def __init__(self,
model_path: str,
name: str,
input_size: Tuple[int] = (128, 256),
**kwargs):
super(DGNetEncoder).__init__()
self.name = name
self.model = ft_netAB(751, norm=False, stride=1, pool='max')
state_dict = torch.load(model_path)
self.model.load_state_dict(state_dict['a'], strict=False)
self.model.classifier1.classifier = nn.Sequential()
self.model.classifier2.classifier = nn.Sequential()
self.model = self.model.eval().cuda()
self.size = input_size
self.transform = T.Compose([
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
"""
def __init__(self, hyperparameters, gpu_ids=[0]):
super(DGNet_Trainer, self).__init__()
lr_g = hyperparameters['lr_g'] #生成器学习率
lr_d = hyperparameters['lr_d'] #判别器学习率
ID_class = hyperparameters['ID_class']
if not 'apex' in hyperparameters.keys():
hyperparameters['apex'] = False
self.fp16 = hyperparameters['apex']
# Initiate the networks
# We do not need to manually set fp16 in the network for the new apex. So here I set fp16=False.
# 构建Es编码+解码过程 gen_a.encode()可以进行编码，gen_b.encode()可以进行解码
self.gen_a = AdaINGen(hyperparameters['input_dim_a'], hyperparameters['gen'], fp16 = False) # auto-encoder for domain a
self.gen_b = self.gen_a # auto-encoder for domain b
# ID_stride，外观编码器池化层的stride
if not 'ID_stride' in hyperparameters.keys():
hyperparameters['ID_stride'] = 2
# 构建外观编码器
if hyperparameters['ID_style']=='PCB':
self.id_a = PCB(ID_class)
elif hyperparameters['ID_style']=='AB': #使用的AB编码器
self.id_a = ft_netAB(ID_class, stride = hyperparameters['ID_stride'], norm=hyperparameters['norm_id'], pool=hyperparameters['pool'])
else:
self.id_a = ft_net(ID_class, norm=hyperparameters['norm_id'], pool=hyperparameters['pool']) # return 2048 now
# 浅拷贝，两者等同
self.id_b = self.id_a
# 鉴别器，使用的是一个多尺寸的鉴别器，即对图片进行几次缩放,并且对每次缩放都会预测，计算总的损失
self.dis_a = MsImageDis(3, hyperparameters['dis'], fp16 = False) # discriminator for domain a
self.dis_b = self.dis_a # discriminator for domain b
# load teachers 加载教师模型
if hyperparameters['teacher'] != "":
teacher_name = hyperparameters['teacher']
print(teacher_name)
# 构建教师模型
teacher_names = teacher_name.split(',')
teacher_model = nn.ModuleList()
teacher_count = 0
for teacher_name in teacher_names:
config_tmp = load_config(teacher_name)
if 'stride' in config_tmp:
stride = config_tmp['stride']
else:
stride = 2
# 网络搭建
model_tmp = ft_net(ID_class, stride = stride)
teacher_model_tmp = load_network(model_tmp, teacher_name)
teacher_model_tmp.model.fc = nn.Sequential() # remove the original fc layer in ImageNet
teacher_model_tmp = teacher_model_tmp.cuda()
if self.fp16:
teacher_model_tmp = amp.initialize(teacher_model_tmp, opt_level="O1")
teacher_model.append(teacher_model_tmp.cuda().eval())
teacher_count +=1
self.teacher_model = teacher_model
# 选择是否使用bn
if hyperparameters['train_bn']:
self.teacher_model = self.teacher_model.apply(train_bn)
# 实例正则化
self.instancenorm = nn.InstanceNorm2d(512, affine=False)
# RGB to one channel
if hyperparameters['single']=='edge':
self.single = to_edge
else:
self.single = to_gray(False)
# Random Erasing when training
if not 'erasing_p' in hyperparameters.keys():
self.erasing_p = 0
else:
self.erasing_p = hyperparameters['erasing_p'] # erasing_p表示随机擦除的概率
# 随机擦除矩形区域的一些像素，数据增强
self.single_re = RandomErasing(probability = self.erasing_p, mean=[0.0, 0.0, 0.0])
if not 'T_w' in hyperparameters.keys():
hyperparameters['T_w'] = 1
