def init_param(model):
for name, param in model.named_parameters():
# skip over the embeddings so that the padding index ones are 0
if 'embed' in name:
continue
elif ('rnn' in name or 'lm' in name) and len(param.size()) >= 2:
init.orthogonal(param)
else:
init.normal(param, 0, 0.01)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
print(classname)
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
init.orthogonal(m.weight.data, gain=1)
elif isinstance(m, nn.Linear):
init.orthogonal(m.weight.data, gain=1)
elif isinstance(m, nn.BatchNorm2d):
init.normal(m.weight.data, 1.0, 0.02)
init.constant_(m.bias.data, 0.0)
def wt_init(self):
for name, param in self.named_parameters():
if 'gru' in name and 'weight' in name:
init.orthogonal(param)
elif 'linear' in name:
init.normal(
param, 0,
math.sqrt(2. /
float(self.hidden_size_1 + self.hidden_size_2 +
self.hidden_size_3)))
def __init__(self, input_size , hidden_size, n_layers=1):
super(Encoder, self).__init__()
self.n_layers = n_layers
self.hidden_size = hidden_size
#self.embedding = nn.Embedding(vocab_size, emb_size)
self.gru = nn.GRU(input_size, hidden_size, dropout=0.2, batch_first=True, bidirectional=True)
for w in self.gru.parameters(): # initialize the gate weights with orthogonal
if w.dim()>1:
weight_init.orthogonal(w)
def initWeights(net, scheme='orthogonal'):
print('Initializing weights. Warning: may overwrite sensitive bias parameters (e.g. batchnorm)')
for e in net.parameters():
if scheme == 'orthogonal':
if len(e.size()) >= 2:
init.orthogonal(e)
elif scheme == 'normal':
init.normal(e, std=1e-2)
elif scheme == 'xavier':
init.xavier_normal(e)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
#print(classname)
if 'Conv' in classname:
init.orthogonal(m.weight.data, gain=1)
elif 'Linear' in classname:
init.orthogonal(m.weight.data, gain=1)
elif 'BatchNorm' in classname:
init.normal(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
# print(classname)
if classname.find("Conv") != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find("Linear") != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find("BatchNorm2d") != -1:
init.uniform(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
def reset_parameters(self):
self.reducer.reset_parameters()
init.kaiming_normal(self.word_linear.weight.data)
init.constant(self.word_linear.bias.data, val=0)
init.kaiming_normal(self.tracker_cell.weight_ih.data)
init.orthogonal(self.tracker_cell.weight_hh.data)
init.constant(self.tracker_cell.bias_ih.data, val=0)
init.constant(self.tracker_cell.bias_hh.data, val=0)
init.kaiming_normal(self.trans_linear.weight.data)
init.constant(self.trans_linear.bias.data, val=0)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
print(classname)
if classname.find('Conv') != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find('Linear') != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find('BatchNorm2d') != -1:
init.normal(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
print(classname)
if classname.find('Conv') != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find('Linear') != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find('BatchNorm2d') != -1:
init.normal_(m.weight.data, 1.0, 0.02)
init.constant_(m.bias.data, 0.0)
def __init__(self,
growth_rate=32,
block_config=(6, 12, 24, 16),
num_init_features=64,
bn_size=4,
drop_rate=0,
num_classes=p_transform["n_labels"]):
super(MyDenseNet, self).__init__()
# First convolution
self.features = nn.Sequential(
OrderedDict([
('conv0',
nn.Conv2d(3,
num_init_features,
kernel_size=7,
stride=2,
padding=3,
bias=False)),
('norm0', nn.BatchNorm2d(num_init_features)),
('relu0', nn.ReLU(inplace=True)),
('pool0', nn.MaxPool2d(kernel_size=3, stride=2, padding=1)),
]))
# Each denseblock
num_features = num_init_features
final_num_features = 0
for i, num_layers in enumerate(block_config):
block = torchvision.models.densenet._DenseBlock(
num_layers=num_layers,
num_input_features=num_features,
bn_size=bn_size,
growth_rate=growth_rate,
drop_rate=drop_rate)
self.features.add_module('denseblock%d' % (i + 1), block)
num_features = num_features + num_layers * growth_rate
if i != len(block_config) - 1:
trans = torchvision.models.densenet._Transition(
num_input_features=num_features,
num_output_features=num_features // 2)
self.features.add_module('transition%d' % (i + 1), trans)
num_features = num_features // 2
# Final batch norm
self.features.add_module('norm5', nn.BatchNorm2d(num_features))
#self.classifier_drop = nn.Dropout(p=0.75)
# Linear layer
self.fc = nn.Linear(num_features, num_features / 4)
init.orthogonal(self.fc.weight, gain=np.sqrt(2.0))
self.classifier = nn.Linear(num_features / 4, num_classes)
self.classifier.weight.data.zero_()
def _weight_init(self, m):
if isinstance(m, nn.LSTM):
for n, p in m.named_parameters():
if 'weight' in n:
# init.xavier_normal(p.data)
init.orthogonal(p.data)
elif 'bias' in n:
p.data.zero_()
def reset_parameters(self):
"""
Initialize parameters following the way proposed in the paper.
