def dot_attention(query, key, value, mask, dropout=0.0):
# query: (batch_size, h, length_q, model_dim/h)
# key: (batch_size, h, length_k, model_dim/h)
# value: (batch_size, h, length_k, model_dim/h)
query_shape = query.shape
query = query.reshape(-3, -2)
key = key.reshape(-3, -2)
value = value.reshape(-3, -2)
# matmul, t: (batch_size*h, length_q, length_k)
t = nd.batch_dot(query, key.swapaxes(1, 2)) / math.sqrt(query.shape[-1])
# masked
# mask PAD and future words
m = nd.full(t.shape, LARGE_NEGATIVE_VALUE)
mask = nd.ones(t.shape) * mask
t = nd.where(mask, t, m)
# softmax
t = nd.softmax(t, axis=-1)
if dropout > 0.0:
t = nd.dropout(t, p=dropout)
# (batch_size, h, length_q, model_dim/h)
return nd.batch_dot(t, value).reshape(query_shape)
def _train_batch(self, batch, gibbs_sampling_steps, learning_rate):
"""Performs k-step Contrastive Divergence (CD-k) learning.
Updates weights and biases.
Keep in mind that most variables are "batch" tensors.
Variable name suffix "_pr" stands for Pr. (probability).
"""
hidden_pr, hidden, dreamed_visible, dreamed_hidden_pr = self.gibbs_sampling_step(
batch)
positive_phase = nd.batch_dot(self._transpose_batch(batch), hidden)
for _ in range(gibbs_sampling_steps - 1):
_, _, dreamed_visible, dreamed_hidden_pr = self.gibbs_sampling_step(
dreamed_visible)
negative_phase = nd.batch_dot(self._transpose_batch(dreamed_visible),
dreamed_hidden_pr)
# make learning rate independent from the batch size
learning_rate = learning_rate / batch.shape[0]
self.weights += learning_rate * nd.sum(positive_phase - negative_phase,
axis=(0, ))
if self.hidden_bias is not None:
self.hidden_bias += learning_rate * nd.sum(
hidden_pr - dreamed_hidden_pr, axis=(0, ))
if self.visible_bias is not None:
self.visible_bias += learning_rate * nd.sum(
batch - dreamed_visible, axis=(0, ))
def forward(self, query, key, value, mask=None):
d = query.shape[-1]
scores = nd.batch_dot(query, key, transpose_b=True) / math.sqrt(d)
attention_weights = nlp.model.attention_cell._masked_softmax(
nd, scores, mask, scores.dtype)
attention_weights = self.dropout(attention_weights)
return nd.batch_dot(attention_weights, value)
def forward(self, input_data):
freq = input_data[:, 0:2].expand_dims(1)
input_data = input_data[:, 2:]
e1_vec_start = FIXED_WORD_LENGTH * DIMENSION
x = input_data[:, :e1_vec_start].reshape(
(input_data.shape[0], FIXED_WORD_LENGTH,
DIMENSION)) # (m, 60, 110)
e1neimask = input_data[:, e1_vec_start:e1_vec_start +
MASK_LENGTH] # (m, 51)
e1edge = input_data[:, e1_vec_start + MASK_LENGTH:e1_vec_start +
MASK_LENGTH + ENTITY_EDGE_VEC_LENGTH].reshape(
(input_data.shape[0], ENTITY_DEGREE,
WORD_DIMENSION * 2)) # (m, 51, 200)
e1neigh = e1edge[:, :, :WORD_DIMENSION]
e2_vec_start = e1_vec_start + MASK_LENGTH + ENTITY_EDGE_VEC_LENGTH