# Setup the optimizers 设置优化器的参数
beta1 = hyperparameters['beta1']
beta2 = hyperparameters['beta2']
dis_params = list(self.dis_a.parameters()) #+ list(self.dis_b.parameters())
gen_params = list(self.gen_a.parameters()) #+ list(self.gen_b.parameters())
# 使用Adam优化器
self.dis_opt = torch.optim.Adam([p for p in dis_params if p.requires_grad],
lr=lr_d, betas=(beta1, beta2), weight_decay=hyperparameters['weight_decay'])
self.gen_opt = torch.optim.Adam([p for p in gen_params if p.requires_grad],
lr=lr_g, betas=(beta1, beta2), weight_decay=hyperparameters['weight_decay'])
# id params
if hyperparameters['ID_style']=='PCB':
ignored_params = (list(map(id, self.id_a.classifier0.parameters() ))
+list(map(id, self.id_a.classifier1.parameters() ))
+list(map(id, self.id_a.classifier2.parameters() ))
+list(map(id, self.id_a.classifier3.parameters() ))
)
base_params = filter(lambda p: id(p) not in ignored_params, self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD([
{'params': base_params, 'lr': lr2},
{'params': self.id_a.classifier0.parameters(), 'lr': lr2*10},
{'params': self.id_a.classifier1.parameters(), 'lr': lr2*10},
{'params': self.id_a.classifier2.parameters(), 'lr': lr2*10},
{'params': self.id_a.classifier3.parameters(), 'lr': lr2*10}
], weight_decay=hyperparameters['weight_decay'], momentum=0.9, nesterov=True)
elif hyperparameters['ID_style']=='AB':
ignored_params = (list(map(id, self.id_a.classifier1.parameters()))
+ list(map(id, self.id_a.classifier2.parameters())))
# 获得基本的配置参数，如学习率
base_params = filter(lambda p: id(p) not in ignored_params, self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD([
{'params': base_params, 'lr': lr2},
{'params': self.id_a.classifier1.parameters(), 'lr': lr2*10},
{'params': self.id_a.classifier2.parameters(), 'lr': lr2*10}
], weight_decay=hyperparameters['weight_decay'], momentum=0.9, nesterov=True)
else:
ignored_params = list(map(id, self.id_a.classifier.parameters() ))
base_params = filter(lambda p: id(p) not in ignored_params, self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD([
{'params': base_params, 'lr': lr2},
{'params': self.id_a.classifier.parameters(), 'lr': lr2*10}
], weight_decay=hyperparameters['weight_decay'], momentum=0.9, nesterov=True)
# 选择各个网络优化的策略
self.dis_scheduler = get_scheduler(self.dis_opt, hyperparameters)
self.gen_scheduler = get_scheduler(self.gen_opt, hyperparameters)
self.id_scheduler = get_scheduler(self.id_opt, hyperparameters)
self.id_scheduler.gamma = hyperparameters['gamma2']
#ID Loss
self.id_criterion = nn.CrossEntropyLoss()
self.criterion_teacher = nn.KLDivLoss(size_average=False)
# Load VGG model if needed
if 'vgg_w' in hyperparameters.keys() and hyperparameters['vgg_w'] > 0:
self.vgg = load_vgg16(hyperparameters['vgg_model_path'] + '/models')
self.vgg.eval()
for param in self.vgg.parameters():
param.requires_grad = False
# save memory
if self.fp16:
# Name the FP16_Optimizer instance to replace the existing optimizer
assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
self.gen_a = self.gen_a.cuda()
self.dis_a = self.dis_a.cuda()
self.id_a = self.id_a.cuda()
self.gen_b = self.gen_a
self.dis_b = self.dis_a
self.id_b = self.id_a
self.gen_a, self.gen_opt = amp.initialize(self.gen_a, self.gen_opt, opt_level="O1")
self.dis_a, self.dis_opt = amp.initialize(self.dis_a, self.dis_opt, opt_level="O1")
self.id_a, self.id_opt = amp.initialize(self.id_a, self.id_opt, opt_level="O1")
