def __init__(self, input_size,n_head=4, identity=False):
super(multiSeqAttnMatch, self).__init__()
self.hidden_size = input_size // n_head
self.w = nn.Parameter(torch.FloatTensor(n_head, input_size, self.hidden_size))
init.xavier_normal(self.w)
self.n_head = n_head
def weights_init_xavier(m):
classname = m.__class__.__name__
# print(classname)
if classname.find('Conv') != -1:
init.xavier_normal(m.weight.data, gain=0.02)
elif classname.find('Linear') != -1:
init.xavier_normal(m.weight.data, gain=0.02)
elif classname.find('BatchNorm2d') != -1:
init.normal(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
def __init__(self, args):
super(BiGRU, self).__init__()
self.hidden_dim = args.hidden_dim
self.batch_size = args.batch_size
self.dropout = nn.Dropout(args.dropout)
self.dropout_embed = nn.Dropout(args.dropout_embed)
self.word_embeddings = nn.Embedding(args.embed_num, args.embedding_dim)
self.bigru = nn.GRU(args.embedding_dim, args.hidden_dim, bidirectional=True, dropout=args.dropout_model)
# self.hidden2label1 = nn.Linear(args.hidden_dim * 2, args.hidden_dim)
# self.hidden2label2 = nn.Linear(args.hidden_dim, args.class_num)
self.hidden2label = nn.Linear(args.hidden_dim * 2, args.class_num)
self.hidden = self.init_hidden(args.batch_size)
pretrained_weight = np.array(args.pretrained_weight)
# print(pretrained_weight.shape)
self.word_embeddings.weight.data.copy_(torch.from_numpy(pretrained_weight))
# weight modify, gru do not have forget gate
init.xavier_normal(self.bigru.all_weights[0][0], gain=np.sqrt(2.0))
init.xavier_normal(self.bigru.all_weights[0][1], gain=np.sqrt(2.0))
init.xavier_normal(self.bigru.all_weights[1][0], gain=np.sqrt(2.0))
init.xavier_normal(self.bigru.all_weights[1][1], gain=np.sqrt(2.0))
# self.bigru.all_weights[0][3].data.fill_(0.1)
# self.bigru.all_weights[0][2].data.fill_(0.1)
# self.bigru.all_weights[1][3].data.fill_(0.1)
# self.bigru.all_weights[1][2].data.fill_(0.1)
self.bn1 = nn.BatchNorm1d(600)
def __init__(self, args):
super(BiLSTM_1, self).__init__()
self.args = args
self.hidden_dim = args.lstm_hidden_dim
self.num_layers = args.lstm_num_layers
V = args.embed_num
D = args.embed_dim
C = args.class_num
self.dropout = nn.Dropout(args.dropout)
self.dropout_embed = nn.Dropout(args.dropout_embed)
if args.max_norm is not None:
print("max_norm = {} ".format(args.max_norm))
self.embed = nn.Embedding(V, D, max_norm=args.max_norm, scale_grad_by_freq=True)
else:
print("max_norm = {} |||||".format(args.max_norm))
self.embed = nn.Embedding(V, D, scale_grad_by_freq=True)
if args.word_Embedding:
pretrained_weight = np.array(args.pretrained_weight)
self.embed.weight.data.copy_(torch.from_numpy(pretrained_weight))
self.bilstm = nn.LSTM(D, self.hidden_dim, num_layers=self.num_layers, bias=True, bidirectional=True,
dropout=self.args.dropout)
print(self.bilstm)
if args.init_weight:
print("Initing W .......")
