def forward(self, input):
weight1 = self.weight * self.scale + (self.weight - self.weight * self.scale).detach()
weight = weight1 + (wage_quantizer.Q(weight1,self.wl_weight) -weight1).detach()
output= F.conv2d(input, weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
output = output/self.scale
output = wage_quantizer.WAGEQuantizer_f(output, self.wl_activate, self.wl_error)
return output
def forward(self, input):
weight1 = self.weight * self.scale + (self.weight - self.weight * self.scale).detach()
weight = weight1 + (wage_quantizer.Q(weight1,self.wl_weight) -weight1).detach()
output = F.linear(input, weight, self.bias)
output = output/self.scale
output = wage_quantizer.WAGEQuantizer_f(output,self.wl_activate, self.wl_error)
return output
def Neural_Sim(self, input, output):
input_file_name = './layer_record/input' + str(self.name) + '.csv'
weight_file_name = './layer_record/weight' + str(self.name) + '.csv'
f = open('./layer_record/trace_command.sh', "a")
f.write(weight_file_name + ' ' + input_file_name + ' ')
weight_q = wage_quantizer.Q(self.weight, self.wl_weight)
write_matrix_weight(weight_q.cpu().data.numpy(), weight_file_name)
if len(self.weight.shape) > 2:
k = self.weight.shape[-1]
write_matrix_activation_conv(
stretch_input(input[0].cpu().data.numpy(), k), None, self.wl_input,
input_file_name)
else:
write_matrix_activation_fc(input[0].cpu().data.numpy(), None,
self.wl_input, input_file_name)
def Neural_Sim(self, input, output):
input_file_name = './layer_record/input' + str(self.name) + '.csv'
weight_file_name = './layer_record/weight' + str(self.name) + '.csv'
weightOld_file_name = './layer_record/weightOld' + str(self.name) + '.csv'
f = open('./layer_record/trace_command.sh', "a")
input_activity = open('./input_activity.csv', "a")
weight_q = wage_quantizer.Q(self.weight,self.wl_weight)
write_matrix_weight( weight_q.cpu().data.numpy(),weight_file_name)
if len(self.weight.shape) > 2:
k=self.weight.shape[-1]
padding = self.padding
stride = self.stride
activity = write_matrix_activation_conv(stretch_input(input[0].cpu().data.numpy(),k,padding,stride),None,self.wl_input,input_file_name)
input_activity.write(str(activity)+",")
else:
activity = write_matrix_activation_fc(input[0].cpu().data.numpy(),None ,self.wl_input, input_file_name)
if (str(self.name) == 'FC2_'):
input_activity.write(str(activity)+"\n")
else:
input_activity.write(str(activity)+",")
f.write(weight_file_name+' '+weightOld_file_name+' '+input_file_name+' '+str(activity)+' ')
def Neural_Sim(self, input, output):
global model_n, FP
print("quantize layer ", self.name)
input_file_name = './layer_record_' + str(model_n) + '/input' + str(
self.name) + '.csv'
weight_file_name = './layer_record_' + str(model_n) + '/weight' + str(
self.name) + '.csv'
f = open('./layer_record_' + str(model_n) + '/trace_command.sh', "a")
f.write(weight_file_name + ' ' + input_file_name + ' ')
if FP:
weight_q = float_quantizer.float_range_quantize(
self.weight, self.wl_weight)
else:
weight_q = wage_quantizer.Q(self.weight, self.wl_weight)
write_matrix_weight(weight_q.cpu().data.numpy(), weight_file_name)
if len(self.weight.shape) > 2:
k = self.weight.shape[-1]
write_matrix_activation_conv(
stretch_input(input[0].cpu().data.numpy(), k), None, self.wl_input,
input_file_name)
else:
write_matrix_activation_fc(input[0].cpu().data.numpy(), None,
self.wl_input, input_file_name)
def forward(self, input):
weight1 = self.weight * self.scale + (
self.weight - self.weight * self.scale).detach()
weight = weight1 + (wage_quantizer.Q(weight1, self.wl_weight) -
weight1).detach()
outputOrignal = F.conv2d(input, weight, self.bias, self.stride,
