def h_swish(x):
"""Hard Swish: MobileNetV3 https://arxiv.org/pdf/1905.02244.pdf
"""
return x * F.relu6(x + 3) / 6
def forward(self, x):
return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0
def hard_sigmoid(x, inplace: bool = False):
if inplace:
return x.add_(3.).clamp_(0., 6.).div_(6.)
else:
return F.relu6(x + 3.) / 6.
def hard_sigmoid(x, inplace=False):
return F.relu6(x + 3, inplace) / 6
def hard_swish(input, inplace=False):
return input * F.relu6(input + 3.).div(6.)
def forward(self, x):
out = F.relu6(x + 3., self.inplace) / 6.
return out * x
def forward(self, input):
# 训练态
if self.training:
self.step += 1
if self.bn:
# 先做普通卷积得到A，以取得BN参数
output = F.conv2d(input=input,
weight=self.weight,
bias=self.bias,
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
# 更新BN统计参数（batch和running）
dims = [dim for dim in range(4) if dim != 1]
self.batch_mean = torch.mean(output, dim=dims)
self.batch_var = torch.var(output, dim=dims)
with torch.no_grad():
if self.first_bn == 0 and torch.equal(
self.running_mean,
torch.zeros_like(
self.running_mean)) and torch.equal(
self.running_var,
torch.zeros_like(self.running_var)):
self.first_bn.add_(1)
self.running_mean.add_(self.batch_mean)
self.running_var.add_(self.batch_var)
else:
self.running_mean.mul_(1 - self.momentum).add_(
self.momentum * self.batch_mean)
self.running_var.mul_(1 - self.momentum).add_(
self.momentum * self.batch_var)
# BN融合
if self.step < self.freeze_step:
if self.bias is not None:
bias = reshape_to_bias(
self.beta + (self.bias - self.batch_mean) *
(self.gamma /
torch.sqrt(self.batch_var + self.eps)))
else:
bias = reshape_to_bias(
self.beta - self.batch_mean *
(self.gamma /
torch.sqrt(self.batch_var + self.eps))) # b融batch
weight = self.weight * reshape_to_weight(
self.gamma /
torch.sqrt(self.batch_var + self.eps)) # w融running
else:
if self.bias is not None:
bias = reshape_to_bias(
self.beta + (self.bias - self.running_mean) *
(self.gamma /
torch.sqrt(self.running_var + self.eps)))
else:
bias = reshape_to_bias(
self.beta - self.running_mean *
(self.gamma / torch.sqrt(self.running_var +
self.eps))) # b融batch
weight = self.weight * reshape_to_weight(
self.gamma /
torch.sqrt(self.running_var + self.eps)) # w融running
else:
bias = self.bias
weight = self.weight
# 测试态
else:
# print(self.running_mean, self.running_var)
# BN融合
if self.bn:
if self.bias is not None:
bias = reshape_to_bias(
self.beta + (self.bias - self.running_mean) *
(self.gamma / torch.sqrt(self.running_var + self.eps)))
else:
bias = reshape_to_bias(
self.beta - self.running_mean *
(self.gamma / torch.sqrt(self.running_var + self.eps))
) # b融running
weight = self.weight * reshape_to_weight(
self.gamma /
torch.sqrt(self.running_var + self.eps)) # w融running
else:
bias = self.bias
weight = self.weight
# 量化A和bn融合后的W
q_weight = self.weight_quantizer(weight)
q_bias = self.bias_quantizer(bias)
if self.quantizer_output == True: # 输出量化参数txt文档
# 创建的quantizer_output输出文件夹
if not os.path.isdir('./quantier_output'):
os.makedirs('./quantier_output')
if not os.path.isdir('./quantier_output/q_weight_out'):
os.makedirs('./quantier_output/q_weight_out')
if not os.path.isdir('./quantier_output/w_scale_out'):
os.makedirs('./quantier_output/w_scale_out')
if not os.path.isdir('./quantier_output/q_weight_max'):
os.makedirs('./quantier_output/q_weight_max')
if not os.path.isdir('./quantier_output/max_weight_count'):
os.makedirs('./quantier_output/max_weight_count')
#######################输出当前层的权重量化因子
weight_scale = self.weight_quantizer.get_scale()
np.savetxt(
('./quantier_output/w_scale_out/scale %f.txt' % time.time()),
weight_scale,
delimiter='\n')
#######################输出当前层的量化权重
q_weight_txt = self.weight_quantizer.get_quantize_value(weight)
q_weight_txt = np.array(q_weight_txt.cpu()).reshape(1, -1)
q_weight_max = [np.max(q_weight_txt)]
# q_weight_max = np.argmax(q_weight_txt)
max_weight_count = [np.sum(abs(q_weight_txt) >= 255)] # 统计该层溢出的数目
np.savetxt(