init.xavier_normal(self.bilstm.all_weights[0][0], gain=np.sqrt(args.init_weight_value))
init.xavier_normal(self.bilstm.all_weights[0][1], gain=np.sqrt(args.init_weight_value))
init.xavier_normal(self.bilstm.all_weights[1][0], gain=np.sqrt(args.init_weight_value))
init.xavier_normal(self.bilstm.all_weights[1][1], gain=np.sqrt(args.init_weight_value))
# print("eeeeeeeeeeeeeeeeeeeeeeee")
# fan_in, fan_out = BiLSTM_1.calculate_fan_in_and_fan_out(self.bilstm.all_weights[1][1])
# print(" in {} out {} ".format(fan_in, fan_out))
# std = np.sqrt(args.init_weight_value) * np.sqrt(2.0 / (fan_in + fan_out))
# print("aaaaaaaaaaaaa {} ".format(std))
# print("self.bilstm.all_weights {} ".format(self.bilstm.all_weights))
# self.bilstm.all_weights[0][3].data.fill_(0)
# self.bilstm.all_weights[0][2].data.fill_(0)
# self.bilstm.all_weights[1][3].data.fill_(0)
# self.bilstm.all_weights[1][2].data.fill_(0)
# self.bilstm.all_weights[0][3].data[20:40].fill_(1)
# self.bilstm.all_weights[0][3].data[0:20].fill_(0)
# self.bilstm.all_weights[0][3].data[40:80].fill_(0)
# # self.bilstm.all_weights[0][3].data[40:].fill_(0)
# self.bilstm.all_weights[0][2].data[20:40].fill_(1)
# self.bilstm.all_weights[0][2].data[0:20].fill_(0)
# self.bilstm.all_weights[0][2].data[40:80].fill_(0)
# # self.bilstm.all_weights[0][2].data[40:].fill_(0)
# self.bilstm.all_weights[1][3].data[20:40].fill_(1)
# self.bilstm.all_weights[1][3].data[0:20].fill_(0)
# self.bilstm.all_weights[1][3].data[40:80].fill_(0)
# # self.bilstm.all_weights[1][3].data[40:].fill_(0)
# self.bilstm.all_weights[1][2].data[20:40].fill_(1)
# self.bilstm.all_weights[1][2].data[0:20].fill_(0)
# self.bilstm.all_weights[1][2].data[40:80].fill_(0)
# # self.bilstm.all_weights[1][2].data[40:].fill_(0)
# self.hidden2label1 = nn.Linear(self.hidden_dim * 2, self.hidden_dim)
# self.hidden2label2 = nn.Linear(self.hidden_dim, C)
self.hidden2label = nn.Linear(self.hidden_dim * 2, C)
self.hidden = self.init_hidden(self.num_layers, args.batch_size)
print("self.hidden", self.hidden)
def __init__(self, args):
super(BiLSTM, self).__init__()
self.hidden_dim = args.hidden_dim
self.batch_size = args.batch_size
self.dropout = nn.Dropout(args.dropout)
self.dropout_embed = nn.Dropout(args.dropout_embed)
self.word_embeddings = nn.Embedding(args.embed_num, args.embedding_dim, max_norm=5.0)
self.lstm = nn.LSTM(args.embedding_dim, args.hidden_dim, bidirectional=True, dropout=args.dropout_model)
self.hidden2label1 = nn.Linear(args.hidden_dim * 2, args.hidden_dim)
self.hidden2label2 = nn.Linear(args.hidden_dim, args.class_num)
self.hidden = self.init_hidden(args.batch_size)
pretrained_weight = np.array(args.pretrained_weight)
# print(pretrained_weight.shape)
self.word_embeddings.weight.data.copy_(torch.from_numpy(pretrained_weight))
# weight modify
init.xavier_normal(self.lstm.all_weights[0][0], gain=np.sqrt(2.0))
init.xavier_normal(self.lstm.all_weights[0][1], gain=np.sqrt(2.0))
init.xavier_normal(self.lstm.all_weights[1][0], gain=np.sqrt(2.0))
init.xavier_normal(self.lstm.all_weights[1][1], gain=np.sqrt(2.0))
self.lstm.all_weights[0][3].data[20:40].fill_(1)
self.lstm.all_weights[0][3].data[0:20].fill_(0)
self.lstm.all_weights[0][3].data[40:80].fill_(0)
self.lstm.all_weights[0][2].data[20:40].fill_(1)
self.lstm.all_weights[0][2].data[0:20].fill_(0)
self.lstm.all_weights[0][2].data[40:80].fill_(0)
self.lstm.all_weights[1][3].data[20:40].fill_(1)
self.lstm.all_weights[1][3].data[0:20].fill_(0)
self.lstm.all_weights[1][3].data[40:80].fill_(0)
self.lstm.all_weights[1][2].data[20:40].fill_(1)
self.lstm.all_weights[1][2].data[0:20].fill_(0)
self.lstm.all_weights[1][2].data[40:80].fill_(0)
def __init__(self, args):
super(DEEP_CNN_MUI, self).__init__()
self.args = args
V = args.embed_num
V_mui = args.embed_num_mui
D = args.embed_dim
C = args.class_num
Ci = 2
Co = args.kernel_num
Ks = args.kernel_sizes
if args.max_norm is not None:
print("max_norm = {} ".format(args.max_norm))
self.embed_no_static = nn.Embedding(V, D, max_norm=args.max_norm, scale_grad_by_freq=True)
self.embed_static = nn.Embedding(V_mui, D, max_norm=args.max_norm, scale_grad_by_freq=True)
else:
print("max_norm = {} ".format(args.max_norm))
self.embed_no_static = nn.Embedding(V, D, scale_grad_by_freq=True)
self.embed_static = nn.Embedding(V_mui, D, scale_grad_by_freq=True)
if args.word_Embedding:
pretrained_weight = np.array(args.pretrained_weight)
self.embed_no_static.weight.data.copy_(torch.from_numpy(pretrained_weight))
pretrained_weight_static = np.array(args.pretrained_weight_static)
self.embed_static.weight.data.copy_(torch.from_numpy(pretrained_weight_static))
# whether to fixed the word embedding
self.embed_no_static.weight.requires_grad = True
# cons layer
self.convs1 = [nn.Conv2d(Ci, D, (K, D), stride=1, padding=(K//2, 0), bias=True) for K in Ks]
self.convs2 = [nn.Conv2d(1, Co, (K, D), stride=1, padding=(K//2, 0), bias=True) for K in Ks]
print(self.convs1)
print(self.convs2)
if args.init_weight:
print("Initing W .......")
