def __init__(self,
in_channels,
out_channels,
weight_init='normal',
bias_init='zeros',
damping=0.03,
loss_scale=1,
frequency=278,
batch_size=32,
has_bias=True,
activation=None):
super(Dense_Thor_GPU, self).__init__()
self.in_channels = Validator.check_positive_int(in_channels)
self.out_channels = Validator.check_positive_int(out_channels)
self.has_bias = Validator.check_bool(has_bias)
self.thor = True
if isinstance(weight_init, Tensor):
if weight_init.ndim != 2 or weight_init.shape[0] != out_channels or \
weight_init.shape[1] != in_channels:
raise ValueError("weight_init shape error")
self.weight = Parameter(initializer(weight_init, [out_channels, in_channels]))
if self.has_bias:
if isinstance(bias_init, Tensor):
if bias_init.ndim != 1 or bias_init.shape[0] != out_channels:
raise ValueError("bias_init shape error")
self.bias = Parameter(initializer(bias_init, [out_channels]))
self.matmul = P.MatMul(transpose_b=True)
self.bias_add = P.BiasAdd()
self.activation = get_activation(activation)
self.activation_flag = self.activation is not None
split_dim = 128
matrix_A_shape, matrix_G_shape = caculate_matmul_shape(self.in_channels, self.out_channels, split_dim)
self.matrix_A_inv = Parameter(Tensor(np.zeros(matrix_A_shape).astype(np.float32)), requires_grad=False)
self.matrix_G_inv = Parameter(Tensor(np.zeros(matrix_G_shape).astype(np.float32)), requires_grad=False)
self.broadcast_to = P.BroadcastTo(matrix_A_shape)
self.cov_step = Parameter(initializer(0, [1], mstype.int32), requires_grad=False)
self.shape = P.Shape()
self.reshape = P.Reshape()
self.transpose = P.Transpose()
self.mul = P.Mul()
self.cube_matmul = P.MatMul(transpose_a=True)
self.loss_scale = Tensor(1 / loss_scale, mstype.float16)
self.batch_size = Tensor(batch_size, mstype.float16)
self.getG = P.InsertGradientOf(self.save_gradient)
self.damping = Parameter(Tensor(damping), requires_grad=False)
self.dampingA = Tensor(np.identity(in_channels), mstype.float32)
self.dampingG = Tensor(np.identity(out_channels), mstype.float32)
self.cast = P.Cast()
self.gather = P.Gather()
self.freq = Tensor(frequency, mstype.int32)
self.axis = 0
self.add = P.Add()
self.sqrt = P.Sqrt()
self.cholesky = P.CholeskyTrsm(split_dim=split_dim)
self.vector_matmul = P.BatchMatMul(transpose_a=True)
def __init__(self):
super(MatInverse, self).__init__()
self.dtype = P.DType()
self.choleskytrsm = P.CholeskyTrsm()
self.matmul = MatMul(transpose_x1=True)
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,
batch_size=32,
bias_init='zeros'):
self.skfac = True
self.hw = kernel_size * kernel_size
kernel_size = twice(kernel_size)
super(Conv2d_SKFAC_GPU, 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.matrix_A_dim = self.in_channels * self.kernel_size[
0] * self.kernel_size[1]
self.matrix_G_dim = self.out_channels
split_dim = 128
self.matrix_A_inv = Parameter(np.zeros(
(self.matrix_A_dim, self.matrix_A_dim)).astype(np.float32),
requires_grad=False)
self.matrix_G_inv = Parameter(np.zeros(
(self.matrix_G_dim, self.matrix_G_dim)).astype(np.float32),
requires_grad=False)
self.cov_step = Parameter(initializer(0, [1], mstype.int32),
requires_grad=False)
self.img2col = P.Im2Col(kernel_size=kernel_size,
stride=stride,
pad_mode="same")
self.matmul = P.MatMul(transpose_a=True)
self.matmul_ = P.MatMul()
self.shape = P.Shape()
self.reshape = P.Reshape()
self.mul = P.Mul()
self.getG = P.InsertGradientOf(self.save_gradient)
self.loss_scale = Tensor(1 / loss_scale, mstype.float16)
self.batch_size = Tensor(batch_size, mstype.float16)
self.transpose = P.Transpose()
self.cast = P.Cast()
self.gather = P.Gather()
self.freq = Tensor(frequency, mstype.int32)
self.axis = 0
self.sqrt = P.Sqrt()
self.reduce_mean = P.ReduceMean(keep_dims=False)
self.damping = Parameter(Tensor(damping), requires_grad=False)
self.dampingA = Tensor(np.identity(batch_size), mstype.float32)
self.dampingG = Tensor(np.identity(batch_size), mstype.float32)
self.I_G = Tensor(np.identity(out_channels), mstype.float32)
self.I_A = Tensor(np.identity(self.matrix_A_dim), mstype.float32)
self.cholesky = P.CholeskyTrsm(split_dim=split_dim)
self.vector_matmul = P.BatchMatMul(transpose_a=True)
self.batch_coefficient = Tensor((1 / 32)**0.5, mstype.float32)
