def forward(self, input):
w = self.weight
a = self.bias
if self.has_sq == True and self.quan == True:
nw = QuantizeNbit.apply(w, self.quan_bit, self.w_scale,
self.w_offset)
na = QuantizeNbit.apply(a, self.quan_bit, self.a_scale,
self.a_offset)
else:
nw = w
na = a
output = F.linear(input, nw, na)
return output
def true_weight(self):
w = self.weight
w_bias = self.bias
nw = 2*(w - (w.max()+w.min())/2)/(w.max()-w.min())
if(w_bias.max()-w_bias.min() == 0):
nw_bias = w_bias
else:
nw_bias = 2*(w_bias - (w_bias.max()+w_bias.min())/2) / \
(w_bias.max()-w_bias.min())
if self.quan == True:
nw = QuantizeNbit.apply(nw, self.quan_bit)
nw_bias = QuantizeNbit.apply(nw_bias, self.quan_bit)
return nw, nw_bias
def forward(self, input):
w = self.weight
w_bias = self.bias
nw = 2*(w - (w.max()+w.min())/2)/(w.max()-w.min())
if(w_bias.max()-w_bias.min() == 0):
nw_bias = w_bias
else:
nw_bias = 2*(w_bias - (w_bias.max()+w_bias.min())/2) / \
(w_bias.max()-w_bias.min())
if self.quan == True:
nw = QuantizeNbit.apply(nw, self.quan_bit)
nw_bias = QuantizeNbit.apply(nw_bias, self.quan_bit)
output = F.linear(input, nw, nw_bias)
return output
def true_weight(self):
w = self.weight
a = self.bias
if self.quan == False:
print('quan=False')
return w, w_bias
if self.has_sq == True and self.quan == True:
nw = QuantizeNbit.apply(
w, self.quan_bit, self.w_scale, self.w_offset)
na = QuantizeNbit.apply(
a, self.quan_bit, self.a_scale, self.a_offset)
return nw, na
else:
print('have not cal the scale and offset, weight is origin weight.')
return w, a
def forward(self, input):
w = self.weight
a = input
if self.has_sq == True and self.quan == True:
nw = QuantizeNbit.apply(w, self.quan_bit, self.scale, self.offset)
else:
nw = w
output = F.conv2d(a, nw, self.bias, self.stride, self.padding,
self.dilation, self.groups)
return output
def true_weight(self):
w = self.weight
#nw = 2*(w - (w.max()+w.min())/2)/(w.max()-w.min())
nw = w - ((w.view(w.size(0), -1).max(-1)[0].view(w.size(0), 1, 1, 1)) +
(w.view(w.size(0), -1).min(-1)[0].view(w.size(0), 1, 1, 1)))/2
nw = 2 * nw
nw = nw / ((w.view(w.size(0), -1).max(-1)[0].view(w.size(0), 1, 1, 1)) -
(w.view(w.size(0), -1).min(-1)[0].view(w.size(0), 1, 1, 1)))
# print(nw)
if self.quan == True:
nw = QuantizeNbit.apply(nw, self.quan_bit)
return nw
def true_weight(self):
w = self.weight
if self.quan == False:
print('quan=False')
return w
if self.has_sq == True and self.quan == True:
nw = QuantizeNbit.apply(w, self.quan_bit, self.scale, self.offset)
return nw
else:
print('have not cal the scale and offset, weight is origin weight.')
return w
def forward(self, input):
w = self.weight
a = input
# 将权重从min,max映射到-1，1
#nw = 2*(w - (w.max()+w.min())/2)/(w.max()-w.min())
nw = w - ((w.view(w.size(0), -1).max(-1)[0].view(w.size(0), 1, 1, 1)) +
(w.view(w.size(0), -1).min(-1)[0].view(w.size(0), 1, 1, 1)))/2
nw = 2 * nw
nw = nw / ((w.view(w.size(0), -1).max(-1)[0].view(w.size(0), 1, 1, 1)) -
(w.view(w.size(0), -1).min(-1)[0].view(w.size(0), 1, 1, 1)))
if self.quan == True:
nw = QuantizeNbit.apply(nw, self.quan_bit)
output = F.conv2d(a, nw, self.bias,
self.stride, self.padding,
self.dilation, self.groups)
return output