"""
init.orthogonal(self.weight_ih.data)
weight_hh_data = torch.eye(self.hidden_size)
weight_hh_data = weight_hh_data.repeat(1, 3)
self.weight_hh.data.set_(weight_hh_data)
# The bias is just set to zero vectors.
if self.use_bias:
init.constant(self.bias.data, val=0)
def init_weights(self,init_type,init_scale):
# Initialize weight matrix
for p in self.parameters():
if init_type=="orthogonal" and p.dim()>=2:
nninit.orthogonal(p)
elif init_type=="uniform":
p.data.uniform_(-init_scale, init_scale)
elif init_type=="xavier_n" and p.dim()>=2:
nninit.xavier_normal(p)
elif init_type=="xavier_u" and p.dim()>=2:
nninit.xavier_uniform(p)
def __init__(self, x_dim, h_dim, act_func, W=None):
super(AutoEncoder, self).__init__()
self.x_dim = x_dim
self.h_dim = h_dim
self.f = act_func
if W is None:
self.W = Parameter(torch.FloatTensor(x_dim, h_dim))
# init.xavier_uniform(self.W)
init.orthogonal(self.W)
else:
self.W = Parameter(torch.FloatTensor(W))
def __init__(self, depth, pretrained=True, cut_at_pooling=False,
num_features=0, norm=False, dropout=0, num_classes=0,
num_diff_features=0, iden_pretrain = False,
model_path='/media/hh/disc_d/hh/open-reid-master/pretrained model/resnet50.pth'):
super(ResNet, self).__init__()
self.depth = depth
self.pretrained = pretrained
self.cut_at_pooling = cut_at_pooling
self.iden_pretrain = iden_pretrain
# Construct base (pretrained) resnet
if depth not in ResNet.__factory:
raise KeyError("Unsupported depth:", depth)
# self.base = ResNet.__factory[depth](pretrained=pretrained)
self.base = baseresnet.ResNet(baseresnet.Bottleneck, [3, 4, 6, 3])
if pretrained is True:
self.base.load_state_dict(torch.load(model_path))
self.relu = nn.ReLU(inplace=True)
if not self.cut_at_pooling:
self.num_features = num_features
self.num_diff_features = num_diff_features
self.norm = norm
self.dropout = dropout
self.has_embedding = num_features > 0
self.num_classes = num_classes
out_planes = self.base.fc.in_features
# Append new layers
if self.has_embedding:
self.feat = nn.Linear(out_planes, self.num_features)
self.feat_bn = nn.BatchNorm1d(self.num_features)
init.kaiming_normal(self.feat.weight, mode='fan_out')
init.constant(self.feat.bias, 0)
init.constant(self.feat_bn.weight, 1)
init.constant(self.feat_bn.bias, 0)
else:
# Change the num_features to CNN output channels
self.num_features = out_planes
if self.dropout > 0:
self.drop = nn.Dropout(self.dropout)
if self.num_diff_features > 0:
self.diff_feat = nn.Linear(self.num_features, self.num_diff_features)
init.orthogonal(self.diff_feat.weight)
init.constant(self.diff_feat.bias, 0)
if self.num_classes > 0:
self.classifier = nn.Linear(self.num_features, self.num_classes)
# init.orthogonal(self.classifier.weight)
init.normal(self.classifier.weight, std=0.001)
init.constant(self.classifier.bias, 0)
if not self.pretrained:
self.reset_params()
def reset_parameters(self):
init.orthogonal(self.alpha_weight_ih.data)
alpha_weight_hh_data = torch.eye(self.hidden_size)
alpha_weight_hh_data = alpha_weight_hh_data.repeat(1, 1)
self.alpha_weight_hh.data.set_(alpha_weight_hh_data)
if self.use_bias:
init.constant(self.alpha_bias.data, val=0)
def init_weights(self):