e2neimask = input_data[:, e2_vec_start:e2_vec_start +
MASK_LENGTH] # (m, 51)
e2edge = input_data[:, e2_vec_start + MASK_LENGTH:e2_vec_start +
MASK_LENGTH + ENTITY_EDGE_VEC_LENGTH].reshape(
(input_data.shape[0], ENTITY_DEGREE,
WORD_DIMENSION * 2)) # (m, 51,200)
e2neigh = e2edge[:, :, :WORD_DIMENSION]
gru = self.gru
x = nd.transpose(x, axes=(1, 0, 2))
h = gru(x)
ht = nd.transpose(h, axes=(1, 0, 2))
gru_out = self.gru_out
y1 = gru_out(ht.expand_dims(1)) # (m,200)
att = self.center_att
e1edge = nd.tanh(e1edge)
e1g = att(e1edge) * freq[:, :, :1] # (m,51,1)
e1g = e1g * e1neimask.expand_dims(2)
e1g = nd.softmax(e1g, axis=1)
e1gt = nd.transpose(e1g, axes=(0, 2, 1)) # (m,1,151)
e1n = nd.batch_dot(e1gt, e1neigh) # (m,1,100)
e1n = e1n.reshape((e1n.shape[0], 100)) # (m,100)
e2edge = nd.tanh(e2edge)
e2g = att(e2edge) * freq[:, :, 1:] # (m,51,1)
e2g = e2g * e2neimask.expand_dims(2)
e2g = nd.softmax(e2g, axis=1)
e2gt = nd.transpose(e2g, axes=(0, 2, 1)) # (m,1,151)
e2n = nd.batch_dot(e2gt, e2neigh) # (m,1,100)
e2n = e2n.reshape((e2n.shape[0], 100)) # (m,100)
center_y = nd.concat(e1n, e2n, dim=1) # (m,200)
center_out = self.center_out
center_y = center_out(center_y)
out = self.output
y4 = nd.concat(y1, center_y, dim=1)
y5 = out(y4)
return y5
def forward(self, x):
"""Forward Relation Module.
Parameters
----------
feat : mxnet.nd.NDArray or mxnet.symbol
(M, 1024) Feature tensor (used to compute q).
ctx_feat : mxnet.nd.NDArray or mxnet.symbol
(N, 1024)Contextual Feature tensor (used to compute k,v).
box: mxnet.nd.NDArray or mxnet.symbol
(M, 4) boxes with corner encoding.
ctx_box: mxnet.nd.NDArray or mxnet.symbol
(N, 4) boxes with corner encoding.
Returns
-------
gt_relation_feat, ctx_relation_feat
(M, 1024).
"""
e = self.dim_k # e = 1024 (feature size)
k, v, q = x.shape[0], x.shape[0], x.shape[
0] # k, v, q = N (number of bounding boxes)
h = self.num_group # h = 16 (Number of groups or num_group for multi head attention)
x = x.reshape(k, h, e)
x = x.reshape(k * h, e)
keys = self.to_keys(x).reshape(k, h, e).transpose(
axes=(1, 0, 2)) # keys : (h, k, e)
values = self.to_values(x).reshape(k, h, e).transpose(
axes=(1, 0, 2)) # values : (h, v, e)
queries = self.to_queries(x).reshape(k, h, e).transpose(
axes=(1, 0, 2)) # queries : (h, q, e)
keys = keys / (self.num_feat**(1 / 4))
queries = queries / (self.num_feat**(1 / 4))
dot = F.batch_dot(lhs=queries,
rhs=keys,
transpose_a=False,
transpose_b=True) # dot : (h, q, k)
attention = F.softmax(dot, axis=2)
out = F.batch_dot(lhs=attention,
rhs=values,
transpose_a=False,
transpose_b=False) # out : (h, q, e)
out = out.transpose(axes=(1, 0, 2)) # out : (q, h, e)
out = out.reshape(q, -1) # out : (q, h*e)
out = self.unify_heads(out) # out : (q, e)