for (conv1, conv2) in zip(self.convs1, self.convs2):
init.xavier_normal(conv1.weight.data, gain=np.sqrt(args.init_weight_value))
init.uniform(conv1.bias, 0, 0)
init.xavier_normal(conv2.weight.data, gain=np.sqrt(args.init_weight_value))
init.uniform(conv2.bias, 0, 0)
# dropout
self.dropout = nn.Dropout(args.dropout)
# linear
in_fea = len(Ks) * Co
self.fc1 = nn.Linear(in_features=in_fea, out_features=in_fea // 2, bias=True)
self.fc2 = nn.Linear(in_features=in_fea // 2, out_features=C, bias=True)
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, args):
super(CNN, self).__init__()
self.args = args
# self.conv1l = nn.Conv2d(3,20,5,stride=1,bias=True)
# init.xavier_normal()
Ci = 1
self.embed = nn.Embedding(args.embed_num, args.embedding_dim)
# print(self.embed)
# pretrained_weight is a numpy matrix of shape (num_embeddings, embedding_dim)
# print(len(args.pretrained_weight))
pretrained_weight = np.array(args.pretrained_weight)
# print(pretrained_weight.shape)
self.embed.weight.data.copy_(torch.from_numpy(pretrained_weight))
# print(self.embed) # 15453 Embedding(15453, 128)
self.convs1 = [nn.Conv2d(Ci, args.kernel_num, (K, args.embedding_dim)) for K in args.kernel_sizes]
# init.xavier_normal([(conv.weight, gain=np.sqrt(2.0) for conv in self.convsl)])
for conv in self.convs1:
init.xavier_normal(conv.weight, gain=np.sqrt(2.0))
# init.normal(conv.weight, mean=0, std=0.1)
# init.constant(conv.bias, 0.1)
# print(self.convs1)
# self.conv13 = nn.Conv2d(Ci, Co, (3, D))
# self.conv14 = nn.Conv2d(Ci, Co, (4, D))
# self.conv15 = nn.Conv2d(Ci, Co, (5, D))
# self.conv16 = nn.Conv2d(Ci, Co, (6, D))
self.dropout = nn.Dropout(args.dropout)
self.fc1 = nn.Linear(len(args.kernel_sizes) * args.kernel_num, args.class_num) # len(Ks)*Co -> C
# self.fc1 = nn.Linear(len(Ks) * Co * 2, C) # len(Ks)*Co*2 -> C
# self.bn = nn.BatchNorm1d(1, momentum=0.5)
self.bn = nn.BatchNorm2d(1)
def __init__(self, args):
super(LSTM, self).__init__()
self.args = args
# print(args)
self.hidden_dim = args.lstm_hidden_dim
self.num_layers = args.lstm_num_layers
V = args.embed_num
D = args.embed_dim
C = args.class_num
if args.max_norm is not None:
print("max_norm = {} ".format(args.max_norm))
self.embed = nn.Embedding(V, D, max_norm=args.max_norm, scale_grad_by_freq=True)
else:
print("max_norm = {} |||||".format(args.max_norm))
self.embed = nn.Embedding(V, D, scale_grad_by_freq=True)
# word embedding
if args.word_Embedding:
pretrained_weight = np.array(args.pretrained_weight)
self.embed.weight.data.copy_(torch.from_numpy(pretrained_weight))
# lstm
self.lstm = nn.LSTM(D, self.hidden_dim, dropout=args.dropout, num_layers=self.num_layers)
if args.init_weight:
print("Initing W .......")