init.orthogonal(self.lstm.weight_ih_l0)
init.uniform(self.lstm.weight_hh_l0, a=-0.01, b=0.01)
embedding_weights = torch.FloatTensor(self.vocab_size, 100)
init.uniform(embedding_weights, a=-0.25, b=0.25)
for id, vec in id_to_vec.items():
embedding_weights[id] = vec
self.embedding.weight.data.copy_(embedding_weights)
def init_bilstm(self, hidden_dim):
init.xavier_uniform(self.encoder.weight_ih_l0)
init.xavier_uniform(self.encoder.weight_ih_l0_reverse)
init.orthogonal(self.encoder.weight_hh_l0)
init.orthogonal(self.encoder.weight_hh_l0_reverse)
bias = self.init_lstm_bias(self.encoder.bias_ih_l0, hidden_dim)
self.encoder.bias_ih_l0 = nn.Parameter(bias.clone())
self.encoder.bias_hh_l0 = nn.Parameter(bias.clone())
self.encoder.bias_ih_l0_reverse = nn.Parameter(bias.clone())
self.encoder.bias_hh_l0_reverse = nn.Parameter(bias.clone())
def init_weights(self):
init.uniform(self.lstm.weight_ih_l0, a=-0.01, b=0.01)
init.orthogonal(self.lstm.weight_hh_l0)
self.lstm.weight_ih_l0.requires_grad = True
self.lstm.weight_hh_l0.requires_grad = True
embedding_weights = torch.FloatTensor(self.vocab_size, self.input_size)
for id, vec in id_to_vec.items():
embedding_weights[id] = vec
self.embedding.weight = nn.Parameter(embedding_weights, requires_grad=True)
def init_weights(self):
init.orthogonal(self.rnn.weight_ih_l0)
init.uniform(self.rnn.weight_hh_l0, a=-0.01, b=0.01)
glove_embeddings = preprocessing.load_glove_embeddings()
embedding_weights = torch.FloatTensor(self.vocab_size, self.input_size)
init.uniform(embedding_weights, a=-0.25, b=0.25)
for k,v in glove_embeddings.items():
embedding_weights[k] = torch.FloatTensor(v)
embedding_weights[0] = torch.FloatTensor([0]*self.input_size)
del self.embedding.weight
self.embedding.weight = nn.Parameter(embedding_weights)
def init_weights(self):
init_range = self._init_range
init_std = self._gru_init_std
self._l1_embedding_layer.weight.data.copy_(
torch.from_numpy(self._l1_emb_vector))
self._l2_embedding_layer.weight.data.copy_(
torch.from_numpy(self._l2_emb_vector))
unk_n_var_1 = self._l1_embedding_layer.weight.data[1:2 + self._nr_unk +
self._var_size, :]
init.normal(unk_n_var_1, 0, 1)
unk_n_var_1 /= torch.norm(unk_n_var_1, p=2, dim=1).unsqueeze(
1) # normalise randomly initialised embeddings
self._l1_embedding_layer.weight.data[0, :] = 0
if not self._emb_trainable:
self._l1_embedding_layer.weight.requires_gard = False # size = entities + ph + non-ent-marker
unk_2 = self._l2_embedding_layer.weight.data[2:2 + self._nr_unk, :]
init.normal(unk_2, 0, 1)
unk_2 /= torch.norm(unk_2, p=2, dim=1).unsqueeze(
1) # normalise randomly initialised embeddings
# ^^^ init unk * 100 embeddings
self._l2_embedding_layer.weight.data[
1, :] = self._l1_embedding_layer.weight.data[1, :]
# ^^^ share @placeholder embedding
self._l2_embedding_layer.weight.data[2 + self._nr_unk: 2 + self._nr_unk + self._var_size, :] = \
self._l1_embedding_layer.weight.data[2 + self._nr_unk: 2 + self._nr_unk + self._var_size, :]
# ^^^ share @entityX embeddings
self._l2_embedding_layer.weight.data[0, :] = 0
if not self._emb_trainable:
self._l2_embedding_layer.weight.requires_gard = False # size = entities + ph + non-ent-marker