return f
###################################################################
# ######################################################################
# Load Collected data Trained model
###load config###
config_path = os.path.join('G:/pose-estimation/tracking/reid/outputs', name,
'config.yaml')
with open(config_path, 'r') as stream:
config = yaml.load(stream)
model_structure = ft_netAB(config['ID_class'],
norm=config['norm_id'],
stride=config['ID_stride'],
pool=config['pool'])
if opt.PCB:
model_structure = PCB(config['ID_class'])
model = load_network(model_structure)
# Remove the final fc layer and classifier layer
model.model.fc = nn.Sequential()
model.classifier1.classifier = nn.Sequential()
model.classifier2.classifier = nn.Sequential()
# Change to test mode
model = model.eval()
if use_gpu:
def __init__(self, hyperparameters, gpu_ids=[0]):
super(DGNet_Trainer, self).__init__()
# 从配置文件获取生成模型和鉴别模型的学习率
lr_g = hyperparameters['lr_g']
lr_d = hyperparameters['lr_d']
# ID 类别
ID_class = hyperparameters['ID_class']
# 是否设置使用fp16，
if not 'apex' in hyperparameters.keys():
hyperparameters['apex'] = False
self.fp16 = hyperparameters['apex']
# Initiate the networks
# We do not need to manually set fp16 in the network for the new apex. So here I set fp16=False.
self.gen_a = AdaINGen(hyperparameters['input_dim_a'],
hyperparameters['gen'],
fp16=False) # auto-encoder for domain a
self.gen_b = self.gen_a # auto-encoder for domain b
'''
ft_netAB : Ea
'''
# ID_stride： 外观编码器池化层的stride
if not 'ID_stride' in hyperparameters.keys():
hyperparameters['ID_stride'] = 2
# id_a ： 外观编码器 -> Ea
if hyperparameters['ID_style'] == 'PCB':
self.id_a = PCB(ID_class)
elif hyperparameters['ID_style'] == 'AB':
self.id_a = ft_netAB(ID_class,
stride=hyperparameters['ID_stride'],
norm=hyperparameters['norm_id'],
pool=hyperparameters['pool'])
else:
self.id_a = ft_net(ID_class,
norm=hyperparameters['norm_id'],
pool=hyperparameters['pool']) # return 2048 now
self.id_b = self.id_a # 对图片b的操作与图片a的操作一致
# 判别器，使用的是一个多尺寸的判别器，就是对图片进行几次缩放，并且对每次缩放都会预测，计算总的损失
# 经过网络3个缩放，，分别为：[batch_size, 1, 64, 32],[batch_size, 1, 32, 16],[batch_size, 1, 16, 8]
self.dis_a = MsImageDis(3, hyperparameters['dis'],
fp16=False) # discriminator for domain a
self.dis_b = self.dis_a # discriminator for domain b
# load teachers
if hyperparameters['teacher'] != "":
teacher_name = hyperparameters['teacher']
print(teacher_name)
# 加载多个老师模型
teacher_names = teacher_name.split(',')
# 构建老师模型
teacher_model = nn.ModuleList() # 初始化为空，接下来开始填充
teacher_count = 0
for teacher_name in teacher_names:
config_tmp = load_config(teacher_name)
# 池化层的stride
if 'stride' in config_tmp:
stride = config_tmp['stride']
else:
stride = 2
# 开始搭建网络
model_tmp = ft_net(ID_class, stride=stride)
teacher_model_tmp = load_network(model_tmp, teacher_name)
teacher_model_tmp.model.fc = nn.Sequential(
) # remove the original fc layer in ImageNet
teacher_model_tmp = teacher_model_tmp.cuda()
# teacher_model_tmp,[3, 224, 224]
# 使用fp16
if self.fp16:
teacher_model_tmp = amp.initialize(teacher_model_tmp,
opt_level="O1")
teacher_model.append(teacher_model_tmp.cuda().eval(
)) # 第一个填充为 teacher_model_tmp.cuda().eval()
teacher_count += 1
self.teacher_model = teacher_model
# 是否使用batchnorm
if hyperparameters['train_bn']:
self.teacher_model = self.teacher_model.apply(train_bn)
# 实例正则化
self.instancenorm = nn.InstanceNorm2d(512, affine=False)