# n = self.lstm.input_size * self.lstm
init.xavier_normal(self.lstm.all_weights[0][0], gain=np.sqrt(args.init_weight_value))
init.xavier_normal(self.lstm.all_weights[0][1], gain=np.sqrt(args.init_weight_value))
# linear
self.hidden2label = nn.Linear(self.hidden_dim, C)
# hidden
self.hidden = self.init_hidden(self.num_layers, args.batch_size)
# dropout
self.dropout = nn.Dropout(args.dropout)
self.dropout_embed = nn.Dropout(args.dropout_embed)
def __init__(self, in_size, out_size, kernel_size=3,stride=1, padding=1, activation=nn.ReLU(), space_dropout=False):
super(UNetUpBlock, self).__init__()
self.conv0 = nn.Conv2d(in_size, out_size, 3, stride=1, padding=1)
self.conv = nn.Conv2d(in_size, out_size, kernel_size, stride=1, padding=1)
self.conv2 = nn.Conv2d(out_size, out_size, kernel_size,stride=1, padding=1)
init.xavier_normal(self.conv0.weight,gain=np.sqrt(2))
init.xavier_normal(self.conv.weight,gain=np.sqrt(2))
init.xavier_normal(self.conv2.weight,gain=np.sqrt(2))
init.constant(self.conv0.bias, 0.1)
init.constant(self.conv.bias, 0.1)
init.constant(self.conv2.bias, 0.1)
self.activation = activation
self.upsampler = nn.Upsample(scale_factor=2)
def __init__(self, in_size, out_size, kernel_size=3, stride=1, padding=1, activation = nn.ReLU(), downsample=True):
super(UNetConvBlock, self).__init__()
self.conv_down = nn.Conv2d(in_size, in_size, kernel_size, stride=2, padding=1)
self.conv = nn.Conv2d(in_size, out_size, kernel_size, stride=1, padding=padding)
self.conv2 = nn.Conv2d(out_size, out_size, kernel_size,stride=1, padding=1)
init.xavier_normal(self.conv_down.weight,gain=np.sqrt(2))
init.xavier_normal(self.conv.weight,gain=np.sqrt(2))
init.xavier_normal(self.conv2.weight,gain=np.sqrt(2))
init.constant(self.conv_down.bias,0.1)
init.constant(self.conv.bias, 0.1)
init.constant(self.conv2.bias, 0.1)
self.activation = activation
self.downsample = downsample
def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
super(MultiHeadAttention, self).__init__()
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.w_qs = nn.Parameter(torch.FloatTensor(n_head, d_model, d_k))
self.w_ks = nn.Parameter(torch.FloatTensor(n_head, d_model, d_k))
self.w_vs = nn.Parameter(torch.FloatTensor(n_head, d_model, d_v))
self.attention = ScaledDotProductAttention(d_model)
self.layer_norm = LayerNormalization(d_model)
self.proj = Linear(n_head*d_v, d_model)
self.dropout = nn.Dropout(dropout)
init.xavier_normal(self.w_qs)
init.xavier_normal(self.w_ks)
init.xavier_normal(self.w_vs)
def __init__(self, **kwargs):
super(TextCNN, self).__init__()
self.input_size = kwargs['input_size']
self.hidden_size = kwargs['hidden_size']
self.output_size = kwargs['output_size']
if 'kernel_num' in kwargs:
self.kernel_num = kwargs['kernel_num']
else:
self.kernel_num = 256
if 'kernel_sizes' in kwargs:
self.kernel_sizes = kwargs['kernel_sizes']
else:
self.kernel_sizes = [1, 2, 3, 4]
if 'embed_size' in kwargs:
self.embed_size = kwargs['embed_size']
else:
self.embed_size = kwargs['hidden_size']
if 'dropout' in kwargs:
self.dropout = kwargs['dropout']
else:
self.dropout = 0.1
if 'wide_conv' in kwargs:
self.wide_conv = kwargs['wide_conv']
else:
self.wide_conv = False
if 'init_weight' in kwargs:
self.init_weight = kwargs['init_weight']
else:
self.init_weight = False
if 'init_weight_value' in kwargs:
self.init_weight_value = kwargs['init_weight_value']
else:
self.init_weight_value = 2.0
if 'batch_normal' in kwargs:
self.batch_normal = kwargs['batch_normal']
else:
self.batch_normal = False
if 'batch_normal_momentum' in kwargs:
self.batch_normal_momentum
else:
self.batch_normal_momentum = 0.1
if 'batch_normal_affine' in kwargs:
self.batch_normal_affine = kwargs['batch_normal_affine']
else:
self.batch_normal_affine = False
Ci = 1 # input channels, 处理文本,一层通道
Co = self.kernel_num # output channel
Ks = self.kernel_sizes # list
if 'max_norm' in kwargs:
self.embed = nn.Embedding(self.input_size,
self.embed_size,