self.padding, self.dilation, self.groups)
bitWeight = int(self.wl_weight)
bitActivation = int(self.wl_input)
if self.inference == 1:
# retention
weight = wage_quantizer.Retention(weight, self.t, self.v,
self.detect, self.target)
# set parameters for Hardware Inference
onoffratio = self.onoffratio
upper = 1
lower = 1 / onoffratio
output = torch.zeros_like(outputOrignal)
del outputOrignal
cellRange = 2**self.cellBit # cell precision is 4
# Now consider on/off ratio
dummyP = torch.zeros_like(weight)
dummyP[:, :, :, :] = (cellRange - 1) * (upper + lower) / 2
for i in range(3):
for j in range(3):
# need to divide to different subArray
numSubArray = int(weight.shape[1] / self.subArray)
# cut into different subArrays
if numSubArray == 0:
mask = torch.zeros_like(weight)
mask[:, :, i, j] = 1
if weight.shape[1] == 3:
# after get the spacial kernel, need to transfer floating weight [-1, 1] to binarized ones
X_decimal = torch.round((2**bitWeight - 1) / 2 *
(weight + 1) + 0) * mask
outputP = torch.zeros_like(output)
outputD = torch.zeros_like(output)
for k in range(int(bitWeight / self.cellBit)):
remainder = torch.fmod(X_decimal,
cellRange) * mask
variation = np.random.normal(
0, self.vari,
list(weight.size())).astype(np.float32)
X_decimal = torch.round(
(X_decimal - remainder) / cellRange) * mask
# Now also consider weight has on/off ratio effects
# Here remainder is the weight mapped to Hardware, so we introduce on/off ratio in this value
# the range of remainder is [0, cellRange-1], we truncate it to [lower, upper]
remainderQ = (upper - lower) * (
remainder - 0
) + (
cellRange - 1
) * lower # weight cannot map to 0, but to Gmin
remainderQ = remainderQ + remainderQ * torch.from_numpy(
variation).cuda()
outputPartial = F.conv2d(
input, remainderQ * mask, self.bias,
self.stride, self.padding, self.dilation,
self.groups)
outputDummyPartial = F.conv2d(
input, dummyP * mask, self.bias,
self.stride, self.padding, self.dilation,
self.groups)
scaler = cellRange**k
outputP = outputP + outputPartial * scaler * 2 / (
1 - 1 / onoffratio)
outputD = outputD + outputDummyPartial * scaler * 2 / (
1 - 1 / onoffratio)
outputP = outputP - outputD
output = output + outputP
else:
# quantize input into binary sequence
inputQ = torch.round((2**bitActivation - 1) / 1 *
(input - 0) + 0)
outputIN = torch.zeros_like(output)
for z in range(bitActivation):
inputB = torch.fmod(inputQ, 2)
inputQ = torch.round((inputQ - inputB) / 2)
outputP = torch.zeros_like(output)
# after get the spacial kernel, need to transfer floating weight [-1, 1] to binarized ones
X_decimal = torch.round(
(2**bitWeight - 1) / 2 *
(weight + 1) + 0) * mask
outputD = torch.zeros_like(output)
for k in range(int(bitWeight / self.cellBit)):
remainder = torch.fmod(
X_decimal, cellRange) * mask
variation = np.random.normal(
0, self.vari,
list(weight.size())).astype(np.float32)
X_decimal = torch.round(
(X_decimal - remainder) /
cellRange) * mask
# Now also consider weight has on/off ratio effects
# Here remainder is the weight mapped to Hardware, so we introduce on/off ratio in this value
# the range of remainder is [0, cellRange-1], we truncate it to [lower, upper]
remainderQ = (upper - lower) * (
remainder - 0
) + (
cellRange - 1
) * lower # weight cannot map to 0, but to Gmin
remainderQ = remainderQ + remainderQ * torch.from_numpy(
variation).cuda()
outputPartial = F.conv2d(
input, remainderQ * mask, self.bias,
self.stride, self.padding,
self.dilation, self.groups)
outputDummyPartial = F.conv2d(
input, dummyP * mask, self.bias,
self.stride, self.padding,
self.dilation, self.groups)