# RGB to one channel
# 因为Es 需要使用灰度图， 所以single 用来将图片转化为灰度图
if hyperparameters['single'] == 'edge':
self.single = to_edge
else:
self.single = to_gray(False)
# Random Erasing when training
# arasing_p 随机擦除的概率
if not 'erasing_p' in hyperparameters.keys():
self.erasing_p = 0
else:
self.erasing_p = hyperparameters['erasing_p']
# 对图片中的某一随机区域进行擦除，具体：将该区域的像素值设置为均值
self.single_re = RandomErasing(probability=self.erasing_p,
mean=[0.0, 0.0, 0.0])
if not 'T_w' in hyperparameters.keys():
hyperparameters['T_w'] = 1
# Setup the optimizers
beta1 = hyperparameters['beta1']
beta2 = hyperparameters['beta2']
dis_params = list(
self.dis_a.parameters()) #+ list(self.dis_b.parameters())
gen_params = list(
self.gen_a.parameters()) #+ list(self.gen_b.parameters())
self.dis_opt = torch.optim.Adam(
[p for p in dis_params if p.requires_grad],
lr=lr_d,
betas=(beta1, beta2),
weight_decay=hyperparameters['weight_decay'])
self.gen_opt = torch.optim.Adam(
[p for p in gen_params if p.requires_grad],
lr=lr_g,
betas=(beta1, beta2),
weight_decay=hyperparameters['weight_decay'])
# id params
# 修改 id_a模型中分类器的学习率
if hyperparameters['ID_style'] == 'PCB':
ignored_params = (
list(map(id, self.id_a.classifier0.parameters())) +
list(map(id, self.id_a.classifier1.parameters())) +
list(map(id, self.id_a.classifier2.parameters())) +
list(map(id, self.id_a.classifier3.parameters())))
base_params = filter(lambda p: id(p) not in ignored_params,
self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD(
[{
'params': base_params,
'lr': lr2
}, {
'params': self.id_a.classifier0.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier1.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier2.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier3.parameters(),
'lr': lr2 * 10
}],
weight_decay=hyperparameters['weight_decay'],
momentum=0.9,
nesterov=True)
elif hyperparameters['ID_style'] == 'AB':
ignored_params = (
list(map(id, self.id_a.classifier1.parameters())) +
list(map(id, self.id_a.classifier2.parameters())))
base_params = filter(lambda p: id(p) not in ignored_params,
self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD(
[{
'params': base_params,
'lr': lr2
}, {
'params': self.id_a.classifier1.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier2.parameters(),
'lr': lr2 * 10
}],
weight_decay=hyperparameters['weight_decay'],
momentum=0.9,
nesterov=True)
else:
ignored_params = list(map(id, self.id_a.classifier.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params,
self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD(
[{
'params': base_params,
'lr': lr2
}, {
'params': self.id_a.classifier.parameters(),
'lr': lr2 * 10
}],
weight_decay=hyperparameters['weight_decay'],
momentum=0.9,
nesterov=True)
# 生成器和判别器中的优化策略（学习率的更新策略）
self.dis_scheduler = get_scheduler(self.dis_opt, hyperparameters)
self.gen_scheduler = get_scheduler(self.gen_opt, hyperparameters)
self.id_scheduler = get_scheduler(self.id_opt, hyperparameters)
self.id_scheduler.gamma = hyperparameters['gamma2']
#ID Loss
self.id_criterion = nn.CrossEntropyLoss()
self.criterion_teacher = nn.KLDivLoss(
size_average=False) # 生成主要特征： Lprim
# Load VGG model if needed
if 'vgg_w' in hyperparameters.keys() and hyperparameters['vgg_w'] > 0:
self.vgg = load_vgg16(hyperparameters['vgg_model_path'] +
'/models')
self.vgg.eval()
for param in self.vgg.parameters():