('./quantier_output/max_weight_count/max_weight_count %f.txt' %
time.time()), max_weight_count)
np.savetxt(('./quantier_output/q_weight_max/max_weight %f.txt' %
time.time()), q_weight_max)
np.savetxt(
('./quantier_output/q_weight_out/weight %f.txt' % time.time()),
q_weight_txt,
delimiter='\n')
# io.savemat('save.mat',{'q_weight_txt':q_weight_txt})
#######################创建输出偏置txt的文件夹
if not os.path.isdir('./quantier_output/q_bias_out'):
os.makedirs('./quantier_output/q_bias_out')
if not os.path.isdir('./quantier_output/b_scale_out'):
os.makedirs('./quantier_output/b_scale_out')
#######################输出当前层偏置的量化因子
bias_scale = self.bias_quantizer.get_scale()
np.savetxt(
('./quantier_output/b_scale_out/scale %f.txt' % time.time()),
bias_scale,
delimiter='\n')
#######################输出当前层的量化偏置
q_bias_txt = self.bias_quantizer.get_quantize_value(bias)
q_bias_txt = np.array(q_bias_txt.cpu()).reshape(1, -1)
np.savetxt(
('./quantier_output/q_bias_out/bias %f.txt' % time.time()),
q_bias_txt,
delimiter='\n')
# 量化卷积
if self.training: # 训练态
output = F.conv2d(
input=input,
weight=q_weight,
# bias=self.bias, # 注意，这里不加bias（self.bias为None）
bias=q_bias,
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
else: # 测试态
output = F.conv2d(
input=input,
weight=q_weight,
bias=q_bias, # 注意，这里加bias，做完整的conv+bn
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
if self.activate == 'leaky':
output = F.leaky_relu(output, 0.125, inplace=True)
elif self.activate == 'relu6':
output = F.relu6(output, inplace=True)
elif self.activate == 'h_swish':
output = output * (F.relu6(output + 3.0, inplace=True) / 6.0)
elif self.activate == 'relu':
output = F.relu(output, inplace=True)
elif self.activate == 'mish':
output = output * F.softplus(output).tanh()
elif self.activate == 'linear':
return output
# pass
else:
print(self.activate + " is not supported !")
if self.quantizer_output == True:
if not os.path.isdir('./quantier_output/q_activation_out'):
os.makedirs('./quantier_output/q_activation_out')
if not os.path.isdir('./quantier_output/a_scale_out'):
os.makedirs('./quantier_output/a_scale_out')
if not os.path.isdir('./quantier_output/q_activation_max'):
os.makedirs('./quantier_output/q_activation_max')
if not os.path.isdir('./quantier_output/max_activation_count'):
os.makedirs('./quantier_output/max_activation_count')
##################输出当前激活的量化因子
activation_scale = self.activation_quantizer.get_scale()
np.savetxt(
('./quantier_output/a_scale_out/scale %f.txt' % time.time()),
activation_scale,
delimiter='\n')
##################输出当前层的量化激活
q_activation_txt = self.activation_quantizer.get_quantize_value(
output)
q_activation_txt = np.array(q_activation_txt.cpu()).reshape(1, -1)
q_activation_max = [np.max(q_activation_txt)] # 统计该层的最大值(即查看是否有溢出)
max_activation_count = [np.sum(abs(q_activation_txt) >= 255)
] # 统计该层溢出的数目
# q_weight_max = np.argmax(q_weight_txt)
np.savetxt((
'./quantier_output/max_activation_count/max_activation_count %f.txt'
% time.time()), max_activation_count)
np.savetxt(
('./quantier_output/q_activation_max/max_activation %f.txt' %
time.time()), q_activation_max)
np.savetxt(
('./quantier_output/q_activation_out/activation %f.txt' %
time.time()),
q_activation_txt,
delimiter='\n')
output = self.activation_quantizer(output)
return output
def forward(self, x):
pi1 = F.relu6(self.pi1(x))
logits = self.pi2(pi1)
v1 = F.relu6(self.v1(x))
values = self.v2(v1)
return logits, values
def relu6_forward_unquant(self, input):
return F.relu6(input)
def forward(self, input):
tic = time.time()
output = F.relu6(input, self.inplace)
toc = time.time()
self.run_time = toc - tic
return output
def relu6_forward(self, input):
return qunsigned(F.relu6(input), self.acti_log2_t, self.acti_bit_width)
def forward(self, x):
if self.inplace:
return x.add_(3.).clamp_(0., 6.).div_(6.)
else:
return F.relu6(x + 3.) / 6.
def hard_swish(x, inplace=False):
if inplace:
return x.mul_(F.relu6(x + 3.) / 6.)
else:
return x * F.relu6(x + 3.) / 6.