return out
def _get_co_attention(as_, bs_, r, lamb=k_lambda):
"""
as_, bs_: (batch_size, seq_len, embed_size)
r: (batch_size, seq_len, seq_len, 5)
"""
e = nd.batch_dot(as_, bs_, transpose_b=True) + lamb * F(
r, ctx) # (batch_size, seq_len, seq_len,)
alpha = nd.softmax(e, axis=2) # alpha_ij = exp(eij) / SUM_k(exp(eik))
beta = nd.softmax(e, axis=1) # beta_ij = exp(ij) / SUM_k(exp(ekj))
beta = nd.transpose(beta,
axes=[0, 2,
1]) # transpose becasue of softmax axis=1
ac = nd.batch_dot(alpha, bs_) #
bc = nd.batch_dot(beta, as_)
return ac, bc, alpha, beta
def forward(self, query, key, value, valid_length):
"""Forward function"""
query, key = self.W_k(query), self.W_q(key)
features = query.expand_dims(axis=2) + key.expand_dims(axis=1)
scores = self.v(features).squeeze(axis=-1)
attention_weights = self.dropout(masked_softmax(scores, valid_length))
return nd.batch_dot(attention_weights, value)
def forward(self, x):
'''
Parameters
----------
x: mx.ndarray, shape is (batch_size, N, C_{r-1}, T_{r-1})
Returns
----------
mx.ndarray, shape is (batch_size, N, num_of_time_filters, T_{r-1})
'''
(batch_size, num_of_vertices,
num_of_features, num_of_timesteps) = x.shape
# shape is (batch_size, T, T)
temporal_At = self.TAt(x)
x_TAt = nd.batch_dot(x.reshape(batch_size, -1, num_of_timesteps),
temporal_At)\
.reshape(batch_size, num_of_vertices,
num_of_features, num_of_timesteps)
# cheb gcn with spatial attention
spatial_At = self.SAt(x_TAt)
spatial_gcn = self.cheb_conv_SAt(x, spatial_At)
# convolution along time axis
time_conv_output = (self.time_conv(spatial_gcn.transpose((0, 2, 1, 3)))
.transpose((0, 2, 1, 3)))
# residual shortcut
x_residual = (self.residual_conv(x.transpose((0, 2, 1, 3)))
.transpose((0, 2, 1, 3)))
return self.ln(nd.relu(x_residual + time_conv_output))
def forward(self, x1, x2):
y1 = self.mlp(x1)
y2 = self.mlp(x2)
# re-shape it
y1 = y1.expand_dims(axis=1) # add dummy dimension
y2 = y2.expand_dims(axis=2) # Y1: (N, 1, C) Y2: (N, C, 1)
return nd.batch_dot(y1, y2)
def forward(self, query, key, value, valid_length):
"""Forward function"""
query, key = self.W_k(query), self.W_q(key)
features = query.expand_dims(axis=2) + key.expand_dims(axis=1)
scores = self.v(features).squeeze(axis=-1)
attention_weights = self.dropout(masked_softmax(scores, valid_length))
return nd.batch_dot(attention_weights, value)
def _calculate_trilinear_similarity(self, context, query, context_max_len,
query_max_len, w4mlu, bias):
"""Implement the computation of trilinear similarity function.
refer https://github.com/NLPLearn/QANet/blob/master/layers.py#L505
The similarity function is:
f(w, q) = W[w, q, w * q]
where w and q represent the word in context and query respectively,
and * operator means hadamard product.
