1. 模型參數初始化
# ————————————————— 利用model.apply(weights_init)實現初始化
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('BatchNorm') != -1:
m.weight.data.fill_(1)
m.bias.data.zero_()
elif classname.find('Linear') != -1:
n = m.weight.size(1)
m.weight.data.normal_(0, 0.01)
m.bias.data = torch.ones(m.bias.data.size())
# ————————————————— 直接放在__init__構造函數中實現初始化
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm1d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight.data)
if m.bias is not None:
m.bias.data.zero_()
# —————————————————
self.weight = Parameter(torch.Tensor(out_features, in_features))
self.bias = Parameter(torch.FloatTensor(out_features))
nn.init.xavier_uniform_(self.weight)
nn.init.zero_(self.bias)
nn.init.constant_(m, initm)
# nn.init.kaiming_uniform_()
# self.weight.data.normal_(std=0.001)
2. 模型參數分組weight_decay
def separate_bn_prelu_params(model, ignored_params=[]):
bn_prelu_params = []
for m in model.modules():
if isinstance(m, nn.BatchNorm2d):
ignored_params += list(map(id, m.parameters()))
bn_prelu_params += m.parameters()
if isinstance(m, nn.BatchNorm1d):
ignored_params += list(map(id, m.parameters()))
bn_prelu_params += m.parameters()
elif isinstance(m, nn.PReLU):
ignored_params += list(map(id, m.parameters()))
bn_prelu_params += m.parameters()
base_params = list(filter(lambda p: id(p) not in ignored_params, model.parameters()))
return base_params, bn_prelu_params, ignored_params
OPTIMIZER = optim.SGD([
{'params': base_params, 'weight_decay': WEIGHT_DECAY},
{'params': fc_head_param, 'weight_decay': WEIGHT_DECAY * 10},
{'params': bn_prelu_params, 'weight_decay': 0.0}
], lr=LR, momentum=MOMENTUM ) # , nesterov=True
Note 1：PReLU(x) = max(0,x) + a * min(0,x). Here a is a learnable parameter. When called without arguments, nn.PReLU() uses a single parameter a across all input channels. If called with nn.PReLU(nChannels), a separate a is used for each input channel.
Note 2: weight decay should not be used when learning a for good performance.
Note 3: The default number of a to learn is 1, the default initial value of a is 0.25.
3. 參數分組weight_decay–其他
第2節中的內容可以滿足一般的參數分組需求，此部分可以滿足更個性化的分組需求。參考：face_evoLVe_Pytorch-master
自定義schedule
def schedule_lr(optimizer):
for params in optimizer.param_groups:
params['lr'] /= 10.
print(optimizer)
方法一：利用model.modules()和obj.__class__ （更普適）
# model.modules()和model.children()的區別：model.modules()會迭代地遍歷模型的所有子層，而model.children()只會遍歷模型下的一層
# 下面的關鍵詞if 'model'，源於模型定義文件。如model_resnet.py中自定義的所有nn.Module子類，都會前綴'model_resnet'，所以可通過這種方式一次性篩選出自定義的模塊
def separate_irse_bn_paras(model):
paras_only_bn = []
paras_no_bn = []
for layer in model.modules():
if 'model' in str(layer.__class__): # eg. a=[1,2] type(a): <class 'list'> a.__class__: <class 'list'>
continue
if 'container' in str(layer.__class__): # 去掉Sequential型的模塊
continue
else:
if 'batchnorm' in str(layer.__class__):
paras_only_bn.extend([*layer.parameters()])
else:
paras_no_bn.extend([*layer.parameters()]) # extend()用於在列表末尾一次性追加另一個序列中的多個值（用新列表擴展原來的列表）
return paras_only_bn, paras_no_bn
方法二：調用modules.parameters和named_parameters()
但是本質上，parameters()是根據named_parameters()獲取，named_parameters()是根據modules()獲取。使用此方法的前提是，須按下文1,2中的方式定義模型，或者利用Sequential+OrderedDict定義模型。
def separate_resnet_bn_paras(model):
all_parameters = model.parameters()
paras_only_bn = []
for pname, p in model.named_parameters():
if pname.find('bn') >= 0:
paras_only_bn.append(p)
paras_only_bn_id = list(map(id, paras_only_bn))
paras_no_bn = list(filter(lambda p: id(p) not in paras_only_bn_id, all_parameters))
return paras_only_bn, paras_no_bn
兩種方法的區別
參數分組的區別，其實對應了模型構造時的區別。舉例：
-
構造ResNet的basic block，在__init__()函數中定義了
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = BatchNorm2d(planes)
self.relu = ReLU(inplace = True)
… -
在forward()中定義
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
… -
對ResNet取model.name_parameters()返回的pname形如：
‘layer1.0.conv1.weight’
‘layer1.0.bn1.weight’
‘layer1.0.bn1.bias’
# layer對應conv2_x, …, conv5_x； '0’對應各layer中的block索引，比如conv2_x有3個block，對應索引爲layer1.0, …, layer1.2； 'conv1’就是__init__()中定義的self.conv1 -
若構造model時採用了Sequential()，則model.name_parameters()返回的pname形如：‘body.3.res_layer.1.weight’，此處的1.weight實際對應了BN的weight，無法通過pname.find(‘bn’)找到該模塊。
self.res_layer = Sequential(
Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False),
BatchNorm2d(depth),
ReLU(depth),
Conv2d(depth, depth, (3, 3), stride, 1, bias=False),
BatchNorm2d(depth)
) -
針對4中的情況，兩種解決辦法：利用OrderedDict修飾Sequential，或利用方法一
downsample = Sequential( OrderedDict([
(‘conv_ds’, conv1x1(self.inplanes, planes * block.expansion, stride)),
(‘bn_ds’, BatchNorm2d(planes * block.expansion)),
]))
# 如此，相應模塊的pname將會帶有’conv_ds’，‘bn_ds’字樣