def forward(self, x):
out = F.relu6(self.bn1(self.conv1(x)))
out = F.relu6(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out = out + self.shortcut(x) if self.stride == 1 else out
return out
def forward(self, x):
return F.relu6(self.bn(self.conv(x)))
def forward(self, input):
# 训练态
if self.training:
if self.bn:
# 先做普通卷积得到A，以取得BN参数
output = F.conv2d(input=input,
weight=self.weight,
bias=self.bias,
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
# 更新BN统计参数（batch和running）
dims = [dim for dim in range(4) if dim != 1]
self.batch_mean = torch.mean(output, dim=dims)
self.batch_var = torch.var(output, dim=dims)
with torch.no_grad():
if self.first_bn == 0 and torch.equal(
self.running_mean,
torch.zeros_like(
self.running_mean)) and torch.equal(
self.running_var,
torch.zeros_like(self.running_var)):
self.first_bn.add_(1)
self.running_mean.add_(self.batch_mean)
self.running_var.add_(self.batch_var)
else:
self.running_mean.mul_(1 - self.momentum).add_(
self.batch_mean * self.momentum)
self.running_var.mul_(1 - self.momentum).add_(
self.batch_var * self.momentum)
# BN融合
if self.bias is not None:
bias = reshape_to_bias(
self.beta + (self.bias - self.batch_mean) *
(self.gamma / torch.sqrt(self.batch_var + self.eps)))
else:
bias = reshape_to_bias(
self.beta - self.batch_mean *
(self.gamma / torch.sqrt(self.batch_var + self.eps))
) # b融batch
weight = self.weight * reshape_to_weight(
self.gamma /
torch.sqrt(self.batch_var + self.eps)) # w融running
# if self.bias is not None:
# bias = reshape_to_bias(
# self.beta + (self.bias - self.running_mean) * (
# self.gamma / torch.sqrt(self.running_var + self.eps)))
# else:
# bias = reshape_to_bias(
# self.beta - self.running_mean * (
# self.gamma / torch.sqrt(self.running_var + self.eps))) # b融batch
# weight = self.weight * reshape_to_weight(
# self.gamma / torch.sqrt(self.running_var + self.eps)) # w融running
else:
bias = self.bias
weight = self.weight
# 测试态
else:
# print(self.running_mean, self.running_var)
if self.bn:
# BN融合
if self.bias is not None:
bias = reshape_to_bias(
self.beta + (self.bias - self.running_mean) *
(self.gamma / torch.sqrt(self.running_var + self.eps)))
else:
bias = reshape_to_bias(
self.beta - self.running_mean * self.gamma /
torch.sqrt(self.running_var + self.eps)) # b融running
weight = self.weight * reshape_to_weight(
self.gamma /
torch.sqrt(self.running_var + self.eps)) # w融running
else:
bias = self.bias
weight = self.weight
# 量化A和bn融合后的W
q_weight = self.weight_quantizer(weight)
q_bias = self.bias_quantizer(bias)
# 量化卷积
if self.training: # 训练态
output = F.conv2d(input=input,
weight=q_weight,
bias=q_bias,
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
# （这里将训练态下，卷积中w融合running参数的效果转为融合batch参数的效果）running ——> batch
# if self.bn:
# output *= reshape_to_activation(
# torch.sqrt(self.running_var + self.eps) / torch.sqrt(self.batch_var + self.eps))
# output += reshape_to_activation(
# self.gamma * (self.running_mean / (self.running_var + self.eps) - self.batch_mean / (
# self.batch_var + self.eps)))
# output += reshape_to_activation(bias)
else: # 测试态
output = F.conv2d(
input=input,
weight=q_weight,
bias=q_bias, # 注意，这里加bias，做完整的conv+bn
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
if self.activate == 'leaky':
output = F.leaky_relu(output, 0.125, inplace=True)
elif self.activate == 'relu6':
output = F.relu6(output, inplace=True)
elif self.activate == 'h_swish':
output = output * (F.relu6(output + 3.0, inplace=True) / 6.0)
elif self.activate == 'relu':
output = F.relu(output, inplace=True)
elif self.activate == 'mish':
output = output * F.softplus(output).tanh()
elif self.activate == 'linear':
# return output
pass
else:
print(self.activate + "%s is not supported !")
output = self.activation_quantizer(output)
return output
def forward(self, x):
out = x * (F.relu6(x + 3) / 6)
return out
def hard_sigmoid(x, inplace=True):
relu = F.relu6(x + 3.0)
return (1.0 / 6) * relu
def forward(self, x):
out = F.relu6(x + 3) / 6
return out
def hard_sigmoid(input, inplace=False):
return F.relu6(input + 3).div(6.)
