def _init_classifier(self):
classifier = nn.Linear(
len(self.dan_module_names) * self.output_dim,
self.owner().num_classes)
init_params(classifier)
self.classifier = classifier
def _init_classifiers(self):
self.classifiers = nn.ModuleList()
for p in range(1, self.part_num + 1):
classifier = nn.Linear(self.output_dim, self.num_classes)
init_params(classifier)
self.classifiers.append(classifier)
def __init__(self, backbone, args, num_classes, **kwargs):
super().__init__(backbone, args, num_classes, **kwargs)
from torchreid.utils.torchtools import init_params
self.final_classifier = nn.Linear(
args['cls_dim'],
num_classes,
)
init_params(self.final_classifier)
def _init_reduction_layer(self):
reduction = nn.Sequential(
nn.Conv2d(self.input_dim,
self.output_dim,
kernel_size=1,
bias=False), nn.BatchNorm2d(self.output_dim),
nn.ReLU(inplace=True))
init_params(reduction)
self.reduction = reduction
def _init_fc_layer(self):
dropout_p = self.args['dropout']
if dropout_p is not None:
dropout_layer = [nn.Dropout(p=dropout_p)]
else:
dropout_layer = []
fc = nn.Sequential(nn.Linear(self.input_dim, self.output_dim),
nn.BatchNorm1d(self.output_dim),
nn.ReLU(inplace=True), *dropout_layer)
init_params(fc)
return fc
def _init_attention_modules(self):
args = self.args
self.dan_module_names = set()
DAN_module_names = {'cam', 'pam'} & set(args['dan_dan'])
use_head = not args['dan_dan_no_head']
self.use_dan = bool(DAN_module_names)
before_module = get_attention_module_instance('identity',
self.output_dim,
use_head=False)
self.dan_module_names.add('before_module')
self.before_module = before_module
if use_head:
init_params(before_module)
if 'cam' in DAN_module_names:
cam_module = get_attention_module_instance('cam',
self.input_dim,
out_dim=self.output_dim,
use_head=use_head)
init_params(cam_module)
self.dan_module_names.add('cam_module')
self.cam_module = cam_module
if 'pam' in DAN_module_names:
pam_module = get_attention_module_instance('pam',
self.input_dim,
out_dim=self.output_dim,
use_head=use_head)
init_params(pam_module)
self.dan_module_names.add('pam_module')
self.pam_module = pam_module
def _init_attention_modules(self):
args = self.args
self.dan_module_names = set()
DAN_module_names = {'cam', 'pam'} & set(args['abd_dan'])
use_head = not args['abd_dan_no_head']
self.use_dan = bool(DAN_module_names)
before_module = get_attention_module_instance('identity',
self.output_dim,
use_head=use_head)
self.dan_module_names.add('before_module')
self.before_module = before_module
if use_head:
init_params(before_module)
if 'cam' in DAN_module_names:
cam_module = get_attention_module_instance('cam',
self.output_dim,
use_head=use_head)
init_params(cam_module)
self.dan_module_names.add('cam_module')
self.cam_module = cam_module
if 'pam' in DAN_module_names:
pam_module = get_attention_module_instance('pam',
self.output_dim,
use_head=use_head)
init_params(pam_module)
self.dan_module_names.add('pam_module')
self.pam_module = pam_module
sum_conv = nn.Sequential(
nn.Dropout2d(0.1, False),
nn.Conv2d(self.output_dim, self.output_dim, kernel_size=1))
init_params(sum_conv)
self.sum_conv = sum_conv
def _init_classifier(self):
classifier = nn.Linear(self.output_dim, self.num_classes)
init_params(classifier)
return classifier