Parameters
-----------
context : NDArray
input tensor with shape `(batch_size, context_sequence_length, hidden_size)`
query : NDArray
input tensor with shape `(batch_size, query_sequence_length, hidden_size)`
context_max_len : int
context_max_len : int
Returns
--------
similarity_mat : NDArray
output tensor with shape `(batch_size, context_sequence_length, query_sequence_length)`
"""
subres0 = nd.tile(self.w4c(context), [1, 1, query_max_len])
subres1 = nd.tile(nd.transpose(self.w4q(query), axes=(0, 2, 1)),
[1, context_max_len, 1])
subres2 = nd.batch_dot(w4mlu * context,
nd.transpose(query, axes=(0, 2, 1)))
similarity_mat = subres0 + subres1 + subres2 + bias
return similarity_mat
def attention(query, key, value, mask=None, dropout=None):
# Q * K.transpose() * value
assert (len(query.shape) == 3)
assert (len(key.shape) == 3)
assert (len(value.shape) == 3)
d_model = query.shape[-1]
scores = nd.batch_dot(query, key, transpose_b=True) / math.sqrt(d_model)
if mask is not None:
val = nd.ones(scores.shape, ctx=cfg.ctx) * (-1e9)
scores = nd.where(mask == 1, scores, val)
p_attn = nd.softmax(scores, axis=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return nd.batch_dot(p_attn, value), p_attn
def forward(self, cur_input, state, encoder_outputs):
# 当循环神经网络有多个隐藏层时，取靠近输出层的单层隐藏状态
single_layer_state = [state[0][-1].expand_dims(0)]
encoder_outputs = encoder_outputs.reshape((self.max_seq_len, -1,
self.encoder_num_hiddens))
hidden_broadcast = nd.broadcast_axis(single_layer_state[0], axis=0,
size=self.max_seq_len)
encoder_outputs_and_hiddens = nd.concat(encoder_outputs,
hidden_broadcast, dim=2)
energy = self.attention(encoder_outputs_and_hiddens)
batch_attention = nd.softmax(energy, axis=0).transpose((1, 2, 0))
batch_encoder_outputs = encoder_outputs.swapaxes(0, 1)
decoder_context = nd.batch_dot(batch_attention, batch_encoder_outputs)
#改这里
input_and_context = nd.concat(nd.expand_dims(self.embedding(cur_input), axis=1),
decoder_context, dim=2)
concat_input = self.rnn_concat_input(input_and_context).reshape((1, -1, 0))
concat_input = self.dropout(concat_input)
state = [nd.broadcast_axis(single_layer_state[0], axis=0,size=self.num_layers)]
output, state = self.rnn(concat_input, state)
output = self.dropout(output)
output = self.out(output).reshape((-3, -1))
return output, state
def calculate_loss(x, y, model, loss, loss_name, class_weight, penalization_coeff):
"""calculate loss value
Args:
x (NDArray): intput of model
y (NDArray): target
model (Block): model
loss (gluon.loss): loss function
loss_name (str): name of loss function
class_weight (NDArray): weight of sample loss value for each category
penalization_coeff (float): Attention penalty coefficient
Returns:
NDArray: output of model
NDArray: loss value
"""
pred, att = model(x)
if loss_name == 'sce':
l = loss(pred, y)
elif loss_name == 'wsce':
l = loss(pred, y, class_weight, class_weight.shape[0])
# penalty
diversity_penalty = nd.batch_dot(att, nd.transpose(att, axes=(0, 2, 1))
) - nd.eye(att.shape[1], ctx=att.context)
l = l + penalization_coeff * diversity_penalty.norm(axis=(1, 2))
return pred, l
def forward(self, x):
if self.routing is not None:
routing_weight = nd.softmax(nd.zeros(shape=(1, 1, self.num_points),
ctx=x.context),
axis=2)
trans = self.stn(x)
x = nd.transpose(x, (0, 2, 1))
x = nd.batch_dot(x, trans)
x = nd.transpose(x, (0, 2, 1))
x = nd.relu(self.bn1(self.conv1(x)))
pointfeat = x
x = nd.relu(self.bn2(self.conv2(x)))
x = self.bn3(self.conv3(x))
if self.routing is not None:
s = nd.sum(x * routing_weight, axis=2, keepdims=True)
# v = Squash(s, axis=1)
for _ in range(self.routing):
routing_weight = routing_weight + nd.sum(
x * s, axis=1, keepdims=True)
c = nd.softmax(routing_weight, axis=2)
s = nd.sum(x * c, axis=2, keepdims=True)
# v = Squash(s, axis=1)
x = s
else:
x = self.mp1(x)
if self.global_feat:
return x, trans
else:
x = x.repeat(self.num_points, axis=2)
return nd.concat(x, pointfeat, dim=1), trans
def matmul(self, x, y, transpose_a=False,transpose_b=False):
x = nd.split(x, self.embedding_size, 2)
y = nd.split(y, self.embedding_size, 2)
res = []
for idx in range(self.embedding_size):
array = nd.batch_dot(x[idx], y[idx], transpose_a,transpose_b=transpose_b)
res.append(array.asnumpy().tolist())
return nd.array(res,ctx=self.ctx)
def augment(points, xforms, r=None):
points_xformed = nd.batch_dot(points, xforms, name='points_xformed')
if r is None:
return points_xformed
jitter_data = r * mx.random.normal(shape=points_xformed.shape)
jitter_clipped = nd.clip(jitter_data, -5 * r, 5 * r, name='jitter_clipped')
return points_xformed + jitter_clipped
def forward(self, decoder_output, encoder_output):
"""TODO: Docstring for forward.