# Add ADC quanization effects here !!!
outputPartialQ = wage_quantizer.LinearQuantizeOut(
outputPartial, self.ADCprecision)
outputDummyPartialQ = wage_quantizer.LinearQuantizeOut(
outputDummyPartial, self.ADCprecision)
scaler = cellRange**k
outputP = outputP + outputPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
outputD = outputD + outputDummyPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
scalerIN = 2**z
outputIN = outputIN + (outputP -
outputD) * scalerIN
output = output + outputIN / (2**bitActivation)
else:
# quantize input into binary sequence
inputQ = torch.round((2**bitActivation - 1) / 1 *
(input - 0) + 0)
outputIN = torch.zeros_like(output)
for z in range(bitActivation):
inputB = torch.fmod(inputQ, 2)
inputQ = torch.round((inputQ - inputB) / 2)
outputP = torch.zeros_like(output)
for s in range(numSubArray):
mask = torch.zeros_like(weight)
mask[:, (s * self.subArray):(s + 1) *
self.subArray, i, j] = 1
# after get the spacial kernel, need to transfer floating weight [-1, 1] to binarized ones
X_decimal = torch.round(
(2**bitWeight - 1) / 2 *
(weight + 1) + 0) * mask
outputSP = torch.zeros_like(output)
outputD = torch.zeros_like(output)
for k in range(int(bitWeight / self.cellBit)):
remainder = torch.fmod(
X_decimal, cellRange) * mask
variation = np.random.normal(
0, self.vari,
list(weight.size())).astype(np.float32)
X_decimal = torch.round(
(X_decimal - remainder) /
cellRange) * mask
# Now also consider weight has on/off ratio effects
# Here remainder is the weight mapped to Hardware, so we introduce on/off ratio in this value
# the range of remainder is [0, cellRange-1], we truncate it to [lower, upper]*(cellRange-1)
remainderQ = (upper - lower) * (
remainder - 0
) + (
cellRange - 1
) * lower # weight cannot map to 0, but to Gmin
remainderQ = remainderQ + remainderQ * torch.from_numpy(
variation).cuda()
outputPartial = F.conv2d(
inputB, remainderQ * mask, self.bias,
self.stride, self.padding,
self.dilation, self.groups)
outputDummyPartial = F.conv2d(
inputB, dummyP * mask, self.bias,
self.stride, self.padding,
self.dilation, self.groups)
# Add ADC quanization effects here !!!
outputPartialQ = wage_quantizer.LinearQuantizeOut(
outputPartial, self.ADCprecision)
outputDummyPartialQ = wage_quantizer.LinearQuantizeOut(
outputDummyPartial, self.ADCprecision)
scaler = cellRange**k
outputSP = outputSP + outputPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
outputD = outputD + outputDummyPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
if (weight.shape[0]
== 256) & (weight.shape[1] == 128):
weightMatrix = (
remainderQ *
mask).cpu().data.numpy()
weight_file_name = './layer_record/weightForLayer3_subarray' + str(
s) + '_weightBitNo_' + str(
k) + ".csv"
cout = weightMatrix.shape[0]
weight_matrix = weightMatrix.reshape(
cout, -1).transpose()
np.savetxt(weight_file_name,
weight_matrix,
delimiter=",",
fmt='%10.5f')
# !!! Important !!! the dummy need to be multiplied by a ratio
outputSP = outputSP - outputD # minus dummy column
outputP = outputP + outputSP
scalerIN = 2**z
outputIN = outputIN + outputP * scalerIN
output = output + outputIN / (2**bitActivation)
output = output / (
2**bitWeight
) # since weight range was convert from [-1, 1] to [-256, 256]
else:
# original WAGE QCov2d
weight1 = self.weight * self.scale + (
self.weight - self.weight * self.scale).detach()
weight = weight1 + (wage_quantizer.Q(weight1, self.wl_weight) -
weight1).detach()
output = F.conv2d(input, weight, self.bias, self.stride,
self.padding, self.dilation, self.groups)
output = output / self.scale
output = wage_quantizer.WAGEQuantizer_f(output, self.wl_activate,
self.wl_error)
return output
def forward(self, input):
weight1 = self.weight * self.scale + (
self.weight - self.weight * self.scale).detach()
weight = weight1 + (wage_quantizer.Q(weight1, self.wl_weight) -
weight1).detach()
outputOrignal = F.linear(input, weight, self.bias)
output = torch.zeros_like(outputOrignal)
bitWeight = int(self.wl_weight)
bitActivation = int(self.wl_input)
if self.inference == 1:
# retention
weight = wage_quantizer.Retention(weight, self.t, self.v,
self.detect, self.target)
# set parameters for Hardware Inference
onoffratio = self.onoffratio
upper = 1
lower = 1 / onoffratio
output = torch.zeros_like(outputOrignal)
cellRange = 2**self.cellBit # cell precision is 4
# Now consider on/off ratio
dummyP = torch.zeros_like(weight)
dummyP[:, :] = (cellRange - 1) * (upper + lower) / 2
# need to divide to different subArray
numSubArray = int(weight.shape[1] / self.subArray)
if numSubArray == 0:
mask = torch.zeros_like(weight)
mask[:, :] = 1
# quantize input into binary sequence
inputQ = torch.round((2**bitActivation - 1) / 1 * (input - 0) +
0)
outputIN = torch.zeros_like(outputOrignal)
for z in range(bitActivation):
inputB = torch.fmod(inputQ, 2)
inputQ = torch.round((inputQ - inputB) / 2)
# after get the spacial kernel, need to transfer floating weight [-1, 1] to binarized ones
X_decimal = torch.round((2**bitWeight - 1) / 2 *
(weight + 1) + 0) * mask
outputP = torch.zeros_like(outputOrignal)
outputD = torch.zeros_like(outputOrignal)
for k in range(int(bitWeight / self.cellBit)):
remainder = torch.fmod(X_decimal, cellRange) * mask
variation = np.random.normal(
0, self.vari,
list(weight.size())).astype(np.float32)
X_decimal = torch.round(
(X_decimal - remainder) / cellRange) * mask
# Now also consider weight has on/off ratio effects
# Here remainder is the weight mapped to Hardware, so we introduce on/off ratio in this value
# the range of remainder is [0, cellRange-1], we truncate it to [lower, upper]
remainderQ = (upper - lower) * (remainder - 0) + (
cellRange -
1) * lower # weight cannot map to 0, but to Gmin
remainderQ = remainderQ + remainderQ * torch.from_numpy(
variation).cuda()
outputPartial = F.linear(input, remainderQ * mask,
self.bias)
outputDummyPartial = F.linear(input, dummyP * mask,
self.bias)