param.requires_grad = False
# save memory
# 保存当前的模型，是为了提高计算效率
if self.fp16:
# Name the FP16_Optimizer instance to replace the existing optimizer
assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
self.gen_a = self.gen_a.cuda()
self.dis_a = self.dis_a.cuda()
self.id_a = self.id_a.cuda()
self.gen_b = self.gen_a
self.dis_b = self.dis_a
self.id_b = self.id_a
self.gen_a, self.gen_opt = amp.initialize(self.gen_a,
self.gen_opt,
opt_level="O1")
self.dis_a, self.dis_opt = amp.initialize(self.dis_a,
self.dis_opt,
opt_level="O1")
self.id_a, self.id_opt = amp.initialize(self.id_a,
self.id_opt,
opt_level="O1")
def __init__(self, hyperparameters, gpu_ids=[0]):
super(DGNet_Trainer, self).__init__()
# 从配置文件获取生成模型的和鉴别模型的学习率
lr_g = hyperparameters['lr_g']
lr_d = hyperparameters['lr_d']
# # ID的类别，这里要注意，不同的数据集都是不一样的，应该是训练数据集的ID数目，非测试集
ID_class = hyperparameters['ID_class']
# 看是否设置使用float16，估计float16可以增加精确度
if not 'apex' in hyperparameters.keys():
hyperparameters['apex'] = False
self.fp16 = hyperparameters['apex']
# Initiate the networks
# We do not need to manually set fp16 in the network for the new apex. So here I set fp16=False.
################################################################################################################
##这里是定义Es和G
# 注意这里包含了两个步骤，Es编码+解码过程，既然解码（论文Figure 2的黄色梯形G）包含到这里了，下面Ea应该不会包含解码过程了
# 因为这里是一个类，如后续gen_a.encode()可以进行编码，gen_b.encode()可以进行解码
self.gen_a = AdaINGen(hyperparameters['input_dim_a'],
hyperparameters['gen'],
fp16=False) # auto-encoder for domain a
self.gen_b = self.gen_a # auto-encoder for domain b
############################################################################################################################################
############################################################################################################################################
##这里是定义Ea
# ID_stride，外观编码器池化层的stride
if not 'ID_stride' in hyperparameters.keys():
hyperparameters['ID_stride'] = 2
# hyperparameters['ID_style']默认为'AB'，论文中的Ea编码器
#这里是设置Ea，有三种模型可以选择
#PCB模型，ft_netAB为改造后的resnet50，ft_net为resnet50
if hyperparameters['ID_style'] == 'PCB':
self.id_a = PCB(ID_class)
elif hyperparameters['ID_style'] == 'AB':
# 这是我们执行的模型，注意的是，id_a返回两个x（表示身份），获得f，具体介绍看函数内部
# 我们使用的是ft_netAB，是代码中Ea编码的过程，也就得到 ap code的过程，除了ap code还会得到两个分类结果
# 现在怀疑，该分类结果，可能就是行人重识别的结果
#ID_class表示有ID_class个不同ID的行人
self.id_a = ft_netAB(ID_class,
stride=hyperparameters['ID_stride'],
norm=hyperparameters['norm_id'],
pool=hyperparameters['pool'])
else:
self.id_a = ft_net(ID_class,
norm=hyperparameters['norm_id'],
pool=hyperparameters['pool']) # return 2048 now
# 这里进行的是浅拷贝，所以我认为他们的权重是一起的，可以理解为一个
self.id_b = self.id_a
############################################################################################################################################################
############################################################################################################################################################
##这里是定义D
# 鉴别器，行人重识别，这里使用的是一个多尺寸的鉴别器，大概就是说，对图片进行几次缩放,并且对每次缩放都会预测，计算总的损失
# 经过网络3个元素，分别大小为[batch_size,1，64，32], [batch_size,1，32，16], [batch_size,1，16，8]
self.dis_a = MsImageDis(3, hyperparameters['dis'],
fp16=False) # discriminator for domain a
self.dis_b = self.dis_a # discriminator for domain b
############################################################################################################################################################