:decoder_output: TODO
:encoder_output: TODO
:returns: TODO
"""
decoder_output = decoder_output.transpose([0, 2, 1])
score = nd.batch_dot(encoder_output, decoder_output)
weight = nd.softmax(score, axis=1)
context = nd.batch_dot(nd.transpose(weight, [0, 2, 1]), encoder_output)
return context, nd.squeeze(weight)
def bilinear(x,
W,
y,
input_size,
seq_len,
batch_size,
num_outputs=1,
bias_x=False,
bias_y=False):
"""Do xWy
Parameters
----------
x : NDArray
(input_size x seq_len) x batch_size
W : NDArray
(num_outputs x ny) x nx
y : NDArray
(input_size x seq_len) x batch_size
input_size : int
input dimension
seq_len : int
sequence length
batch_size : int
batch size
num_outputs : int
number of outputs
bias_x : bool
whether concat bias vector to input x
bias_y : bool
whether concat bias vector to input y
Returns
-------
output : NDArray
[seq_len_y x seq_len_x if output_size == 1 else seq_len_y x num_outputs x seq_len_x]
x batch_size
"""
if bias_x:
x = nd.concat(x, nd.ones((1, seq_len, batch_size)), dim=0)
if bias_y:
y = nd.concat(y, nd.ones((1, seq_len, batch_size)), dim=0)
ny = input_size + bias_y
# W: (num_outputs x ny) x nx
lin = nd.dot(W, x)
if num_outputs > 1:
lin = reshape_fortran(lin, (ny, num_outputs * seq_len, batch_size))
y = y.transpose([2, 1, 0]) # May cause performance issues
lin = lin.transpose([2, 1, 0])
blin = nd.batch_dot(lin, y, transpose_b=True)
blin = blin.transpose([2, 1, 0])
if num_outputs > 1:
blin = reshape_fortran(blin,
(seq_len, num_outputs, seq_len, batch_size))
return blin
def forward(self, x, spatial_attention, cheb_polynomials):
'''
Chebyshev graph convolution operation
Parameters
----------
x: mx.ndarray, graph signal matrix
shape is (batch_size, N, F, T_{r-1}), F is the num of features
spatial_attention: mx.ndarray, shape is (batch_size, N, N)
spatial attention scores
Returns
----------
mx.ndarray, shape is (batch_size, N, self.num_of_filters, T_{r-1})
'''
(batch_size, num_of_vertices,
num_of_features, num_of_timesteps) = x.shape
self.Theta.shape = (self.K, num_of_features, self.num_of_filters)
self.Theta._finish_deferred_init()
cur_context = x.context
outputs = []
for time_step in range(num_of_timesteps):
# shape is (batch_size, V, F)
graph_signal = x[:, :, :, time_step]
output = nd.zeros(shape=(batch_size, num_of_vertices,
self.num_of_filters), ctx=x.context)
for k in range(self.K):
# shape of T_k is (V, V)
T_k = cheb_polynomials[k].tostype('default').as_in_context(cur_context)
# print("T_K: ", T_k)
# shape of T_k_with_at is (batch_size, V, V)
T_k_with_at = T_k * spatial_attention
# T_k_with_at = T_k.as_in_context(cur_context) * spatial_attention
# print("T_k_with_at: ", T_k_with_at)
# shape of theta_k is (F, num_of_filters)
theta_k = self.Theta.data(cur_context)[k]
# shape is (batch_size, V, F)
# rhs = nd.batch_dot(T_k_with_at.transpose((0, 2, 1)).tostype('csr'),
# graph_signal)
# print("T_k_with_at: ", T_k_with_at)
# print("graph signal: ", graph_signal)
rhs = nd.batch_dot(T_k_with_at, graph_signal)