# DONE: initialise non-zero locations
# TODO: randomise in forward step?
gain = init.calculate_gain('tanh')
for p in self._recurrent_layer.parameters():
if p.dim() == 1:
p.data.normal_(0, init_std)
else:
init.orthogonal(p.data, gain)
for p in self._question_recurrent_layer.parameters():
if p.dim() == 1:
p.data.normal_(0, init_std)
else:
init.orthogonal(p.data, gain)
# self._embedding_projection_layer.weight.data.uniform_(-init_range, init_range)
self._output_layer.weight.data.uniform_(-init_range, init_range)
self._output_layer.bias.data.fill_(0)
self._mix_matrix.data.uniform_(-init_range, init_range)
def _initialize_weights(self):
"""" ""
if type == 'Tanh':
self.nonlinearity = 'tanh'
elif type == 'ReLU':
self.nonlinearity = 'relu'
elif type == 'Leaky':
self.nonlinearity = 'leaky_relu'
"""
init.orthogonal(self.conv1.weight,
init.calculate_gain(self.nonlinearity))
init.orthogonal(self.conv2.weight,
init.calculate_gain(self.nonlinearity))
def weights_init_orthogonal(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
init.orthogonal(m.weight.data, gain=1)
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('Linear') != -1:
init.orthogonal(m.weight.data, gain=1)
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('BatchNorm2d') != -1:
init.constant_(m.weight.data, 1.0)
init.constant_(m.bias.data, 0.0)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
# print(classname)
if classname.find('Conv') != -1:
try:
init.orthogonal(m.weight.data, gain=1)
except AttributeError:
weights_init_normal(m)
elif classname.find('Linear') != -1:
init.orthogonal(m.weight.data, gain=1)
elif classname.find('BatchNorm2d') != -1:
init.normal(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
def init_weights(self, init_type, init_scale):
# Initialize weight matrix
for p in self.parameters():
if init_type == "orthogonal" and p.dim() >= 2:
nninit.orthogonal(p)
elif init_type == "uniform":
p.data.uniform_(-init_scale, init_scale)
elif init_type == "xavier_n" and p.dim() >= 2:
nninit.xavier_normal(p)
elif init_type == "xavier_u" and p.dim() >= 2:
nninit.xavier_uniform(p)
# Initialize bias for the linear layer
self.hidden2tag.bias.data.fill_(0.0)
def torch_weight_init(m):
"""
Usage:
model = Model()
model.apply(weight_init)
"""
if isinstance(m, nn.Conv1d):
init.normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.Conv2d):
init.xavier_normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.Conv3d):
init.xavier_normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.ConvTranspose1d):
init.normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.ConvTranspose2d):
init.xavier_normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.ConvTranspose3d):
init.xavier_normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.BatchNorm1d):
init.normal(m.weight.data, mean=1, std=0.02)
init.constant(m.bias.data, 0)
elif isinstance(m, nn.BatchNorm2d):
init.normal(m.weight.data, mean=1, std=0.02)
init.constant(m.bias.data, 0)
elif isinstance(m, nn.BatchNorm3d):
init.normal(m.weight.data, mean=1, std=0.02)
init.constant(m.bias.data, 0)
elif isinstance(m, nn.Linear):
init.xavier_normal(m.weight.data)
init.normal(m.bias.data)
elif isinstance(m, nn.LSTM):
for param in m.parameters():
if len(param.shape) >= 2:
init.orthogonal(param.data)
else:
init.normal(param.data)
elif isinstance(m, nn.LSTMCell):
for param in m.parameters():
if len(param.shape) >= 2:
init.orthogonal(param.data)
else:
init.normal(param.data)
elif isinstance(m, nn.GRU):
for param in m.parameters():
if len(param.shape) >= 2:
init.orthogonal(param.data)
else:
init.normal(param.data)
elif isinstance(m, nn.GRUCell):
for param in m.parameters():
if len(param.shape) >= 2:
init.orthogonal(param.data)
else:
init.normal(param.data)
def reset_parameters(self):
"""
Initialize parameters following the way proposed in the paper.