############################################################################################################################################################
# load teachers
# 加载老师模型
# teacher:老师模型名称。对于DukeMTMC，您可以设置“best - duke”
if hyperparameters['teacher'] != "":
#teacher_name=best
teacher_name = hyperparameters['teacher']
print(teacher_name)
#有这个操作，我怀疑是可以加载多个教师模型
teacher_names = teacher_name.split(',')
#构建老师模型
teacher_model = nn.ModuleList()
teacher_count = 0
# 默认只有一个teacher_name='teacher_name'，所以其加载的模型配置文件为项目根目录models/best/opts.yaml模型
for teacher_name in teacher_names:
# 加载配置文件models/best/opts.yaml
config_tmp = load_config(teacher_name)
if 'stride' in config_tmp:
#stride=1
stride = config_tmp['stride']
else:
stride = 2
# 老师模型加载，老师模型为ft_net为resnet50
model_tmp = ft_net(ID_class, stride=stride)
teacher_model_tmp = load_network(model_tmp, teacher_name)
# 移除原本的全连接层
teacher_model_tmp.model.fc = nn.Sequential(
) # remove the original fc layer in ImageNet
teacher_model_tmp = teacher_model_tmp.cuda()
# summary(teacher_model_tmp, (3, 224, 224))
#使用浮点型
if self.fp16:
teacher_model_tmp = amp.initialize(teacher_model_tmp,
opt_level="O1")
teacher_model.append(teacher_model_tmp.cuda().eval())
teacher_count += 1
self.teacher_model = teacher_model
# 选择是否使用bn
if hyperparameters['train_bn']:
self.teacher_model = self.teacher_model.apply(train_bn)
############################################################################################################################################################
# 实例正则化
self.instancenorm = nn.InstanceNorm2d(512, affine=False)
# RGB to one channel
# 默认设置signal=gray，Es的输入为灰度图
if hyperparameters['single'] == 'edge':
self.single = to_edge
else:
self.single = to_gray(False)
# Random Erasing when training
#earsing_p表示随机擦除的概率
if not 'erasing_p' in hyperparameters.keys():
self.erasing_p = 0
else:
self.erasing_p = hyperparameters['erasing_p']
#随机擦除矩形区域的一些像素，应该类似于数据增强
self.single_re = RandomErasing(probability=self.erasing_p,
mean=[0.0, 0.0, 0.0])
# 设置T_w为1,T_w为primary feature learning loss的权重系数
if not 'T_w' in hyperparameters.keys():
hyperparameters['T_w'] = 1
################################################################################################
# Setup the optimizers
# 设置优化器参数
beta1 = hyperparameters['beta1']
beta2 = hyperparameters['beta2']
dis_params = list(
self.dis_a.parameters()) #+ list(self.dis_b.parameters())
gen_params = list(
self.gen_a.parameters()) #+ list(self.gen_b.parameters())
#使用Adams优化器，用Adams训练Es，G,D
self.dis_opt = torch.optim.Adam(
[p for p in dis_params if p.requires_grad],
lr=lr_d,
betas=(beta1, beta2),
weight_decay=hyperparameters['weight_decay'])
self.gen_opt = torch.optim.Adam(
[p for p in gen_params if p.requires_grad],
lr=lr_g,
betas=(beta1, beta2),
weight_decay=hyperparameters['weight_decay'])
# id params
# 因为ID_style默认为AB，所以这里不执行
if hyperparameters['ID_style'] == 'PCB':
ignored_params = (
list(map(id, self.id_a.classifier0.parameters())) +
list(map(id, self.id_a.classifier1.parameters())) +
list(map(id, self.id_a.classifier2.parameters())) +
list(map(id, self.id_a.classifier3.parameters())))
base_params = filter(lambda p: id(p) not in ignored_params,
self.id_a.parameters())
lr2 = hyperparameters['lr2']
#Ea 的优化器
self.id_opt = torch.optim.SGD(
[{
'params': base_params,
'lr': lr2
}, {