# print("rhs: ", rhs)
# print("theta_k: ", theta_k)
output = output + nd.dot(rhs, theta_k)
outputs.append(output.expand_dims(-1))
return nd.relu(nd.concat(*outputs, dim=-1))
def forward(self, x):
'''
Parameters
----------
x: mx.ndarray, x^{(r - 1)}_h
shape is (batch_size, N, C_{r-1}, T_{r-1})
Returns
----------
E_normalized: mx.ndarray, S', spatial attention scores
shape is (batch_size, T_{r-1}, T_{r-1})
'''
_, num_of_vertices, num_of_features, num_of_timesteps = x.shape
# defer shape
self.U_1.shape = (num_of_vertices, )
self.U_2.shape = (num_of_features, num_of_vertices)
self.U_3.shape = (num_of_features, )
self.b_e.shape = (1, num_of_timesteps, num_of_timesteps)
self.V_e.shape = (num_of_timesteps, num_of_timesteps)
for param in [self.U_1, self.U_2, self.U_3, self.b_e, self.V_e]:
param._finish_deferred_init()
# print(x)
# print(self.U_1.data())
# print(self.U_2.data())
# print("=========================")
# print("Context of Variables")
# print("x::{0}, D1::{1}, D2::{2}".format(x.context, self.U_1.data().context, self.U_2.data().context))
# print("=========================")
# print(x)
# compute temporal attention scores
# shape is (N, T, V)
# context.current_context()
# print("temporal context", context.current_context())
cur_context = x.context
lhs = nd.dot(nd.dot(x.transpose((0, 3, 2, 1)), self.U_1.data(cur_context)),
self.U_2.data(cur_context))
# shape is (N, V, T)
rhs = nd.dot(self.U_3.data(cur_context), x.transpose((2, 0, 1, 3)))
product = nd.batch_dot(lhs, rhs)
E = nd.dot(self.V_e.data(cur_context),
nd.sigmoid(product + self.b_e.data(cur_context))
.transpose((1, 2, 0))).transpose((2, 0, 1))
# normailzation
E = E - nd.max(E, axis=1, keepdims=True)
exp = nd.exp(E)
E_normalized = exp / nd.sum(exp, axis=1, keepdims=True)
return E_normalized
def forward(self, cur_input, state, encoder_outputs):
# 当RNN为多层时，取最靠近输出层的单层隐含状态。
# state.shape is [(1, batch_size, decoder_hidden_dim)]
single_layer_state = [state[0][-1].expand_dims(0)]
# encoder_outputs.shape is (max_seq_len, batch_size * encoder_hidden_dim)
encoder_outputs = encoder_outputs.reshape(
(self.max_seq_len, -1, self.encoder_hidden_dim))
# single_layer_state尺寸: [(1, batch_size, decoder_hidden_dim)]
# hidden_broadcast尺寸: (max_seq_len, batch_size, decoder_hidden_dim)
hidden_broadcast = nd.broadcast_axis(single_layer_state[0],
axis=0,
size=self.max_seq_len)
# encoder_outputs_and_hiddens尺寸:
# (max_seq_len, batch_size, encoder_hidden_dim + decoder_hidden_dim)
encoder_outputs_and_hiddens = nd.concat(encoder_outputs,
hidden_broadcast,
dim=2)
# energy尺寸: (max_seq_len, batch_size, 1)
energy = self.attention(encoder_outputs_and_hiddens)
# batch_attention尺寸: (batch_size, 1, max_seq_len)
batch_attention = nd.softmax(energy, axis=0).transpose((1, 2, 0))
# batch_encoder_outputs尺寸: (batch_size, max_seq_len, encoder_hidden_dim)
batch_encoder_outputs = encoder_outputs.swapaxes(0, 1)
# decoder_context尺寸: (batch_size, 1, encoder_hidden_dim)
decoder_context = nd.batch_dot(batch_attention, batch_encoder_outputs)