"""
# The input-to-hidden weight matrix is initialized orthogonally.
init.orthogonal(self.weight_ih.data)
# The hidden-to-hidden weight matrix is initialized as an identity
# matrix.
weight_hh_data = torch.eye(self.hidden_size)
weight_hh_data = weight_hh_data.repeat(1, 3)
self.weight_hh.data.set_(weight_hh_data)
# The bias is just set to zero vectors.
init.constant(self.bias.data, val=0)
def __init__(self, input_size, hidden_size, vocab_size, n_layers=1):
super(Decoder, self).__init__()
self.n_layers = n_layers
self.hidden_size = hidden_size
self.vocab_size=vocab_size
#self.embedding = nn.Embedding(vocab_size, emb_size)
self.gru = nn.GRU(input_size, hidden_size, dropout=0.2, batch_first=True)
for w in self.gru.parameters(): # initialize the gate weights with orthogonal
if w.dim()>1:
weight_init.orthogonal(w)
self.out = nn.Linear(hidden_size, vocab_size)
self.softmax = nn.LogSoftmax()
self.dropout=nn.Dropout(p=0.3)
def __init__(self, in_dim, out_dim, p_dropout=0.0):
super().__init__()
self.input_dim = in_dim
self.output_dim = out_dim
self.dropout = p_dropout
self.W_i = nn.Linear(in_dim, out_dim)
init.xavier_uniform(self.W_i.weight)
self.W_i.bias = nn.Parameter(torch.FloatTensor(out_dim).zero_())
self.U_i = nn.Linear(out_dim, out_dim, bias=False)
init.orthogonal(self.U_i.weight)
self.W_f = nn.Linear(in_dim, out_dim)
init.xavier_uniform(self.W_f.weight)
self.W_f.bias = nn.Parameter(torch.FloatTensor(out_dim).fill_(1.0))
self.U_f = nn.Linear(out_dim, out_dim)
init.orthogonal(self.U_f.weight)
self.W_c = nn.Linear(in_dim, out_dim)
init.xavier_uniform(self.W_c.weight)
self.W_c.bias = nn.Parameter(torch.FloatTensor(out_dim).fill_(0.0))
self.U_c = nn.Linear(out_dim, out_dim)
init.orthogonal(self.U_c.weight)
self.W_o = nn.Linear(in_dim, out_dim)
init.xavier_uniform(self.W_o.weight)
self.W_o.bias = nn.Parameter(torch.FloatTensor(out_dim).fill_(0.0))
self.U_o = nn.Linear(out_dim, out_dim)
init.orthogonal(self.U_o.weight)
def weights_init_orthogonal(m):
classname = m.__class__.__name__
# print(classname)
if classname.find('Conv2d') != -1 or classname.find('ConvTranspose') != -1:
init.orthogonal(m.weight.data, gain=1)
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('Linear') != -1:
init.orthogonal(m.weight.data, gain=1)
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('BatchNorm2d') != -1:
init.normal_(m.weight.data, 1.0, 0.02)
init.constant_(m.bias.data, 0.0)
def __init__(self, num_features=0, norm=False, dropout=0,
num_diff_features=0):
super(Trip_embedding, self).__init__()
self.num_features = num_features
self.num_diff_features = num_diff_features
self.norm = norm
self.dropout = dropout
if self.dropout > 0:
self.drop = nn.Dropout(self.dropout)
if self.num_diff_features > 0:
self.diff_feat = nn.Linear(self.num_features, self.num_diff_features)
init.orthogonal(self.diff_feat.weight)
init.constant(self.diff_feat.bias, 0)
def initialize_encoder(self):
"""Manual weight/bias initialization.
