def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
pad_mode='same',
padding=0,
dilation=1,
group=1,
data_format='NCHW',
has_bias=False,
weight_init='normal',
damping=0.03,
loss_scale=1,
frequency=278,
bias_init='zeros'):
self.thor = True
ksizes = (1, kernel_size, kernel_size, 1)
self.hw = kernel_size * kernel_size
strides = (1, stride, stride, 1)
kernel_size = twice(kernel_size)
super(Conv2d_Thor, self).__init__(
in_channels,
out_channels,
kernel_size,
stride,
pad_mode,
padding,
dilation,
group,
data_format,
has_bias,
weight_init,
bias_init,
)
self.conv2d = P.Conv2D(out_channel=self.out_channels,
kernel_size=self.kernel_size,
mode=1,
pad_mode=self.pad_mode,
pad=self.padding,
stride=self.stride,
dilation=self.dilation,
group=self.group
)
self.img2col = P.CusImg2Col(ksizes=ksizes, strides=strides)
self.cube_matmul = P.CusMatMulCube(transpose_a=True)
self.matrix_combine = P.CusMatrixCombine()
self.cholesky = P.CusCholeskyTrsm()
self.transpose02314 = P.CusTranspose02314()
self.matrix_A_dim = self.in_channels * self.kernel_size[0] * self.kernel_size[1]
self.matrix_G_dim = self.out_channels
self.matrix_A_device_shape, self.matrix_A_device_dim = caculate_device_shape(self.matrix_A_dim,
self.in_channels, True)
self.matrix_G_device_shape, self.matrix_G_device_dim = caculate_device_shape(self.matrix_G_dim,
self.in_channels, False)
self.matrix_A_device_temp_shape = (
self.matrix_A_device_shape[0], self.matrix_A_device_shape[2], self.matrix_A_device_shape[1],
self.matrix_A_device_shape[3])
self.matrix_G_device_temp_shape = (
self.matrix_G_device_shape[0], self.matrix_G_device_shape[2], self.matrix_G_device_shape[1],
self.matrix_G_device_shape[3])
self.matrix_A_inv = Parameter(
Tensor(np.reshape(np.identity(self.matrix_A_device_dim).astype(np.float16), self.matrix_A_device_shape)),
name='matrix_A_inv', requires_grad=False)
self.A_inv_max = Parameter(initializer(0, [1], mstype.float32), name="A_inv_max", requires_grad=False)
self.matrix_G_inv = Parameter(
Tensor(np.reshape(np.identity(self.matrix_G_device_dim).astype(np.float16), self.matrix_G_device_shape)),
name="matrix_G_inv", requires_grad=False)
self.G_inv_max = Parameter(initializer(0, [1], mstype.float32), name="G_inv_max", requires_grad=False)
self.fake_G = Tensor(
np.reshape(np.identity(self.matrix_G_device_dim).astype(np.float16), self.matrix_G_device_shape))
self.shape = P.Shape()
self.reshape = P.Reshape()
self.transpose = P.Transpose()
self.cov_step = Parameter(initializer(0, [1], mstype.int32), name="cov_step", requires_grad=False)
self.mul = P.Mul()
self.cast = P.Cast()
self.damping = Tensor(damping)
self.vector_matmul = P.CusBatchMatMul()
self.diag_block_dim = 128
self.channels_slice_flag = False
if self.in_channels % C0 != 0:
self.channels_slice_flag = True
self.padA_flag = False
if (self.matrix_A_dim // self.diag_block_dim) * self.diag_block_dim != self.matrix_A_dim \
and self.matrix_A_dim > self.diag_block_dim:
self.padA_flag = True
pad_dim = self.diag_block_dim - self.matrix_A_dim % self.diag_block_dim
self.padA = P.Pad(((0, pad_dim), (0, pad_dim)))
self.device_shape_pad_flag = False
if self.matrix_A_dim != self.matrix_A_device_dim:
self.device_shape_pad_flag = True
self.device_shape_pad = P.Pad(((0, 0), (0, C0 - self.in_channels), (0, 0), (0, C0 - self.in_channels)))
self.slice = P.Slice()
self.gather = P.GatherV2()
self.freq = Tensor(frequency, mstype.int32)
self.loss_scale = Tensor(1 / loss_scale, mstype.float16)
self.axis = 0
dampingA_dim = self.matrix_A_dim
if (self.matrix_A_dim % self.diag_block_dim) != 0 and self.matrix_A_dim > self.diag_block_dim:
dampingA_dim = (self.matrix_A_dim // self.diag_block_dim + 1) * self.diag_block_dim
dampingG_dim = self.matrix_G_dim