# cur_input尺寸: (batch_size,)
# input_and_context尺寸: (batch_size, 1, decoder_hidden_dim + encoder_hidden_dim )
input_and_context = nd.concat(nd.expand_dims(self.embedding(cur_input),
axis=1),
decoder_context,
dim=2)
# concat_input尺寸: (1, batch_size, decoder_hidden_dim)
concat_input = self.rnn_concat_input(input_and_context).reshape(
(1, -1, 0))
concat_input = self.dropout(concat_input)
# 当RNN为多层时，用单层隐含状态初始化各个层的隐含状态。
state = [
nd.broadcast_axis(single_layer_state[0],
axis=0,
size=self.num_layers)
]
# XXX 注意：state 是 [nd.NDArray]
output, state = self.rnn(concat_input, state)
output = self.dropout(output)
output = self.out(output)
output = nd.reshape(output, (-3, -1))
# output尺寸: (batch_size * 1, output_dim)
return output, state
def predict(self,x):
h=self.e(x[:, 0])
r=self.r(x[:, 1])
t=self.e(x[:, 2])
t=t.reshape(-1,self.dim,1)
r=r.reshape(-1,self.dim,self.dim)
tr=nd.batch_dot(r,t)
tr=tr.reshape(-1,self.dim)
score = nd.sum(h*tr,-1)
return -score
def forward(self, input_data):
x = nd.transpose(input_data, axes=(1, 0, 2))
h = nd.transpose(self.gru(x), axes=(1, 0, 2)) # (m,60,100)
h = nd.tanh(h)
g = self.att(h) # (m,60,1)
g = nd.softmax(g, axis=1)
gt = nd.transpose(g, axes=(0, 2, 1)) # (m,1,60)
n = nd.batch_dot(gt, h)
y = self.att_out(n)
return self.output(y)
def forward(self, x_left, x_right):
x_left = self.embed_left(x_left)
x_right = self.embed_right(x_right)
embed_cross = nd.expand_dims(
nd.batch_dot(x_left, x_right, transpose_b=True), 3)
embed_cross = nd.transpose(embed_cross, (0, 3, 1, 2))
embed_cross = self._conv_block(embed_cross)
embed_pool = self.pool(embed_cross)
out = self.output_layer(embed_pool)
return out
def forward(self, emb_a, emb_b):
# emb_a: batch_size*seq_len_a*emb_size, emb_b: batch_size*seq_len_b*emb_size
# self.W: emb_size*emb_size
# After the evaluation, the shape is batch_size*seq_len_a*emb_size_b
dot_product = nd.batch_dot(nd.dot(emb_a, self.W.data()), \
nd.transpose(emb_b, axes=(0, 2, 1)))
# this softmax is subject to servere numerical unstability,
# add a work around
G_ab = nd.softmax(dot_product -
nd.max(dot_product, axis=1, keepdims=True),
axis=1)
return G_ab
def compute_curvature(nn_pts):
nn_pts_mean = nd.mean(nn_pts, axis=2, keepdims=True) # (N, P, 1, 3)
nn_pts_demean = nn_pts - nn_pts_mean # (N, P, K, 3)
nn_pts_NPK31 = nd.expand_dims(nn_pts_demean, axis=-1)
covariance_matrix = nd.batch_dot(nn_pts_NPK31,
nn_pts_NPK31,
transpose_b=True) # (N, P, K, 3, 3)
covariance_matrix_mean = nd.mean(covariance_matrix, axis=2,
keepdims=False) # (N, P, 3, 3)
eigvals = compute_eigenvals(covariance_matrix_mean) # (N, P, 3)
curvature = nd.min(eigvals, axis=-1) / (nd.sum(eigvals, axis=-1) + 1e-8)
return curvature
def bdd_message_func(self, edges):
"""Message function for block-diagonal-decomposition regularizer"""
ctx = edges.src['h'].context
if edges.src['h'].dtype in (np.int32, np.int64) and len(edges.src['h'].shape) == 1:
raise TypeError('Block decomposition does not allow integer ID feature.')