max_norm=kwargs['max_norm'])
else:
self.embed = nn.Embedding(self.input_size,
self.embed_size,
scale_grad_by_freq=True)
if 'word_embedding' in kwargs:
pretrained_weight = torch.from_numpy(kwargs['word_embedding'])
self.embed.weight.data.copy_(pretrained_weight)
self.embed.weight.requires_grad = True
if self.wide_conv is True:
self.convs1 = [
nn.Conv2d(in_channels=Ci,
out_channels=Co,
kernel_size=(K, self.embed_size),
stride=(1, 1),
padding=(K // 2, 0),
dilation=1,
bias=True) for K in Ks
]
else:
self.convs1 = [
nn.Conv2d(in_channels=Ci,
out_channels=Co,
kernel_size=(K, self.embed_size),
bias=True) for K in Ks
]
if self.init_weight:
for conv in self.convs1:
init.xavier_normal(conv.weight.data,
gain=np.sqrt(self.init_weight_value))
fanin, fanout = self.cal_fanin_fanout(conv.weight.data)
std = np.sqrt(self.init_weight_value) * np.sqrt(
2.0 / (fanin + fanout))
init.uniform(conv.bias, 0, 0)
self.dropout = nn.Dropout(self.dropout)
in_fea = len(Ks) * Co
self.f1 = nn.Linear(in_fea, in_fea // 2, bias=True)
self.f2 = nn.Linear(in_fea // 2, self.output_size, bias=True)
if self.batch_normal:
self.convs1_bn = nn.BatchNorm2d(
num_features=Co,
momentum=self.batch_normal_momentum,
affine=self.batch_normal_affine)
self.f1_bn = nn.BatchNorm1d(num_features=in_fea // 2,
momentum=self.batch_normal_momentum,
affine=self.batch_normal_affine)
self.f2_bn = nn.BatchNorm1d(num_features=self.output_size,
momentum=self.batch_normal_momentum,
affine=self.batch_normal_affine)
def __init__(self, d_in, d_out, bias=True):
super(Linear, self).__init__()
self.linear = nn.Linear(d_in, d_out, bias=bias)
init.xavier_normal(self.linear.weight)
def __init__(self, args):
super(CNN_MUI, self).__init__()
self.args = args
V = args.embed_num
V_mui = args.embed_num_mui
D = args.embed_dim
C = args.class_num
Ci = 2
Co = args.kernel_num
Ks = args.kernel_sizes
if args.max_norm is not None:
print("max_norm = {} ".format(args.max_norm))
self.embed_no_static = nn.Embedding(V, D, max_norm=args.max_norm, scale_grad_by_freq=True)
self.embed_static = nn.Embedding(V_mui, D, max_norm=args.max_norm, scale_grad_by_freq=True)
# self.embed_static = nn.Embedding(V, D, max_norm=args.max_norm, scale_grad_by_freq=True)
else:
print("max_norm = {} ".format(args.max_norm))
self.embed_no_static = nn.Embedding(V, D, scale_grad_by_freq=True)
self.embed_static = nn.Embedding(V_mui, D, scale_grad_by_freq=True)
# self.embed_static = nn.Embedding(V, D, scale_grad_by_freq=True)
if args.word_Embedding:
pretrained_weight = np.array(args.pretrained_weight)
self.embed_no_static.weight.data.copy_(torch.from_numpy(pretrained_weight))
pretrained_weight_static = np.array(args.pretrained_weight_static)
self.embed_static.weight.data.copy_(torch.from_numpy(pretrained_weight_static))
# whether to fixed the word embedding
self.embed_no_static.weight.requires_grad = True
# self.embed_static.weight.requires_grad = False
if args.wide_conv is True:
print("using wide convolution")
self.convs1 = [nn.Conv2d(in_channels=Ci, out_channels=Co, kernel_size=(K, D), stride=(1, 1),
padding=(K//2, 0), bias=True) for K in Ks]
else:
print("using narrow convolution")
self.convs1 = [nn.Conv2d(in_channels=Ci, out_channels=Co, kernel_size=(K, D), bias=True) for K in Ks]
# self.convs1 = [nn.Conv2d(Ci, D, (K, D), stride=1, padding=(K // 2, 0)) for K in Ks]
print(self.convs1)
if args.init_weight:
print("Initing W .......")
for conv in self.convs1:
init.xavier_normal(conv.weight.data, gain=np.sqrt(args.init_weight_value))
init.uniform(conv.bias, 0, 0)
'''
self.conv13 = nn.Conv2d(Ci, Co, (3, D))
self.conv14 = nn.Conv2d(Ci, Co, (4, D))
self.conv15 = nn.Conv2d(Ci, Co, (5, D))
'''
self.dropout = nn.Dropout(args.dropout)
in_fea = len(Ks) * Co
self.fc1 = nn.Linear(in_features=in_fea, out_features=in_fea // 2, bias=True)
self.fc2 = nn.Linear(in_features=in_fea // 2, out_features=C, bias=True)
if args.batch_normalizations is True:
print("using batch_normalizations in the model......")
