def get_ops(self):
mul_dtypes = (self.data.dtype, FixedPoint(16, 15))
rshift_dtype = FixedPoint(self.data.dtype.bits + 16,
self.data.dtype.frac_bits + 15)
cmp_dtypes = (self.data.dtype)
return {
Ops.MUL(mul_dtypes): self.data.size,
Ops.RSHIFT(rshift_dtype): self.data.size,
Ops.CMP(cmp_dtypes): self.data.size
}
def get_ops(self):
num = 1
for i in range(len(self.data.shape)):
num *= self.data.shape[i]
add = num
dtypes = (self.data.dtype, self.weights.dtype)
return {Ops.ADD(dtypes): add}
def get_ops(self):
num = 1
for i in range(len(self.data.shape)):
if i != self.dim:
num *= self.data.shape[i]
add = num
dtypes = (self.output_loss.dtype, self.data.dtype)
return {Ops.ADD(dtypes): add}
def get_ops(self):
num = 1
for i in range(len(self.data.shape) - 1):
num *= self.data.shape[i]
cout = self.output_loss[0].shape[-1]
cin = self.data.shape[-1]
mac = cin * \
cout * \
num
dtypes = (self.output_loss[0].dtype, self.data.dtype,
self.output_tensors.dtype)
return {Ops.MAC(dtypes): mac}
def get_ops(self):
num = 1
for i in range(len(self.data.shape) - 1):
num *= self.data.shape[i]
cout = self.output_loss[0].shape[-1]
cin = self.data.shape[-1]
mul = cin * \
cout * \
num
add = num
# return {Ops.MUL: mul, Ops.ADD: add}
dtypes = (self.output_loss[0].dtype, self.data.dtype,
self.output_tensors.dtype)
return {Ops.MAC(dtypes): mul}
def get_ops(self):
num = 1
for i in range(len(self.output_tensors.shape) - 3):
num *= self.data.shape[i]
cin = self.data.shape[-3]
hin = self.data.shape[-2]
win = self.data.shape[-1]
hfil = self.pooling_kernel[-2]
wfil = self.pooling_kernel[-1]
CMP = hfil * wfil * \
hin * win * cin * \
num
dtypes = (self.data.dtype)
return {Ops.CMP(dtypes): CMP}
def get_ops(self):
num = 1
for i in range(len(self.data.shape) - 3):
num *= self.data.shape[i]
cout = self.output_loss[0].shape[-3]
cin = self.data.shape[-3]
hin = self.data.shape[-2]
win = self.data.shape[-1]
hfil = self.weights.shape[-2]
wfil = self.weights.shape[-1]
mac = (wfil * hfil * cout * \
cin * hin * win * \
num)/self.group
dtypes = (self.output_loss[0].dtype, self.weights.dtype,
self.output_tensors.dtype)
return {Ops.MAC(dtypes): mac}
def get_ops(self):
num = 1
for i in range(len(self.data.shape) - 3):
num *= self.data.shape[i]
cout = self.output_loss[0].shape[-3]
cin = self.data.shape[-3]
hout = self.output_loss[0].shape[-2]
wout = self.output_loss[0].shape[-1]
hfil = self.weights.shape[-2]
wfil = self.weights.shape[-1]
mul = (hout * wout * \
cout * cin * hfil * wfil * \
num) / self.group
add = (hout * wout * \
num) / self.group
# return {Ops.MUL: mul, Ops.ADD: add}
dtypes = (self.output_loss[0].dtype, self.data.dtype,
self.output_tensors.dtype)
return {Ops.MAC(dtypes): mul}
def get_ops(self):
ops = self.data.size
sub_dtypes = (self.data.dtype, self.mean.dtype)
mul_dtypes = (self.data.dtype, self.scale.dtype)
return {Ops.SUB(sub_dtypes): ops, Ops.MUL(sub_dtypes): ops}