def prune_model_keep_size(model, prune_idx, CBL_idx, CBLidx2mask):
pruned_model = deepcopy(model)
activations = []
for i, model_def in enumerate(model.module_defs):
if model_def['type'] == 'convolutional' or model_def[
'type'] == 'depthwise' or model_def['type'] == 'se':
activation = torch.zeros(int(model_def['filters'])).cuda()
if i in prune_idx:
mask = torch.from_numpy(CBLidx2mask[i]).cuda()
bn_module = pruned_model.module_list[i][1]
bn_module.weight.data.mul_(mask)
if hasattr(pruned_model.module_list[i], 'activation'):
ac_module = pruned_model.module_list[i][2]
if ac_module.__class__.__name__ == "LeakyReLU":
activation = F.leaky_relu(
(1 - mask) * bn_module.bias.data, 0.1)
elif ac_module.__class__.__name__ == "ReLU6":
activation = F.relu6((1 - mask) * bn_module.bias.data,
inplace=True)
elif ac_module.__class__.__name__ == "HardSwish":
x = (1 - mask) * bn_module.bias.data
activation = x * (F.relu6(x + 3.0, inplace=True) / 6.0)
elif ac_module.__class__.__name__ == "ReLU":
activation = F.relu((1 - mask) * bn_module.bias.data,
0.1)
elif ac_module.__class__.__name__ == "Mish":
x = (1 - mask) * bn_module.bias.data
activation = x * F.softplus(x).tanh()
else:
activation = (1 - mask) * bn_module.bias.data
else:
activation = (1 - mask) * bn_module.bias.data
update_activation(i, pruned_model, activation, CBL_idx)
bn_module.bias.data.mul_(mask)
activations.append(activation)
elif model_def['type'] == 'shortcut':
actv1 = activations[i - 1]
from_layer = int(model_def['from'][0])
actv2 = activations[i + from_layer]
activation = actv1 + actv2
update_activation(i, pruned_model, activation, CBL_idx)
activations.append(activation)
elif model_def['type'] == 'route':
# spp不参与剪枝，其中的route不用更新，仅占位
from_layers = [int(s) for s in model_def['layers']]
activation = None
if len(from_layers) == 1:
activation = activations[
i +
from_layers[0] if from_layers[0] < 0 else from_layers[0]]
if 'groups' in model_def:
activation = activation[(activation.shape[0] // 2):]
update_activation(i, pruned_model, activation, CBL_idx)
elif len(from_layers) == 2:
actv1 = activations[i + from_layers[0]]
actv2 = activations[
i +
from_layers[1] if from_layers[1] < 0 else from_layers[1]]
activation = torch.cat((actv1, actv2))
update_activation(i, pruned_model, activation, CBL_idx)
activations.append(activation)
elif model_def['type'] == 'upsample':
# activation = torch.zeros(int(model.module_defs[i - 1]['filters'])).cuda()
activations.append(activations[i - 1])
elif model_def['type'] == 'yolo':
activations.append(None)
elif model_def['type'] == 'maxpool': # 区分spp和tiny
if model.module_defs[i + 1]['type'] == 'route':
activations.append(None)
else:
activation = activations[i - 1]
update_activation(i, pruned_model, activation, CBL_idx)
activations.append(activation)
return pruned_model
def hard_swish(self, x, inplace):
inner = F.relu6(x + 3.).div_(6.)
return x.mul_(inner) if inplace else x.mul(inner)
def hard_swish(x):
return x * F.relu6(x + 3.) / 6.
def backward_hook(module, grad_in, grad_out):
# Cut off negative gradients
if isinstance(module, nn.ReLU):
return (F.relu6(grad_in[0]), )
def hard_sigmoid(x):
return F.relu6(x + 3.) / 6.
def forward(self, x):
out = F.relu6(x + 3, inplace=True) / 6
return out
def forward(self, x):
return F.relu6(x + 3, inplace=self.inplace) / 6
def forward(self, x):
return F.relu6(x + 3.0, inplace=True) / 6.0
def forward(self, x):
pl_1 = F.relu6(self.policy_layer_1(x))
policy = F.softmax(self.policy_layer_2(pl_1), dim=1)
vl_1 = F.relu6(self.value_layer_1(x))
value = self.value_layer_2(vl_1)
return policy, value
def forward(self, input):
if self.bn:
# BN融合
if self.bias is not None:
bias = reshape_to_bias(
self.beta + (self.bias - self.running_mean) *
(self.gamma / torch.sqrt(self.running_var + self.eps)))
else:
bias = reshape_to_bias(
self.beta - self.running_mean * self.gamma /
torch.sqrt(self.running_var + self.eps)) # b融running
weight = self.weight * reshape_to_weight(
self.gamma /
torch.sqrt(self.running_var + self.eps)) # w融running
else:
bias = self.bias
weight = self.weight
# 量化A和bn融合后的W
q_weight = self.weight_quantizer(weight)
q_bias = self.bias_quantizer(bias)
if self.quantizer_output == True: # 输出量化参数txt文档
# 创建的quantizer_output输出文件夹
if not os.path.isdir('./quantizer_output'):
os.makedirs('./quantizer_output')
if not os.path.isdir('./quantizer_output/q_weight_out'):
os.makedirs('./quantizer_output/q_weight_out')
if not os.path.isdir('./quantizer_output/w_scale_out'):
os.makedirs('./quantizer_output/w_scale_out')
if not os.path.isdir('./quantizer_output/q_weight_max'):
os.makedirs('./quantizer_output/q_weight_max')
if not os.path.isdir('./quantizer_output/max_weight_count'):
os.makedirs('./quantizer_output/max_weight_count')
if not os.path.isdir('./quantizer_output/q_weight_reorder'):
os.makedirs('./quantizer_output/q_weight_reorder')
if not os.path.isdir('./quantizer_output/q_bias_reorder'):
os.makedirs('./quantizer_output/q_bias_reorder')
if self.layer_idx == -1:
#######################输出当前层的权重量化因子
weight_scale = -self.weight_quantizer.get_scale()
np.savetxt(
('./quantizer_output/w_scale_out/w_scale_%s' % self.name),
weight_scale,
delimiter='\n')
#######################输出当前层的量化权重
q_weight_txt = self.weight_quantizer.get_quantize_value(weight)
#############权重重排序
w_para = q_weight_txt # 重排序参数
if self.reorder == True:
# print("use weights reorder!")
shape_output = w_para.shape[0]
shape_input = w_para.shape[1]
num_TN = int(shape_input / self.TN)
remainder_TN = shape_input % self.TN
num_TM = int(shape_output / self.TM)
remainder_TM = shape_output % self.TM
first = True
reorder_w_para = None
if self.activate == 'linear':
print('layer-linear reorder!')
for k in range(num_TN):
temp = w_para[0:remainder_TM,
k * self.TN:(k + 1) * self.TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1],
temp.shape[2] * temp.shape[3])
temp = temp.permute(2, 0, 1).contiguous().view(-1)
if first:
reorder_w_para = temp.clone().cpu().data.numpy(
)
first = False
else:
reorder_w_para = np.append(
reorder_w_para,
temp.cpu().data.numpy())
else:
for j in range(num_TM):
if shape_input == 3 or shape_input == 1: # 第一层
print('The first layer~~~~~~~~~~~~')
temp = w_para[j * self.TM:(j + 1) * self.TM,
num_TN *
self.TN:num_TN * self.TN +
remainder_TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1],
temp.shape[2] * temp.shape[3])
fill = torch.zeros(self.TM, self.TN,
temp.shape[2]).to(
temp.device)
fill[:, 0:remainder_TN, :] = temp
temp = fill.permute(2, 0,
1).contiguous().view(-1)
if first: # 创建数组存储
reorder_w_para = temp.clone().cpu(
).data.numpy()
first = False
else:
reorder_w_para = np.append(
reorder_w_para,
temp.cpu().data.numpy())
else:
for k in range(num_TN):
temp = w_para[j * self.TM:(j + 1) *
self.TM,
k * self.TN:(k + 1) *
self.TN, :, :]
# #合并成论文图10(a)的TM*TN*(K2)的张量格式
temp = temp.view(
temp.shape[0], temp.shape[1],
temp.shape[2] * temp.shape[3])
# 转换为图10(b)的重排序格式
temp = temp.permute(
2, 0, 1).contiguous().view(-1)
if first:
reorder_w_para = temp.clone().cpu(
).data.numpy()
first = False
else:
reorder_w_para = np.append(
reorder_w_para,
temp.cpu().data.numpy())
w_para_flatten = reorder_w_para
# print(reorder_w_para.size)
#####验证重排序结果的正确性
'''if w_para_flatten.size == w_para.shape[0] * w_para.shape[1] * w_para.shape[2] * w_para.shape[3]:
print("weights convert correctly!")
else:
print("weights convert mismatchingly!")'''
q_weight_reorder = w_para_flatten
q_weight_reorder = np.array(q_weight_reorder).reshape(
1, -1)
np.savetxt(
('./quantizer_output/q_weight_reorder/w_reorder_%s.txt'
% self.name),
q_weight_reorder,
delimiter='\n')
################权重重排序结束
q_weight_txt = np.array(q_weight_txt.cpu()).reshape(1, -1)
q_weight_max = [np.max(q_weight_txt)]
# q_weight_max = np.argmax(q_weight_txt)
max_weight_count = [
np.sum(abs(q_weight_txt) >= (1 << (self.w_bits - 1)) - 1)
] # 统计该层溢出的数目
np.savetxt(
('./quantizer_output/max_weight_count/max_w_count_%s.txt' %
self.name), max_weight_count)
np.savetxt(('./quantizer_output/q_weight_max/max_w_%s.txt' %
self.name), q_weight_max)
np.savetxt(('./quantizer_output/q_weight_out/q_weight_%s.txt' %
self.name),
q_weight_txt,
delimiter='\n')
# io.savemat('save.mat',{'q_weight_txt':q_weight_txt})
#######################创建输出偏置txt的文件夹
if not os.path.isdir('./quantizer_output/q_bias_out'):
os.makedirs('./quantizer_output/q_bias_out')
if not os.path.isdir('./quantizer_output/b_scale_out'):
os.makedirs('./quantizer_output/b_scale_out')
#######################输出当前层偏置的量化因子
bias_scale = -self.bias_quantizer.get_scale()
np.savetxt(('./quantizer_output/b_scale_out/b_scale_%s.txt' %
self.name),
bias_scale,
delimiter='\n')
#######################输出当前层的量化偏置
q_bias_txt = self.bias_quantizer.get_quantize_value(bias)
q_bias_txt = np.array(q_bias_txt.cpu()).reshape(1, -1)
np.savetxt(('./quantizer_output/q_bias_out/q_bias_%s.txt' %
self.name),
q_bias_txt,
delimiter='\n')
#############偏置重排序
if self.reorder == True:
b_para = np.zeros(2048, dtype=int)
b_para[0:q_bias_txt.size] = q_bias_txt
# print(b_para.shape)
# b_para = np.array(b_para.cpu()).reshape(1, -1)
np.savetxt(
('./quantizer_output/q_bias_reorder/q_b_reorder_%s.txt'
% self.name),
b_para,
delimiter='\n')
######权重和偏置的重排序数据的二进制文件保存
bias_weight_reorder = np.append(b_para, q_weight_reorder)
wb_flat = bias_weight_reorder.astype(np.int8)
writer = open(
'./quantizer_output/q_weight_reorder/%s_bias_weight_q_bin'
% self.name, "wb")
writer.write(wb_flat)
writer.close()
################偏置重排序结束
elif int(self.name[1:4]) == self.layer_idx:
#######################输出当前层的权重量化因子
weight_scale = -self.weight_quantizer.get_scale()
np.savetxt(
('./quantizer_output/w_scale_out/w_scale_%s' % self.name),
weight_scale,
delimiter='\n')
#######################输出当前层的量化权重
q_weight_txt = self.weight_quantizer.get_quantize_value(weight)
#############权重重排序
w_para = q_weight_txt # 重排序参数
if self.reorder == True:
# print("use weights reorder!")
shape_output = w_para.shape[0]
shape_input = w_para.shape[1]
num_TN = int(shape_input / self.TN)
remainder_TN = shape_input % self.TN
num_TM = int(shape_output / self.TM)
remainder_TM = shape_output % self.TM
first = True
reorder_w_para = None
if self.activate == 'linear':
print('layer-linear reorder!')
for k in range(num_TN):
temp = w_para[0:remainder_TM,
k * self.TN:(k + 1) * self.TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1],
temp.shape[2] * temp.shape[3])
temp = temp.permute(2, 0, 1).contiguous().view(-1)
if first:
reorder_w_para = temp.clone().cpu().data.numpy(
)
first = False
else:
reorder_w_para = np.append(
reorder_w_para,
temp.cpu().data.numpy())
else:
for j in range(num_TM):
if shape_input == 3 or shape_input == 1: # 第一层
print('The first layer~~~~~~~~~~~~')
temp = w_para[j * self.TM:(j + 1) * self.TM,
num_TN *
self.TN:num_TN * self.TN +
remainder_TN, :, :]
temp = temp.view(temp.shape[0], temp.shape[1],
temp.shape[2] * temp.shape[3])
fill = torch.zeros(self.TM, self.TN,
temp.shape[2]).to(
temp.device)
fill[:, 0:remainder_TN, :] = temp
temp = fill.permute(2, 0,
1).contiguous().view(-1)
if first: # 创建数组存储
reorder_w_para = temp.clone().cpu(
).data.numpy()
first = False
else:
reorder_w_para = np.append(
reorder_w_para,
temp.cpu().data.numpy())
else:
for k in range(num_TN):
temp = w_para[j * self.TM:(j + 1) *
self.TM,
k * self.TN:(k + 1) *
self.TN, :, :]
# #合并成论文图10(a)的TM*TN*(K2)的张量格式
temp = temp.view(
temp.shape[0], temp.shape[1],
temp.shape[2] * temp.shape[3])
# 转换为图10(b)的重排序格式
temp = temp.permute(
2, 0, 1).contiguous().view(-1)
if first:
reorder_w_para = temp.clone().cpu(
).data.numpy()
first = False
else:
reorder_w_para = np.append(
reorder_w_para,
temp.cpu().data.numpy())
w_para_flatten = reorder_w_para
# print(reorder_w_para.size)
#####验证重排序结果的正确性
'''if w_para_flatten.size == w_para.shape[0] * w_para.shape[1] * w_para.shape[2] * w_para.shape[3]:
print("weights convert correctly!")
else:
print("weights convert mismatchingly!")'''