'params': self.id_a.classifier0.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier1.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier2.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier3.parameters(),
'lr': lr2 * 10
}],
weight_decay=hyperparameters['weight_decay'],
momentum=0.9,
nesterov=True)
# 这里是我们执行的代码
elif hyperparameters['ID_style'] == 'AB':
# 忽略的参数，应该是适用于'PCB'或者其他的，但是不适用于'AB'的
ignored_params = (
list(map(id, self.id_a.classifier1.parameters())) +
list(map(id, self.id_a.classifier2.parameters())))
# 获得基本的配置参数，如学习率
base_params = filter(lambda p: id(p) not in ignored_params,
self.id_a.parameters())
lr2 = hyperparameters['lr2']
#对Ea使用SGD
self.id_opt = torch.optim.SGD(
[{
'params': base_params,
'lr': lr2
}, {
'params': self.id_a.classifier1.parameters(),
'lr': lr2 * 10
}, {
'params': self.id_a.classifier2.parameters(),
'lr': lr2 * 10
}],
weight_decay=hyperparameters['weight_decay'],
momentum=0.9,
nesterov=True)
else:
ignored_params = list(map(id, self.id_a.classifier.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params,
self.id_a.parameters())
lr2 = hyperparameters['lr2']
self.id_opt = torch.optim.SGD(
[{
'params': base_params,
'lr': lr2
}, {
'params': self.id_a.classifier.parameters(),
'lr': lr2 * 10
}],
weight_decay=hyperparameters['weight_decay'],
momentum=0.9,
nesterov=True)
# 选择各个网络的优化
self.dis_scheduler = get_scheduler(self.dis_opt, hyperparameters)
self.gen_scheduler = get_scheduler(self.gen_opt, hyperparameters)
self.id_scheduler = get_scheduler(self.id_opt, hyperparameters)
self.id_scheduler.gamma = hyperparameters['gamma2']
#ID Loss
#交叉熵损失函数
self.id_criterion = nn.CrossEntropyLoss()
# KL散度
self.criterion_teacher = nn.KLDivLoss(size_average=False)
# Load VGG model if needed
if 'vgg_w' in hyperparameters.keys() and hyperparameters['vgg_w'] > 0:
self.vgg = load_vgg16(hyperparameters['vgg_model_path'] +
'/models')
self.vgg.eval()
for param in self.vgg.parameters():
param.requires_grad = False
# save memory
if self.fp16:
# Name the FP16_Optimizer instance to replace the existing optimizer
assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
self.gen_a = self.gen_a.cuda()
self.dis_a = self.dis_a.cuda()
self.id_a = self.id_a.cuda()
self.gen_b = self.gen_a
self.dis_b = self.dis_a
self.id_b = self.id_a
self.gen_a, self.gen_opt = amp.initialize(self.gen_a,
self.gen_opt,
opt_level="O1")
self.dis_a, self.dis_opt = amp.initialize(self.dis_a,
self.dis_opt,
opt_level="O1")
self.id_a, self.id_opt = amp.initialize(self.id_a,
self.id_opt,
opt_level="O1")
gallery_cam,gallery_label = get_id(gallery_path)
query_cam,query_label = get_id(query_path)
######################################################################
# Load Collected data Trained model
print('-------test-----------')
###load config###
config_path = os.path.join('../outputs',name,'config.yaml')
with open(config_path, 'r') as stream:
config = yaml.load(stream)
model_path, output_dim = get_model_stats()
model_structure = ft_netAB(output_dim, norm=config['norm_id'], stride=config['ID_stride'], pool=config['pool'])
model = load_network(model_structure, model_path)
# Remove the final fc layer and classifier layer
model.model.fc = nn.Sequential()
model.classifier1.classifier = nn.Sequential()
model.classifier2.classifier = nn.Sequential()
# Change to test mode
model = model.eval()
if use_gpu:
model = model.cuda()
# Extract feature
with torch.no_grad():