def construct(self, s_t_hat, encoder_outputs, encoder_feature,
enc_padding_mask, coverage):
b, t_k, n = encoder_outputs.shape
dec_fea = self.decode_proj(s_t_hat) # (B, 2 * hidden_dim)
dec_fea_expand = P.ExpandDims()(dec_fea, 1)
dec_fea_expand = P.BroadcastTo()(dec_fea_expand, (b, t_k, n))
att_features = encoder_feature + dec_fea_expand
if self.is_coverage:
coverage_input = coverage.view(-1, 1) # (B * t_k, 1)
coverage_feature = self.W_c(
coverage_input) # (B * t_k, 2 * hidden_dim)
att_features = att_features + coverage_feature
e = P.Tanh()(att_features) # (B * t_k, 2 * hidden_dim)
scores = self.v(e) # (B * t_k, 1)
scores = scores.view(-1, t_k) # (B, t_k)
attn_dist_ = P.Softmax(1)(scores) * enc_padding_mask # (B, t_k)
normalization_factor = P.ReduceSum(True)(attn_dist_, 1)
attn_dist = attn_dist_ / normalization_factor
attn_dist = P.ExpandDims()(attn_dist, 1) # (B, 1, t_k)
c_t = P.BatchMatMul(attn_dist, encoder_outputs) # (B, 1, n)
c_t = c_t.view(-1, self.hidden_dim * 2) # (B, 2 * hidden_dim)
attn_dist = attn_dist.view(-1, t_k)
if self.is_coverage:
coverage = coverage.view(-1, t_k)
coverage = coverage + attn_dist
return c_t, attn_dist, coverage
def variable_recurrent(self, x, h, seq_length):
time_step = range(x.shape[0])
h_t = h
if self.is_lstm:
hidden_size = h[0].shape[-1]
zero_output = P.ZerosLike()(h_t[0])
else:
hidden_size = h.shape[-1]
zero_output = P.ZerosLike()(h_t)
seq_length = P.BroadcastTo((hidden_size, -1))(seq_length)
seq_length = P.Transpose()(seq_length, (1, 0))
outputs = []
state_t = h_t
for t in time_step:
h_t = self.cell(x[t], state_t)
seq_cond = seq_length > t
if self.is_lstm:
state_t_0 = P.Select()(seq_cond, h_t[0], state_t[0])
state_t_1 = P.Select()(seq_cond, h_t[1], state_t[1])
output = P.Select()(seq_cond, h_t[0], zero_output)
state_t = (state_t_0, state_t_1)
else:
state_t = P.Select()(seq_cond, h_t, state_t)
output = P.Select()(seq_cond, h_t, zero_output)
outputs.append(output)
outputs = P.Stack()(outputs)
return outputs, state_t
def variable_recurrent(self, x, h, seq_length, w_ih, w_hh, b_ih, b_hh):
'''recurrent steps with sequence length'''
time_step = x.shape[0]
h_t = h
if self.is_lstm:
hidden_size = h[0].shape[-1]
zero_output = P.ZerosLike()(h_t[0])
else:
hidden_size = h.shape[-1]
zero_output = P.ZerosLike()(h_t)
seq_length = P.Cast()(seq_length, mindspore.float32)
seq_length = P.BroadcastTo((hidden_size, -1))(seq_length)
seq_length = P.Cast()(seq_length, mindspore.int32)
seq_length = P.Transpose()(seq_length, (1, 0))
outputs = []
state_t = h_t
t = 0
while t < time_step:
x_t = x[t:t + 1:1]
x_t = P.Squeeze(0)(x_t)
h_t = self.cell(x_t, state_t, w_ih, w_hh, b_ih, b_hh)
seq_cond = seq_length > t
if self.is_lstm:
state_t_0 = P.Select()(seq_cond, h_t[0], state_t[0])
state_t_1 = P.Select()(seq_cond, h_t[1], state_t[1])
output = P.Select()(seq_cond, h_t[0], zero_output)
state_t = (state_t_0, state_t_1)
else:
state_t = P.Select()(seq_cond, h_t, state_t)
output = P.Select()(seq_cond, h_t, zero_output)
outputs.append(output)
t += 1
outputs = P.Stack()(outputs)
return outputs, state_t
def construct(self, inputs, targets):
"""
Args:
- inputs: feature matrix with shape (batch_size, feat_dim)
- targets: ground truth labels with shape (num_classes)
"""
n = inputs.shape[0]
# Compute pairwise distance, replace by the official when merged
pow = P.Pow()
sum = P.ReduceSum(keep_dims=True)
expand = P.BroadcastTo((n, n))
transpose = P.Transpose()
mul = P.Mul()
add = P.Add()
sqrt = P.Sqrt()
equal = P.Equal()
cat = P.Concat()
ones_like = P.OnesLike()
dist = pow(inputs, 2)
dist = sum(dist, axis=1)
dist = expand(dist)
dist = dist + transpose(dist, (1, 0))
temp1 = P.matmul(inputs, transpose(inputs, (1, 0)))
temp1 = mul(-2, temp1)
dist = add(dist, temp1)
dist = P.composite.clip_by_value(
dist, clip_value_min=1e-12, clip_value_max=100000000
) # for numerical stability, clip_value_max=? why must set?
dist = sqrt(dist)
# For each anchor, find the hardest positive and negative
targets = expand(targets)
mask = equal(targets, transpose(targets, (1, 0)))
dist_ap = []
dist_an = []
# only for debugging
#####################
# print("dist is")
# print(dist.shape)
# print(dist)
# print("mask is")
# print(mask.shape)
# print(mask)
# print(mask[0])
#####################
for i in range(n):
minval = -1.0
maxval = -1.0
for j in range(n):
if mask[i][j] and dist[i][j] > maxval:
maxval = dist[i][j]
if not mask[i][j] and (dist[i][j] < minval or minval == -1):
minval = dist[i][j]
if (not isinstance(minval, Tensor)
or not isinstance(maxval, Tensor) or minval == -1.0
or maxval == -1.0):
if self.error_msg is not None:
print("Error Msg", file=self.error_msg)
print("mask {} is".format(i), file=self.error_msg)
print(mask[i], file=self.error_msg)
print("dist is:", file=self.error_msg)
print(dist[i], file=self.error_msg)
print(maxval, file=self.error_msg)
print(minval, file=self.error_msg)
print(type(maxval), file=self.error_msg)
print(type(minval), file=self.error_msg)
self.error_msg.flush()
# assert minval != -1.0 and isinstance(minval, Tensor)
# assert maxval != -1.0 and isinstance(maxval, Tensor)
dist_ap.append(maxval.asnumpy())
dist_an.append(minval.asnumpy())
dist_ap = Tensor(dist_ap, ms.float32)
dist_an = Tensor(dist_an, ms.float32)
# only for debugging
#####################
# print(dist_ap)
# print(dist_ap.shape)
# print(dist_an)
#####################
# Compute ranking hinge loss
y = ones_like(dist_an)
loss = self.ranking_loss(dist_an, dist_ap, y)
# # compute accuracy
# correct = torch.ge(dist_an, dist_ap).sum().item()
return loss
# class GradOriTripletLoss(nn.Cell)
# def __init__(self, net):
# super(GradOriTripletLoss, self).__init__()
# self.net = net
# self.grad_op = P.GradOperation(get_all=True)
#
# def construct(self, inputs, targets):
# gradient_function = self.grad_op(self.net)
# return gradient_function(inputs, targets)