self.convs1_bn = nn.BatchNorm2d(num_features=Co, momentum=args.bath_norm_momentum,
affine=args.batch_norm_affine)
self.fc1_bn = nn.BatchNorm1d(num_features=in_fea//2, momentum=args.bath_norm_momentum,
affine=args.batch_norm_affine)
self.fc2_bn = nn.BatchNorm1d(num_features=C, momentum=args.bath_norm_momentum,
affine=args.batch_norm_affine)
def weights_init(m):
for _, mi in m._modules.items():
if isinstance(mi, nn.Conv2d) or isinstance(m, nn.Linear):
xavier_normal(mi.weight.data)
if mi.bias is not None:
xavier_normal(mi.bias.data)
def __init__(self, d_in, d_out, bias=True):
super(XavierLinear, self).__init__()
self.linear = nn.Linear(d_in, d_out, bias=bias)
init.xavier_normal(self.linear.weight)
def __init__(self, vocab_size, hidden_size):
super(AnswerModule, self).__init__()
self.z = nn.Linear(2 * hidden_size, vocab_size)
init.xavier_normal(self.z.state_dict()['weight'])
self.dropout = nn.Dropout(0.1)
def __init__(self):
super(Discriminator, self).__init__()
self.conv1 = nn.Conv2d(1,32,kernel_size=3,stride=2,padding=1) # 256x256
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1) # 128x128
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1) # 64x64
self.conv4 = nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1) # 32x32
self.conv5 = nn.Conv2d(256, 256, kernel_size=3, stride=2, padding=1) # 16x16
self.conv6 = nn.Conv2d(256, 512, kernel_size=3, stride=2, padding=1) # 8x8
self.conv7 = nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=1) # 4x4
self.conv8 = nn.Conv2d(512, 1024, kernel_size=4, stride=1, padding=0) # 1x1
self.bn1 = nn.BatchNorm2d(32)
self.bn2 = nn.BatchNorm2d(64)
self.bn3 = nn.BatchNorm2d(128)
self.bn4 = nn.BatchNorm2d(256)
self.bn5 = nn.BatchNorm2d(256)
self.bn6 = nn.BatchNorm2d(512)
self.bn7 = nn.BatchNorm2d(512)
self.sigmoid = nn.Sigmoid()
self.lrelu = nn.LeakyReLU(negative_slope=0.2)
init.xavier_normal(self.conv1.weight, gain=np.sqrt(2))
init.constant(self.conv1.bias, 0.1)
init.xavier_normal(self.conv2.weight, gain=np.sqrt(2))
init.constant(self.conv2.bias, 0.1)
init.xavier_normal(self.conv3.weight, gain=np.sqrt(2))
init.constant(self.conv3.bias, 0.1)
init.xavier_normal(self.conv4.weight, gain=np.sqrt(2))
init.constant(self.conv4.bias, 0.1)
init.xavier_normal(self.conv5.weight, gain=np.sqrt(2))
init.constant(self.conv5.bias, 0.1)
init.xavier_normal(self.conv6.weight, gain=np.sqrt(2))
init.constant(self.conv6.bias, 0.1)
init.xavier_normal(self.conv7.weight, gain=np.sqrt(2))
init.constant(self.conv7.bias, 0.1)
init.xavier_normal(self.conv8.weight, gain=np.sqrt(2))
init.constant(self.conv8.bias, 0.1)
def weight_init(m):
if isinstance(m, nn.Conv2d):
init.xavier_normal(m.weight)
init.constant(m.bias, 0)
def weights_init(m):
classname = m.__class__.__name__
if 'Linear' in classname:
init.xavier_normal(m.weight.data)
init.constant(m.bias, 0.0)
def init_weight(self):
init.xavier_normal(self.img_embed.weight)
init.xavier_normal(self.att_embed.weight)
def init_weights(layer):
if isinstance(layer, nn.Linear):
xavier_normal(layer.weight.data)
def xavier_init(model):
for param in model.parameters():
if len(param.size()) == 2:
xavier_normal(param)
def __init__(self, dimensions, **kwargs):
super(VAE, self).__init__()
assert len(dimensions) > 1
# unpack dimension of vae
self.embedding_dim = dimensions[0]
self.hidden_dims = dimensions[1:-1]
self.latent_dim = dimensions[-1]
self.dec_final_act = kwargs['decoder_final_activation']
self.device = torch.device('cuda' if (
torch.cuda.is_available()) else 'cpu')
self.is_logits = kwargs.get('logits', False)
if self.is_logits:
self.resconstruction_loss = nn.modules.loss.MSELoss()
else:
self.resconstruction_loss = self.binary_cross_entropy
# Construct layers for encoder and decoder block
# Encoder block
self.enc_hidden_layers = nn.Sequential()
self.enc_hidden_layers.add_module(
'hidden_layer_0', nn.Linear(self.embedding_dim,
self.hidden_dims[0]))
self.enc_hidden_layers.add_module('h_layer_act_0', nn.ReLU())
for i, _ in enumerate(self.hidden_dims[:-1]):
self.enc_hidden_layers.add_module(
'hidden_layer_{}'.format(i + 1),
nn.Linear(self.hidden_dims[i], self.hidden_dims[i + 1])),
self.enc_hidden_layers.add_module('h_layer_act_{}'.format(i + 1),
nn.ReLU())
# define mean and log variance of vae
self.z_mean = nn.Linear(self.hidden_dims[-1], self.latent_dim)
self.z_log_var = nn.Linear(self.hidden_dims[-1], self.latent_dim)
# ~Encoder block
# Decoder block
dec_hidden_layers = nn.Sequential()
dec_hidden_layers.add_module(
'hidden_layer_0', nn.Linear(self.latent_dim, self.hidden_dims[-1]))
dec_hidden_layers.add_module('h_layer_act_0', nn.ReLU())
reversed_hidden_dims = list(reversed(self.hidden_dims))
for i, _ in enumerate(reversed(self.hidden_dims)):
if i == (len(reversed_hidden_dims) - 1):
dec_hidden_layers.add_module(
'hidden_layer_{}'.format(i + 1),
nn.Linear(reversed_hidden_dims[i], self.embedding_dim)),
else:
dec_hidden_layers.add_module(
'hidden_layer_{}'.format(i + 1),
nn.Linear(reversed_hidden_dims[i],
reversed_hidden_dims[i + 1])),
dec_hidden_layers.add_module('h_layer_act_{}'.format(i + 1),
nn.ReLU())
# Final activation function of decoder depends on data
if self.dec_final_act is not None:
if self.dec_final_act == 'sigmoid':
dec_hidden_layers.add_module('dec_final_act', nn.Sigmoid())
elif self.dec_final_act == 'tanh':
dec_hidden_layers.add_module('dec_final_act', nn.Tanh())
elif self.dec_final_act == 'relu':
dec_hidden_layers.add_module('dec_final_act', nn.ReLU())
else:
pass
self.decoder = dec_hidden_layers
# ~ Decoder block
for m in self.modules():
if isinstance(m, nn.Linear):
init.xavier_normal(m.weight.data)
if m.bias is not None:
m.bias.data.zero_()
self.to(self.device)
def main():
print("Loading data from '%s'" % opt.data)
dataset = torch.load(opt.data)
if opt.model_type == 'nmt':
if dataset.get("type", "text") not in ["bitext", "text"]:
print("WARNING: The provided dataset is not bilingual!")
