def construct(self, x):
square_sum = self.hyper_map(get_square_sum, x)
global_norm = F.sqrt(F.addn(square_sum))
cond = self.greater_equal(global_norm, self.clip_norm)
global_norm = F.select(cond, global_norm, self.clip_norm)
clip_x = self.hyper_map(F.partial(apply_global_norm, self.clip_norm, global_norm), x)
return clip_x
def energy_fn(R: Tensor):
dr = pairwise_displacement(R)
# TODO: plus \epsilon is not accurate, a safe mask is better
dr = F.sqrt(reduce_sum(dr * dr, -1) + 1.1920928955078125e-07)
U = relu(1 - dr)
U = reduce_sum(U * U) * 0.5 * 0.5
return U
def construct(self, x):
mean = self.mean(x, -1)
variance = self.mean(F.square(self.sub(x, mean)))
output = self.div(self.sub(x, mean),
F.sqrt(self.add(variance, self.eps)))
rescaled_output = self.add(self.mul(output, self.gamma), self.beta)
return rescaled_output
def construct(self, grads):
square_sum = self.hyper_map(get_square_sum, grads,
self.allreduce_group_size)
square_reduce_sum = F.addn(square_sum)
stage_square_reduce_sum = self.allreduce(square_reduce_sum)
global_square_reduce_sum = self.allreduce2(stage_square_reduce_sum)
global_norms = F.sqrt(global_square_reduce_sum)
return global_norms
def construct(self, x, X_train):
# Tile input x to match the number of samples in X_train
x_tile = self.tile(x, (128, 1))
square_diff = F.square(x_tile - X_train)
square_dist = self.sum(square_diff, 1)
dist = F.sqrt(square_dist)
# -dist mean the bigger the value is, the nearer the samples are
values, indices = self.topk(-dist, self.k)
return indices
def construct(self, output_hm, output_wh, output_off, output_kps, hm, reg_mask, ind, wh, wight_mask, hm_offset,
hps_mask, landmarks):
"""
Construct method.
"""
hm_loss = self.cls_loss(output_hm, hm) # 1. focal loss, center points
wh_loss = self.reg_loss(output_wh, ind, wh, wight_mask) # 2. weight and height
off_loss = self.reg_loss(output_off, ind, hm_offset, wight_mask) # 3. offset
lm_loss = self.reg_loss_cmask(output_kps, hps_mask, ind, landmarks) # 4. landmark loss
loss = self.hm_weight * hm_loss + self.wh_weight * wh_loss + \
self.off_weight * off_loss + self.lm_weight * lm_loss
# depend is needed when wight_mask and reg_mask is not been used
F.depend(loss, F.sqrt(F.cast(wight_mask, mstype.float32)))
F.depend(loss, F.sqrt(F.cast(reg_mask, mstype.float32)))
# add print when you want to see loss detail and do debug
#self.print('hm_loss=', hm_loss, 'wh_loss=', wh_loss, 'off_loss=', off_loss, 'lm_loss=', lm_loss, 'loss=', loss)
return loss
def construct(self, x1, x2, y):
F.same_type_shape(x1, x2)
_check_reduced_shape_valid(F.shape(x1), F.shape(y), (1, ),
self.cls_name)
# if target > 0, 1-cosine(x1, x2)
# else, max(0, cosine(x1, x2)-margin)
prod_sum = self.reduce_sum(x1 * x2, (1, ))
square1 = self.reduce_sum(F.square(x1), (1, ))
square2 = self.reduce_sum(F.square(x2), (1, ))
denom = F.sqrt(square1) * F.sqrt(square2)
cosine = prod_sum / denom
pos_value = 1.0 - cosine
neg_value = self.maximum(cosine - self.margin, 0.0)
zeros = F.zeros_like(cosine)
pos_part = F.select(y == 1, pos_value, zeros)
neg_part = F.select(y == -1, neg_value, zeros)
output_unreduced = pos_part + neg_part
return self.get_loss(output_unreduced)
def construct(self, x):
x_origin_shape = self.shape(x)
x_target_shape = x_origin_shape[:-1]
x_shape = x_origin_shape + (1, )
x = self.reshape(x, x_shape)
x = self.reshape(x, x_target_shape)
mean = self.mean(x, -1)
variance = self.mean(F.square(self.sub(x, mean)))
output = self.div(self.sub(x, mean),
F.sqrt(self.add(variance, self.eps)))
rescaled_output = self.add(self.mul(output, self.gamma), self.beta)
output_shape = self.shape(rescaled_output) + (1, )
rescaled_output = self.reshape(rescaled_output, output_shape)
return rescaled_output
def construct(self, grads):
square_sum = self.hyper_map(get_square_sum, grads)
global_norms = F.sqrt(
F.addn(square_sum) / F.scalar_to_array(len(square_sum)))
return global_norms
def construct(self, x):
mean = self.mean(x, -1)
variance = self.mean(F.square(x - mean), -1)
output = (x - mean) / F.sqrt(variance + self.eps)
rescaled_output = output * self.gamma + self.beta
return rescaled_output
def construct(self, logits, label):
rmse_loss = F.sqrt(self.MSELoss(logits, label))
return rmse_loss
def construct(self, grads):
square_sum_dp = self.hyper_map(get_square_sum, grads, self.values)
global_norms = F.sqrt(P.AllReduce()(F.addn(square_sum_dp)))
return global_norms
def construct(self, logits, label):
_check_shape(logits.shape, label.shape)
rmse_loss = F.sqrt(self.MSELoss(logits, label))
return rmse_loss
