def HswishGrad(y_grad, x):
"""
HswishGrad
Args:
y_grad:
x:
Returns:
"""
shape = x.shape
res0 = tvm.compute(shape, lambda *i: tvm.if_then_else(x(*i) <= -3, 0, y_grad(*i) * (2 * x(*i) + 3) / 6))
res6 = tvm.compute(shape, lambda *i: tvm.if_then_else(x(*i) >= 3, y_grad(*i), res0(*i)))
return res6
def topi_nn_hsigmoid(x):
"""
topi hsigmoid
Args:
x:
Returns:
"""
return tvm.compute(
x.shape, lambda *i: tvm.if_then_else(
x(*i) <= -3, 0, tvm.if_then_else(x(*i) >= 3, 1, (x(*i) + 3) / 6)))
def ReLU6Grad(y_grad, x):
"""
Computes Gradients of Rectified Linear 6.
Args:
y_grad (tvm.tensor.Tensor): Tensor of type float16, float32, gradients backpropagated to the ReLU6 op.
x (tvm.tensor.Tensor): Tensor of type float16/float32, inputs that where passed to the ReLU6 op, or its outputs.
Returns:
tvm.tensor.Tensor, has same type and shape as x.
"""
shape = x.shape
dtype = x.dtype
zero = tvm.const(0, dtype)
six = tvm.const(6, dtype)
res0 = tvm.compute(shape, lambda *i: tvm.if_then_else(x(*i) >= zero, x(*i), zero))
res6 = tvm.compute(shape, lambda *i: tvm.if_then_else(x(*i) >= six, zero, res0(*i)))
res = tvm.compute(shape, lambda *i: tvm.if_then_else(res6(*i) == zero, zero, y_grad(*i)))
return res
def topi_nn_HSwish(x):
"""
topi HSwish
Args:
x:
Returns:
"""
return tvm.compute(
x.shape, lambda *i: tvm.if_then_else(
x(*i) <= -3, 0,
tvm.if_then_else(x(*i) >= 3, x(*i),
x(*i) * (x(*i) + 3) / 6)))
def HsigmoidGrad(y_grad, x):
"""
HsigmoidGrad
Args:
y_grad:
x:
Returns:
"""
return tvm.compute(
x.shape, lambda *i: tvm.if_then_else(
x(*i) <= -3, 0, tvm.if_then_else(x(*i) >= 3, 0,
y_grad(*i) / 6)))
def topi_nn_relu6(x):
"""topi nn relu6."""
return tvm.compute(
x.shape, lambda *i: tvm.min(tvm.max(x(*i), tvm.const(0, x.dtype)),
tvm.const(6, x.dtype)))