"""
xavier_normal(self.gru_enc_f.weight_ih)
orthogonal(self.gru_enc_f.weight_hh)
self.gru_enc_f.bias_ih.data.zero_()
self.gru_enc_f.bias_hh.data.zero_()
xavier_normal(self.gru_enc_b.weight_ih)
orthogonal(self.gru_enc_b.weight_hh)
self.gru_enc_b.bias_ih.data.zero_()
self.gru_enc_b.bias_hh.data.zero_()
def weights_init_orthogonal(m):
classname = m.__class__.__name__
# print('initializing [%s] ...' % classname)
if classname.find('Conv') != -1:
init.orthogonal(m.weight.data, gain=1)
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('Linear') != -1:
init.orthogonal(m.weight.data, gain=1)
if m.bias is not None:
m.bias.data.zero_()
elif classname.find('BatchNorm2d') != -1:
init.normal(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
def init_func(m):
classname = m.__class__.__name__
if hasattr(m, 'weight') and (classname.find('Conv') != -1 or classname.find('Linear') != -1):
if init_type == 'normal':
init.normal(m.weight.data, 0.0, gain)
elif init_type == 'xavier':
init.xavier_normal(m.weight.data, gain=gain)
elif init_type == 'kaiming':
init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
elif init_type == 'orthogonal':
init.orthogonal(m.weight.data, gain=gain)
else:
raise NotImplementedError('initialization method [%s] is not implemented' % init_type)
if hasattr(m, 'bias') and m.bias is not None:
init.constant(m.bias.data, 0.0)
elif classname.find('BatchNorm2d') != -1:
init.normal(m.weight.data, 1.0, gain)
init.constant(m.bias.data, 0.0)
def init_fun(m):
classname = m.__class__.__name__
if (classname.find('Conv') == 0 or classname.find('Linear') == 0) and hasattr(m, 'weight'):
# print m.__class__.__name__
if init_type == 'gaussian':
init.normal(m.weight.data, 0.0, 0.02)
elif init_type == 'xavier':
init.xavier_normal(m.weight.data, gain=math.sqrt(2))
elif init_type == 'kaiming':
init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
elif init_type == 'orthogonal':
init.orthogonal(m.weight.data, gain=math.sqrt(2))
elif init_type == 'default':
pass
else:
assert 0, "Unsupported initialization: {}".format(init_type)
if hasattr(m, 'bias') and m.bias is not None:
init.constant(m.bias.data, 0.0)
def __init__(self, margin=0, num_feature=128):
super(AdaptTripletLoss, self).__init__()
self.margin = margin
self.ranking_loss = nn.MarginRankingLoss(margin=margin)
self.softmargin_loss = nn.SoftMarginLoss()
self.num_classes = num_feature
self.adp_metric_embedding1 = nn.Linear(3*self.num_classes, 3*self.num_classes, bias=False)
self.adp_metric_embedding1_bn = nn.BatchNorm1d(3*self.num_classes)
self.adp_metric_embedding2 = nn.Linear(3*self.num_classes, 2*self.num_classes, bias=False)
self.adp_metric_embedding2_bn = nn.BatchNorm1d(2 * self.num_classes)
self.adp_metric_embedding3 = nn.Linear(2*self.num_classes, 2*self.num_classes, bias=False)
self.adp_metric_embedding3_bn = nn.BatchNorm1d(2 * self.num_classes)
self.adp_metric_embedding4 = nn.Linear(2*self.num_classes, 2*self.num_classes, bias=False)
# self.adp_metric_embedding2_bn = nn.BatchNorm1d(self.num_classes)
# init.constant(self.adp_metric_embedding1.bias,0)
# init.constant(self.adp_metric_embedding2.bias,0)
init.kaiming_normal(self.adp_metric_embedding1.weight, mode='fan_out')
init.kaiming_normal(self.adp_metric_embedding2.weight, mode='fan_out')
init.kaiming_normal(self.adp_metric_embedding3.weight, mode='fan_out')
init.orthogonal(self.adp_metric_embedding4.weight)
def reset_parameters(self):
"""
Initialize parameters following the way proposed in the paper.
"""
# The input-to-hidden weight matrix is initialized orthogonally.
init.orthogonal(self.weight_ih.data)
# The hidden-to-hidden weight matrix is initialized as an identity
# matrix.
weight_hh_data = torch.eye(self.hidden_size)
weight_hh_data = weight_hh_data.repeat(4, 1)
self.weight_hh.data.set_(weight_hh_data)
# The bias is just set to zero vectors.
init.constant(self.bias.data, val=0)
# Initialization of BN parameters.
self.bn_ih.reset_parameters()
self.bn_hh.reset_parameters()
self.bn_c.reset_parameters()
self.bn_ih.bias.data.fill_(0)
self.bn_hh.bias.data.fill_(0)
self.bn_ih.weight.data.fill_(0.1)
self.bn_hh.weight.data.fill_(0.1)
self.bn_c.weight.data.fill_(0.1)
def _initialize_weights(self):
init.orthogonal(self.conv1.weight, init.calculate_gain('relu'))
init.orthogonal(self.conv2.weight, init.calculate_gain('relu'))
init.orthogonal(self.conv3.weight, init.calculate_gain('relu'))
init.orthogonal(self.conv4.weight)
def _initialize_orthogonal(conv):
prelu_gain = math.sqrt(2)
init.orthogonal(conv.weight, gain=prelu_gain)
if conv.bias is not None:
conv.bias.data.zero_()