if (self.matrix_G_dim % self.diag_block_dim) != 0 and self.matrix_G_dim > self.diag_block_dim:
dampingG_dim = (self.matrix_G_dim // self.diag_block_dim + 1) * self.diag_block_dim
self.dampingA = Tensor(np.identity(dampingA_dim), mstype.float32)
self.dampingG = Tensor(np.identity(dampingG_dim), mstype.float32)
self.fused_abs_max1 = P.CusFusedAbsMax1([self.matrix_A_dim, self.matrix_A_dim])
self.fused_abs_max2 = P.CusFusedAbsMax1()
self.log = P.Log()
self.exp = P.Exp()
self.sqrt = P.Sqrt()
self.getG = P.InsertGradientOf(self.save_gradient)
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
pad_mode='same',
padding=0,
dilation=1,
group=1,
has_bias=False,
weight_init='normal',
bias_init='zeros'):
kernel_size = twice(kernel_size)
stride = twice(stride)
self._dilation = dilation
dilation = twice(dilation)
super(Conv2d_Thor,
self).__init__(in_channels, out_channels, kernel_size, stride,
pad_mode, padding, dilation, group, has_bias,
weight_init, bias_init)
self.conv2d = P.Conv2D(out_channel=self.out_channels,
kernel_size=self.kernel_size,
mode=1,
pad_mode=self.pad_mode,
pad=self.padding,
stride=self.stride,
dilation=self.dilation,
group=self.group)
self._init_depthwise_conv2d(weight_init)
self.bias_add = P.BiasAdd()
self.thor = True
self.hw = kernel_size[0] * kernel_size[1]
self.matrix_A_dim = self.in_channels * self.kernel_size[
0] * self.kernel_size[1]
self.matrix_G_dim = self.out_channels
self.shape = P.Shape()
self.reshape = P.Reshape()
self.mul = P.Mul()
self.cast = P.Cast()
self.A_normalizer = Parameter(initializer(0, [1], mstype.float32),
name="A_normalizer",
requires_grad=False)
self.G_normalizer = Parameter(initializer(0, [1], mstype.float32),
name="G_normalizer",
requires_grad=False)
self.is_Ascend = True
if context.get_context("device_target") == "Ascend":
ksizes = (1, kernel_size[0], kernel_size[1], 1)
strides = (1, stride[0], stride[1], 1)
self.img2col = P.CusImg2Col(ksizes=ksizes, strides=strides)
self.cube_matmul = P.CusMatMulCube(transpose_a=True)
self.transpose02314 = P.CusTranspose02314()
dampingA_dim = self.matrix_A_dim
self.diag_block_dim = 128
if (self.matrix_A_dim % self.diag_block_dim
) != 0 and self.matrix_A_dim > self.diag_block_dim:
dampingA_dim = (self.matrix_A_dim // self.diag_block_dim +
1) * self.diag_block_dim
dampingG_dim = self.matrix_G_dim
if (self.matrix_G_dim % self.diag_block_dim
) != 0 and self.matrix_G_dim > self.diag_block_dim:
dampingG_dim = (self.matrix_G_dim // self.diag_block_dim +
1) * self.diag_block_dim
self.matrix_A_cov = Parameter(Tensor(
np.zeros([dampingA_dim, dampingA_dim]).astype(np.float32)),
name='matrix_A',
requires_grad=False)
self.matrix_G_cov = Parameter(Tensor(
np.zeros([dampingG_dim, dampingG_dim]).astype(np.float32)),
name='matrix_G',
requires_grad=False)
self.channels_slice_flag = False
self.C0 = 16
if self.in_channels % self.C0 != 0:
self.channels_slice_flag = True
self.padA_flag = False
if (self.matrix_A_dim // self.diag_block_dim) * self.diag_block_dim != self.matrix_A_dim \
and self.matrix_A_dim > self.diag_block_dim:
self.padA_flag = True
pad_dim = self.diag_block_dim - self.matrix_A_dim % self.diag_block_dim
self.padA = P.Pad(((0, pad_dim), (0, pad_dim)))
self.slice = P.Slice()
else:
self.is_Ascend = False
self.img2col = P.Im2Col(kernel_size=kernel_size,
stride=stride,
pad_mode="same")
self.matmul = P.MatMul(transpose_b=True)
self.reduce_mean = P.ReduceMean(keep_dims=False)
self.matrix_A_cov = Parameter(Tensor(
np.zeros([self.matrix_A_dim,
self.matrix_A_dim]).astype(np.float32)),
name='matrix_A',
requires_grad=False)
self.matrix_G_cov = Parameter(Tensor(
np.zeros([self.matrix_G_dim,
self.matrix_G_dim]).astype(np.float32)),
name='matrix_G',
requires_grad=False)
self.getG = P.InsertGradientOf(self.save_gradient)