q_weight_reorder = w_para_flatten
q_weight_reorder = np.array(q_weight_reorder).reshape(
1, -1)
np.savetxt(
('./quantizer_output/q_weight_reorder/w_reorder_%s.txt'
% self.name),
q_weight_reorder,
delimiter='\n')
################权重重排序结束
q_weight_txt = np.array(q_weight_txt.cpu()).reshape(1, -1)
q_weight_max = [np.max(q_weight_txt)]
# q_weight_max = np.argmax(q_weight_txt)
max_weight_count = [
np.sum(abs(q_weight_txt) >= (1 << (self.w_bits - 1)) - 1)
] # 统计该层溢出的数目
np.savetxt(
('./quantizer_output/max_weight_count/max_w_count_%s.txt' %
self.name), max_weight_count)
np.savetxt(('./quantizer_output/q_weight_max/max_w_%s.txt' %
self.name), q_weight_max)
np.savetxt(('./quantizer_output/q_weight_out/q_weight_%s.txt' %
self.name),
q_weight_txt,
delimiter='\n')
# io.savemat('save.mat',{'q_weight_txt':q_weight_txt})
#######################创建输出偏置txt的文件夹
if not os.path.isdir('./quantizer_output/q_bias_out'):
os.makedirs('./quantizer_output/q_bias_out')
if not os.path.isdir('./quantizer_output/b_scale_out'):
os.makedirs('./quantizer_output/b_scale_out')
#######################输出当前层偏置的量化因子
bias_scale = -self.bias_quantizer.get_scale()
np.savetxt(('./quantizer_output/b_scale_out/b_scale_%s.txt' %
self.name),
bias_scale,
delimiter='\n')
#######################输出当前层的量化偏置
q_bias_txt = self.bias_quantizer.get_quantize_value(bias)
q_bias_txt = np.array(q_bias_txt.cpu()).reshape(1, -1)
np.savetxt(('./quantizer_output/q_bias_out/q_bias_%s.txt' %
self.name),
q_bias_txt,
delimiter='\n')
#############偏置重排序
if self.reorder == True:
b_para = np.zeros(2048, dtype=int)
b_para[0:q_bias_txt.size] = q_bias_txt
# print(b_para.shape)
# b_para = np.array(b_para.cpu()).reshape(1, -1)
np.savetxt(
('./quantizer_output/q_bias_reorder/q_b_reorder_%s.txt'
% self.name),
b_para,
delimiter='\n')
######权重和偏置的重排序数据的二进制文件保存
bias_weight_reorder = np.append(b_para, q_weight_reorder)
wb_flat = bias_weight_reorder.astype(np.int8)
writer = open(
'./quantizer_output/q_weight_reorder/%s_bias_weight_q_bin'
% self.name, "wb")
writer.write(wb_flat)
writer.close()
################偏置重排序结束
# 量化卷积
output = F.conv2d(
input=input,
weight=q_weight,
bias=q_bias, # 注意，这里加bias，做完整的conv+bn
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups)
if self.activate == 'leaky':
output = F.leaky_relu(output,
0.125 if not self.maxabsscaler else 0.25,
inplace=True)
elif self.activate == 'relu6':
output = F.relu6(output, inplace=True)
elif self.activate == 'h_swish':
output = output * (F.relu6(output + 3.0, inplace=True) / 6.0)
elif self.activate == 'relu':
output = F.relu(output, inplace=True)
elif self.activate == 'mish':
output = output * F.softplus(output).tanh()
elif self.activate == 'linear':
# return output
pass
else:
print(self.activate + "%s is not supported !")
if self.quantizer_output == True:
if not os.path.isdir('./quantizer_output/q_activation_out'):
os.makedirs('./quantizer_output/q_activation_out')
if not os.path.isdir('./quantizer_output/a_scale_out'):
os.makedirs('./quantizer_output/a_scale_out')
if not os.path.isdir('./quantizer_output/q_activation_max'):
os.makedirs('./quantizer_output/q_activation_max')
if not os.path.isdir('./quantizer_output/max_activation_count'):
os.makedirs('./quantizer_output/max_activation_count')
if not os.path.isdir('./quantizer_output/q_activation_reorder'):
os.makedirs('./quantizer_output/q_activation_reorder')
if self.layer_idx == -1:
##################输出当前激活的量化因子
activation_scale = -self.activation_quantizer.get_scale()
np.savetxt(('./quantizer_output/a_scale_out/a_scale_%s.txt' %
self.name),
activation_scale,
delimiter='\n')
##################输出当前层的量化激活
q_activation_txt = self.activation_quantizer.get_quantize_value(
output)
a_para = q_activation_txt
#############输入特征图重排序
if self.reorder == True:
# 重排序参数
# print("use activation reorder!")
shape_input = a_para.shape[1]
num_TN = int(shape_input / self.TN)
remainder_TN = shape_input % self.TN
first = True
reorder_a_para = None
if self.activate == 'linear':
print('layer-linear reorder!')
temp = a_para[:, 0:remainder_TN, :, :]
temp = temp.view(temp.shape[1], temp.shape[2],
temp.shape[3])
temp = temp.permute(2, 1, 0).contiguous().view(-1)
if first:
reorder_a_para = temp.clone().cpu().data.numpy()
first = False
else:
reorder_a_para = np.append(reorder_a_para,
temp.cpu().data.numpy())
else:
for k in range(num_TN):
temp = a_para[:,
k * self.TN:(k + 1) * self.TN, :, :]
temp = temp.view(temp.shape[1], temp.shape[2],
temp.shape[3])
temp = temp.permute(2, 1, 0).contiguous().view(-1)
if first:
reorder_a_para = temp.clone().cpu().data.numpy(
)
first = False
else:
reorder_a_para = np.append(
reorder_a_para,
temp.cpu().data.numpy())
a_para_flatten = reorder_a_para
#####验证重排序结果的正确性
'''if a_para_flatten.size == a_para.shape[0] * a_para.shape[1] * a_para.shape[2] * a_para.shape[3]:
print("activation convert correctly!")