# Add ADC quanization effects here !!!
outputPartialQ = wage_quantizer.LinearQuantizeOut(
outputPartial, self.ADCprecision)
outputDummyPartialQ = wage_quantizer.LinearQuantizeOut(
outputDummyPartial, self.ADCprecision)
scaler = cellRange**k
outputP = outputP + outputPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
outputD = outputD + outputDummyPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
scalerIN = 2**z
outputIN = outputIN + (outputP - outputD) * scalerIN
output = output + outputIN / (2**bitActivation)
else:
inputQ = torch.round((2**bitActivation - 1) / 1 * (input - 0) +
0)
outputIN = torch.zeros_like(outputOrignal)
for z in range(bitActivation):
inputB = torch.fmod(inputQ, 2)
inputQ = torch.round((inputQ - inputB) / 2)
outputP = torch.zeros_like(outputOrignal)
for s in range(numSubArray):
mask = torch.zeros_like(weight)
mask[:,
(s * self.subArray):(s + 1) * self.subArray] = 1
# after get the spacial kernel, need to transfer floating weight [-1, 1] to binarized ones
X_decimal = torch.round((2**bitWeight - 1) / 2 *
(weight + 1) + 0) * mask
outputSP = torch.zeros_like(outputOrignal)
outputD = torch.zeros_like(outputOrignal)
for k in range(int(bitWeight / self.cellBit)):
remainder = torch.fmod(X_decimal, cellRange) * mask
variation = np.random.normal(
0, self.vari,
list(remainder.size())).astype(np.float32)
X_decimal = torch.round(
(X_decimal - remainder) / cellRange) * mask
# Now also consider weight has on/off ratio effects
# Here remainder is the weight mapped to Hardware, so we introduce on/off ratio in this value
# the range of remainder is [0, cellRange-1], we truncate it to [lower, upper]*(cellRange-1)
remainderQ = (upper - lower) * (remainder - 0) + (
cellRange - 1
) * lower # weight cannot map to 0, but to Gmin
remainderQ = remainderQ + remainderQ * torch.from_numpy(
variation).cuda()
outputPartial = F.linear(input, remainderQ * mask,
self.bias)
outputDummyPartial = F.linear(
input, dummyP * mask, self.bias)
# Add ADC quanization effects here !!!
outputPartialQ = wage_quantizer.LinearQuantizeOut(
outputPartial, self.ADCprecision)
outputDummyPartialQ = wage_quantizer.LinearQuantizeOut(
outputDummyPartial, self.ADCprecision)
scaler = cellRange**k
outputSP = outputSP + outputPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
outputD = outputD + outputDummyPartialQ * scaler * 2 / (
1 - 1 / onoffratio)
outputSP = outputSP - outputD # minus dummy column
outputP = outputP + outputSP
scalerIN = 2**z
outputIN = outputIN + outputP * scalerIN
output = output + outputIN / (2**bitActivation)
output = output / (2**bitWeight)
else:
# original WAGE QCov2d
weight1 = self.weight * self.scale + (
self.weight - self.weight * self.scale).detach()
weight = weight1 + (wage_quantizer.Q(weight1, self.wl_weight) -
weight1).detach()
output = F.linear(input, weight, self.bias)
output = output / self.scale
output = wage_quantizer.WAGEQuantizer_f(output, self.wl_activate,
self.wl_error)
return output
def pre_save_old_weight(oldWeight, name, wl_weight):
if not os.path.exists('./layer_record'):
os.makedirs('./layer_record')
weight_file_name = './layer_record/Oldweight' + str(name) + '.csv'
weight_q = wage_quantizer.Q(oldWeight, wl_weight)
write_matrix_weight(weight_q.cpu().data.numpy(), weight_file_name)
