def forward(self, x):
"""
x: N x C x T
"""
if x.dim() != 3:
raise RuntimeError("{} accept 3D tensor as input".format(self.__name__))
# N x 1 x 1
mean = flow.mean(x, (1, 2), keepdim=True)
var = flow.mean((x - mean) ** 2, (1, 2), keepdim=True)
# N x C x T
if self.elementwise_affine:
x = self.gamma * (x - mean) / flow.sqrt(var + self.eps) + self.beta
else:
x = (x - mean) / flow.sqrt(var + self.eps)
return x
def forward(self, inputs, targets):
"""
Args:
inputs (torch.Tensor): feature matrix with shape (batch_size, feat_dim).
targets (torch.LongTensor): ground truth labels with shape (num_classes).
"""
n = inputs.size(0)
# Compute pairwise distance, replace by the official when merged
dist = flow.pow(inputs, 2).sum(dim=1).expand(n, n)
dist = dist + flow.transpose(dist, dim0=1, dim1=0)
temp1 = -2 * flow.matmul(inputs, flow.transpose(inputs, dim0=1,
dim1=0))
dist = flow.add(dist, temp1)
dist = flow.sqrt(flow.clamp(dist, min=1e-12))
# For each anchor, find the hardest positive and negative
mask = targets.expand(n, n).eq(
flow.transpose(targets.expand(n, n), dim0=1, dim1=0))
dist_ap, dist_an = [], []
y1 = flow.zeros((1, n), dtype=flow.float32).to("cuda")
y2 = flow.Tensor(np.exp(100 * np.ones((1, n)))).to("cuda")
for i in range(n):
temp_dist = flow.slice(dist, [(i, i + 1, 1)])
temp_mask = flow.slice(mask, [(i, i + 1, 1)])
temp_mask_rev = flow.slice(1 - mask, [(i, i + 1, 1)])
dist_ap.append(temp_mask.where(temp_dist, y1).max().unsqueeze(0))
dist_an.append(
temp_mask_rev.where(temp_dist, y2).min().unsqueeze(0))
dist_ap = flow.cat(dist_ap)
dist_an = flow.cat(dist_an)
# Compute ranking hinge loss
y = flow.ones_like(dist_an)
return self.ranking_loss(dist_an, dist_ap, y)
def test_sqrt(test_case):
input_arr = np.random.randn(3, 2, 5, 7)
np_out = np.sqrt(input_arr)
x = flow.Tensor(input_arr)
of_out = flow.sqrt(input=x)
test_case.assertTrue(
np.allclose(of_out.numpy(), np_out, 1e-5, 1e-5, equal_nan=True))
def forward(self, x):
if hasattr(self, "qi"):
self.qi.update(x)
x = self.qi.fake_quantize_tensor(x)
if self.training:
y = flow.nn.functional.conv2d(
x,
self.conv_module.weight,
self.conv_module.bias,
stride=self.conv_module.stride,
padding=self.conv_module.padding,
dilation=self.conv_module.dilation,
groups=self.conv_module.groups,
)
y = y.permute(1, 0, 2, 3) # NCHW -> CNHW
y = y.view(self.conv_module.out_channels, -1) # CNHW -> C,NHW
mean = y.mean(1).detach()
var = y.var(1).detach()
self.bn_module.running_mean = (
self.bn_module.momentum * self.bn_module.running_mean +
(1 - self.bn_module.momentum) * mean)
self.bn_module.running_var = (
self.bn_module.momentum * self.bn_module.running_var +
(1 - self.bn_module.momentum) * var)
else:
mean = flow.Tensor(self.bn_module.running_mean)
var = flow.Tensor(self.bn_module.running_var)
std = flow.sqrt(var + self.bn_module.eps)
weight, bias = self.fold_bn(mean, std)
self.qw.update(weight.data)
x = flow.nn.functional.conv2d(
x,
self.qw.fake_quantize_tensor(weight),
bias,
stride=self.conv_module.stride,
padding=self.conv_module.padding,
dilation=self.conv_module.dilation,
groups=self.conv_module.groups,
)
if hasattr(self, "qo"):
self.qo.update(x)
x = self.qo.fake_quantize_tensor(x)
return x
def gradient_penalty(self, y, x):
"""Compute gradient penalty: (L2_norm(dy/dx) - 1)**2."""
weight = flow.ones(y.size()).to(self.device)
dydx = flow.autograd.grad(outputs=y,
inputs=x,
out_grads=weight,
retain_graph=True,
create_graph=True)[0]
dydx = dydx.view(dydx.size(0), -1)
dydx_l2norm = flow.sqrt(flow.sum(dydx**2, dim=1))
return flow.mean((dydx_l2norm - 1)**2)
def _forward(self, x):
axis = 1
params_shape = [x.shape[axis]]
weight = self.weight
bias = self.bias
nd_params_shape = [1] * len(x.shape)
nd_params_shape[axis] = params_shape[0]
mean = x.mean(2, keepdim=True)
variance = x.var(2, unbiased=False, keepdim=True)
normalized = (x - mean) / flow.sqrt(variance + self.eps)
if self.weight is not None and params_shape[0] == self.weight.nelement(
):
weight = flow.reshape(self.weight, shape=nd_params_shape)
if self.bias is not None and params_shape[0] == self.bias.nelement():
bias = flow.reshape(self.bias, shape=nd_params_shape)
if self.weight is not None:
normalized = normalized * weight
if self.bias is not None:
normalized = normalized + bias
return normalized
def forward(self, input: Tensor) -> Tensor:
assert (len(input.shape) >=
3), "The dimensions of input tensor must larger than 2"
assert (input.shape[1] == self.num_channels
), "The channels of input tensor must equal num_channels"
origin_shape = input.shape
reshape_to_1d = flow.reshape(
input, shape=[origin_shape[0], self.num_groups, -1])
mean = flow.mean(reshape_to_1d, dim=2, keepdim=True)
variance = flow.var(reshape_to_1d, dim=2, unbiased=False, keepdim=True)
normalized = (reshape_to_1d - mean) / flow.sqrt(variance + self.eps)
normalized = flow.reshape(
normalized, shape=[origin_shape[0], self.num_channels, -1])
if self.weight is not None:
normalized = normalized * self.weight.reshape(
1, self.num_channels, 1)
if self.bias is not None:
normalized = normalized + self.bias.reshape(
1, self.num_channels, 1)
res = flow.reshape(normalized, shape=tuple(input.shape))
return res
def _sqrt(self):
return flow.sqrt(self)
def forward(self, x):
mean = x.mean(-1, keepdim=True)
std = (x - mean).pow(2).mean(-1, keepdim=True)
x = (x - mean) / flow.sqrt(std + self.eps)
return self.weight * x + self.bias
