np.float64(batch.expect).reshape(-1, 1))
loss, accuracy = graph.train()
loss_sum += loss