elif opt.model_type == 'lm':
if dataset.get("type", "text") != 'monotext':
print("WARNING: The provided dataset is not monolingual!")
else:
raise NotImplementedError('Not valid model type %s' % opt.model_type)
dict_checkpoint = (opt.train_from
if opt.train_from else opt.train_from_state_dict)
if dict_checkpoint:
print('Loading dicts from checkpoint at %s' % dict_checkpoint)
checkpoint = torch.load(dict_checkpoint)
if opt.model_type == 'nmt':
assert checkpoint.get('type', None) is None or \
checkpoint['type'] == "nmt", \
"The loaded model is not neural machine translation!"
elif opt.model_type == 'lm':
assert checkpoint['type'] == "lm", \
"The loaded model is not a language model!"
dataset['dicts'] = checkpoint['dicts']
trainData = onmt.Dataset(dataset['train']['src'],
dataset['train']['tgt'],
opt.batch_size,
opt.gpus,
data_type=dataset.get("type", "text"))
validData = onmt.Dataset(dataset['valid']['src'],
dataset['valid']['tgt'],
opt.batch_size,
opt.gpus,
volatile=True,
data_type=dataset.get("type", "text"))
dicts = dataset['dicts']
model_opt = checkpoint['opt'] if dict_checkpoint else opt
if dicts.get('tgt', None) is None:
# Makes the code compatible with the language model
dicts['tgt'] = dicts['src']
if opt.model_type == 'nmt':
print(' * vocabulary size. source = %d; target = %d' %
(dicts['src'].size(), dicts['tgt'].size()))
elif opt.model_type == 'lm':
print(' * vocabulary size = %d' % (dicts['src'].size()))
print(' * number of training sentences. %d' % len(dataset['train']['src']))
print(' * maximum batch size. %d' % opt.batch_size)
print('Building model...')
if opt.model_type == 'nmt':
decoder = onmt.Decoders.getDecoder(model_opt.decoder_type)(
model_opt, dicts['tgt'])
encoder = onmt.Encoders.getEncoder(model_opt.encoder_type)(
model_opt, dicts['src'])
model = onmt.Models.NMTModel(encoder, decoder)
elif opt.model_type == 'lm':
model = onmt.LanguageModel.LM(model_opt, dicts['src'])
generator = nn.Sequential(
nn.Linear(model_opt.rnn_size, dicts['tgt'].size()), nn.LogSoftmax())
if opt.train_from:
print('Loading model from checkpoint at %s' % opt.train_from)
chk_model = checkpoint['model']
generator_state_dict = chk_model.generator.state_dict()
model_state_dict = {
k: v
for k, v in chk_model.state_dict().items() if 'generator' not in k
}
model.load_state_dict(model_state_dict)
generator.load_state_dict(generator_state_dict)
opt.start_epoch = checkpoint['epoch'] + 1
if opt.train_from_state_dict:
print('Loading model from state_dict at %s' %
opt.train_from_state_dict)
model.load_state_dict(checkpoint['model'])
generator.load_state_dict(checkpoint['generator'])
model_opt.start_epoch = opt.start_epoch
model_opt.epochs = opt.epochs
if len(opt.gpus) >= 1:
model.cuda()
generator.cuda()
else:
model.cpu()
generator.cpu()
if len(opt.gpus) > 1:
model = nn.DataParallel(model, device_ids=opt.gpus, dim=1)
generator = nn.DataParallel(generator, device_ids=opt.gpus, dim=0)
model_opt["gpus"] = opt.gpus
model.generator = generator
if not opt.train_from_state_dict and not opt.train_from:
for p in model.parameters():
#p.data.uniform_(-opt.param_init, opt.param_init)
if len(p.data.size()) > 1:
init.xavier_normal(p.data)
else:
p.data.uniform_(-opt.param_init, opt.param_init)
model.initialize_parameters(opt.param_init)
model.load_pretrained_vectors(opt)
if (not opt.train_from_state_dict
and not opt.train_from) or opt.change_optimizer:
optim = onmt.Optim(opt.optim,
opt.learning_rate,
opt.max_grad_norm,
lr_decay=opt.learning_rate_decay,