weight = self.weight.data(ctx)[edges.data['type'], :].reshape(
-1, self.submat_in, self.submat_out)
node = edges.src['h'].reshape(-1, 1, self.submat_in)
msg = nd.batch_dot(node, weight).reshape(-1, self.out_feat)
if 'norm' in edges.data:
msg = msg * edges.data['norm']
return {'msg': msg}
def make_dynamic_dec(T, values_L):
values_T = nd.array(np.linspace(1, T, num=T), ctx=values_L.context)
values_T = nd.expand_dims(nd.expand_dims(values_T, axis=0), axis=2)
values_T = nd.broadcast_axis(values_T, axis=0, size=values_L.shape[0])
values_TL = nd.batch_dot(values_T, values_L, transpose_b=True)
values_sin = nd.sin(values_TL)
values_cos = nd.cos(values_TL)
return nd.concat(values_sin, values_cos, dim=2)
def forward(self, feature, data):
""" Forward process of a HyperDense layer
Args:
feature: a NDArray with shape [n, d]
data: a NDArray with shape [n, b, pre_d]
Returns:
output: a NDArray with shape [n, b, d]
"""
weight = self.w_mlp(feature) # [n, pre_hidden_size * hidden_size]
weight = nd.reshape(weight, (-1, self.pre_hidden_size, self.hidden_size))
bias = nd.reshape(self.b_mlp(feature), shape=(-1, 1, 1)) # [n, 1, 1]
return nd.batch_dot(data, weight) + bias
def forward(self, x, time, context):
hid = []
hid.append(x)
# m_i = sum A_ij * x_ij + T_A_i
Ain_c = self.A(context)
Ain_t = self.T_A(time)
Ain = Ain_c + Ain_t
# c_i = sum B_ij * u + T_B_i
Bin_c = self.B(context)
Bin_t = self.T_B(time)
Bin = Bin_c + Bin_t
for h in xrange(self.nhop):
hid3dim = hid[-1].expand_dims(1)
Aout = nd.batch_dot(hid3dim, Ain.swapaxes(1,2))
Aout2dim = Aout.reshape((-1, self.mem_size))
P = nd.softmax(Aout2dim, axis=1)
Prob3dim = P.expand_dims(1)
Bout = nd.batch_dot(Prob3dim, Bin)
Bout2dim = Bout.reshape((-1, self.edim))
Cout = self.C(hid[-1])
Dout = Bout2dim + Cout
if self.lindim == self.edim:
hid.append(Dout)
elif self.lindim == 0:
hid.append(nd.relu(Dout))
else:
F = Dout[:, :self.lindim]
G = Dout[:, self.lindim:]
K = nd.relu(G)
hid.append(nd.concat(F, K, dim=1))
z = self.W(hid[-1])
return z
def forward(self, query, key, value, valid_length=None):
"""Forward function"""
d = query.shape[-1]
scores = nd.batch_dot(query, key, transpose_b=True) / math.sqrt(d)
attention_weights = self.dropout(masked_softmax(scores, valid_length))
return nd.batch_dot(attention_weights, value)