else:
print("activation convert mismatchingly!")'''
q_activation_reorder = a_para_flatten
q_activation_reorder = np.array(
q_activation_reorder).reshape(1, -1)
np.savetxt((
'./quantizer_output/q_activation_reorder/a_reorder_%s.txt'
% self.name),
q_activation_reorder,
delimiter='\n')
###保存重排序的二进制文件
activation_flat = q_activation_reorder.astype(np.int8)
writer = open(
'./quantizer_output/q_activation_reorder/%s_activation_q_bin'
% self.name, "wb")
writer.write(activation_flat)
writer.close()
##########特征图重排序结束
q_activation_txt = np.array(q_activation_txt.cpu()).reshape(
1, -1)
q_activation_max = [np.max(q_activation_txt)
] # 统计该层的最大值(即查看是否有溢出)
max_activation_count = [
np.sum(
abs(q_activation_txt) >= (1 << (self.w_bits - 1)) - 1)
] # 统计该层溢出的数目
# q_weight_max = np.argmax(q_weight_txt)
np.savetxt((
'./quantizer_output/max_activation_count/max_a_count_%s.txt'
% self.name), max_activation_count)
np.savetxt(
('./quantizer_output/q_activation_max/q_a_max_%s.txt' %
self.name), q_activation_max)
np.savetxt(
('./quantizer_output/q_activation_out/q_activation_%s.txt'
% self.name),
q_activation_txt,
delimiter='\n')
elif int(self.name[1:4]) == self.layer_idx:
##################输出当前激活的量化因子
activation_scale = -self.activation_quantizer.get_scale()
np.savetxt(('./quantizer_output/a_scale_out/a_scale_%s.txt' %
self.name),
activation_scale,
delimiter='\n')
##################输出当前层的量化激活
q_activation_txt = self.activation_quantizer.get_quantize_value(
output)
a_para = q_activation_txt
#############输入特征图重排序
if self.reorder == True:
# 重排序参数
# print("use activation reorder!")
shape_input = a_para.shape[1]
num_TN = int(shape_input / self.TN)
remainder_TN = shape_input % self.TN
first = True
reorder_a_para = None
if self.activate == 'linear':
print('layer-linear reorder!')
temp = a_para[:, 0:remainder_TN, :, :]
temp = temp.view(temp.shape[1], temp.shape[2],
temp.shape[3])
temp = temp.permute(2, 1, 0).contiguous().view(-1)
if first:
reorder_a_para = temp.clone().cpu().data.numpy()
first = False
else:
reorder_a_para = np.append(reorder_a_para,
temp.cpu().data.numpy())
else:
for k in range(num_TN):
temp = a_para[:,
k * self.TN:(k + 1) * self.TN, :, :]
temp = temp.view(temp.shape[1], temp.shape[2],
temp.shape[3])
temp = temp.permute(2, 1, 0).contiguous().view(-1)
if first:
reorder_a_para = temp.clone().cpu().data.numpy(
)
first = False
else:
reorder_a_para = np.append(
reorder_a_para,
temp.cpu().data.numpy())
a_para_flatten = reorder_a_para
#####验证重排序结果的正确性
'''if a_para_flatten.size == a_para.shape[0] * a_para.shape[1] * a_para.shape[2] * a_para.shape[3]:
print("activation convert correctly!")
else:
print("activation convert mismatchingly!")'''
q_activation_reorder = a_para_flatten
q_activation_reorder = np.array(
q_activation_reorder).reshape(1, -1)
np.savetxt((
'./quantizer_output/q_activation_reorder/a_reorder_%s.txt'
% self.name),
q_activation_reorder,
delimiter='\n')
###保存重排序的二进制文件
activation_flat = q_activation_reorder.astype(np.int8)
writer = open(
'./quantizer_output/q_activation_reorder/%s_activation_q_bin'
% self.name, "wb")
writer.write(activation_flat)
writer.close()
##########特征图重排序结束
q_activation_txt = np.array(q_activation_txt.cpu()).reshape(
1, -1)
q_activation_max = [np.max(q_activation_txt)
] # 统计该层的最大值(即查看是否有溢出)
max_activation_count = [
np.sum(
abs(q_activation_txt) >= (1 << (self.w_bits - 1)) - 1)
] # 统计该层溢出的数目
# q_weight_max = np.argmax(q_weight_txt)
np.savetxt((
'./quantizer_output/max_activation_count/max_a_count_%s.txt'
% self.name), max_activation_count)
np.savetxt(
('./quantizer_output/q_activation_max/q_a_max_%s.txt' %
self.name), q_activation_max)
np.savetxt(
('./quantizer_output/q_activation_out/q_activation_%s.txt'
% self.name),
q_activation_txt,
delimiter='\n')
output = self.activation_quantizer(output)
return output
def forward(self, x):
return F.relu6(self.bn(self.conv(x)))