start_decay_at=opt.start_decay_at)
optim.set_parameters(model.parameters())
model_opt.learning_rate = opt.learning_rate
model_opt.learning_rate_decay = opt.learning_rate_decay
model_opt.save_each = opt.save_each
else:
print('Loading optimizer from checkpoint:')
optim = checkpoint['optim']
optim.optimizer.load_state_dict(
checkpoint['optim'].optimizer.state_dict())
optim.set_parameters(model.parameters())
nParams = sum([p.nelement() for p in model.parameters()])
print('* number of parameters: %d' % nParams)
if opt.train_from or opt.train_from_state_dict:
print(model_opt)
model_opt.use_learning_rate_decay = opt.use_learning_rate_decay
trainModel(model, trainData, validData, dataset, optim, model_opt)
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
xavier_normal(m.weight.data)
xavier_normal(m.bias.data)
def init_linear(linear):
init.xavier_normal(linear.weight)
linear.bias.data.zero_()
def __init__(self, args):
super(CNN_Text, self).__init__()
self.args = args
V = args.embed_num
D = args.embed_dim
C = args.class_num
Ci = 1
Co = args.kernel_num
Ks = args.kernel_sizes
if args.max_norm is not None:
print("max_norm = {} ".format(args.max_norm))
self.embed = nn.Embedding(V, D, max_norm=args.max_norm, scale_grad_by_freq=True)
# self.embed.weight.data.uniform(-0.1, 0.1)
else:
print("max_norm = {} ".format(args.max_norm))
self.embed = nn.Embedding(V, D, scale_grad_by_freq=True)
if args.word_Embedding:
pretrained_weight = np.array(args.pretrained_weight)
self.embed.weight.data.copy_(torch.from_numpy(pretrained_weight))
# fixed the word embedding
self.embed.weight.requires_grad = True
print("dddd {} ".format(self.embed.weight.data.size()))
if args.wide_conv is True:
print("using wide convolution")
self.convs1 = [nn.Conv2d(in_channels=Ci, out_channels=Co, kernel_size=(K, D), stride=(1, 1),
padding=(K//2, 0), dilation=1, bias=False) for K in Ks]
else:
print("using narrow convolution")
self.convs1 = [nn.Conv2d(in_channels=Ci, out_channels=Co, kernel_size=(K, D), bias=True) for K in Ks]
# self.convs1 = [nn.Conv2d(Ci, D, (K, D), stride=1, padding=(K // 2, 0)) for K in Ks]
print(self.convs1)
# for con in self.convs1:
# print("PP {} ".format(con.weight))
if args.init_weight:
print("Initing W .......")
for conv in self.convs1:
init.xavier_normal(conv.weight.data, gain=np.sqrt(args.init_weight_value))
fan_in, fan_out = CNN_Text.calculate_fan_in_and_fan_out(conv.weight.data)
print(" in {} out {} ".format(fan_in, fan_out))
std = np.sqrt(args.init_weight_value) * np.sqrt(2.0 / (fan_in + fan_out))
print("aaaaaaaaaaaaa {} ".format(std))
# init.uniform(conv.bias, 0, 0)
self.dropout = nn.Dropout(args.dropout)
self.dropout_embed = nn.Dropout(args.dropout_embed)
in_fea = len(Ks) * Co
# self.fc1 = nn.Linear(in_features=in_fea, out_features=in_fea // 2, bias=True)
# self.fc2 = nn.Linear(in_features=in_fea // 2, out_features=C, bias=True)
self.fc = nn.Linear(in_features=in_fea, out_features=C, bias=True)
# whether to use batch normalizations
if args.batch_normalizations is True:
print("using batch_normalizations in the model......")
self.convs1_bn = nn.BatchNorm2d(num_features=Co, momentum=args.bath_norm_momentum,
affine=args.batch_norm_affine)
self.fc1_bn = nn.BatchNorm1d(num_features=in_fea//2, momentum=args.bath_norm_momentum,
affine=args.batch_norm_affine)
self.fc2_bn = nn.BatchNorm1d(num_features=C, momentum=args.bath_norm_momentum,
affine=args.batch_norm_affine)
def weight_init(m):
if isinstance(m, nn.Conv2d):
init.xavier_normal(m.weight)
init.constant(m.bias, 0)