def __init__(self, config): super(Model, self).__init__() self.config = config self.bilstm1 = nn.LSTM(input_size=config.emb_size, hidden_size=config.hidden_size, bidirectional=True, batch_first=True) self.bilstm2 = nn.LSTM(input_size=config.emb_size, hidden_size=config.hidden_size, bidirectional=True, batch_first=True) self.linear1 = nn.Linear(2*config.hidden_size, 2*config.hidden_size) self.linear2 = nn.Linear(2*config.hidden_size, 2*config.hidden_size) self.linear3 = nn.Linear(2*config.hidden_size, 2*config.hidden_size) self.fuse_linear1 = nn.Linear(4*config.hidden_size, 2*config.hidden_size) self.fuse_linear2 = nn.Linear(4*config.hidden_size, 2*config.hidden_size) self.bilinear = nn.Bilinear(2*config.hidden_size, 2*config.hidden_size, 1) self.align_linear1 = nn.Linear(2*config.hidden_size, 1) self.align_linear2 = nn.Linear(2*config.hidden_size, 1)
def __init__(self, args, rank_mode=False):
super(InteractionAggregator, self).__init__()
self.hidden_size = args.hidden_size
self.position_enc = nn.Embedding.from_pretrained(get_sinusoid_encoding_table(11, self.hidden_size,
padding_idx=0), freeze=True)
self.tanh = nn.Tanh()
self.relu = nn.ReLU(inplace=False)
self.softmax = nn.Softmax(dim=1)
self.rank_mode = rank_mode
self.M = nn.Bilinear(self.hidden_size, self.hidden_size, self.hidden_size)
nn.init.xavier_normal_(self.M.weight)
self.w_attn1 = nn.Linear(self.hidden_size, self.hidden_size)
nn.init.xavier_normal_(self.w_attn1.weight)
if self.rank_mode:
self.w_attn2 = nn.Linear(self.hidden_size, self.hidden_size)
nn.init.xavier_normal_(self.w_attn2.weight)
def __init__(self, word_vocab, rel_vocab, config):
super(RelationRanking, self).__init__()
self.config = config
rel1_vocab, rel2_vocab = rel_vocab
self.word_embed = Embeddings(word_vec_size=config.d_word_embed,
dicts=word_vocab)
self.rel1_embed = Embeddings(word_vec_size=config.d_rel_embed,
dicts=rel1_vocab)
self.rel2_embed = Embeddings(word_vec_size=config.d_rel_embed,
dicts=rel2_vocab)
# print(self.rel_embed.word_lookup_table.weight.data)
#rel_embed的初始化待改 rel_embed.lookup_table.weight.data.normal_(0, 0.1)
if self.config.rnn_type.lower() == 'gru':
self.rnn = nn.GRU(input_size=config.d_word_embed,
hidden_size=config.d_hidden,
num_layers=config.n_layers,
dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
else:
self.rnn = nn.LSTM(input_size=config.d_word_embed,
hidden_size=config.d_hidden,
num_layers=config.n_layers,
dropout=config.dropout_prob,
bidirectional=config.birnn,
batch_first=True)
self.dropout = nn.Dropout(p=config.dropout_prob)
seq_in_size = config.d_hidden
if self.config.birnn:
seq_in_size *= 2
self.question_attention = MLPWordSeqAttention(
input_size=config.d_rel_embed, seq_size=seq_in_size)
self.bilinear = nn.Bilinear(seq_in_size,
config.d_rel_embed,
1,
bias=False)
self.seq_out = nn.Sequential(
# nn.BatchNorm1d(seq_in_size),
self.dropout,
nn.Linear(seq_in_size, config.d_rel_embed))
def __init__(self, args):
super(NeuralNetwork, self).__init__()
self.args = args
self.patience = 0
self.init_clip_max_norm = 5.0
self.optimizer = None
self.best_result = [0, 0, 0, 0, 0, 0]
self.metrics = Metrics(self.args.score_file_path)
self.device = torch.device(
'cuda:0' if torch.cuda.is_available() else 'cpu')
config_class, model_class, tokenizer_class = MODEL_CLASSES[
args.model_type]
self.bert_config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
finetuning_task="classification",
num_labels=1)
self.bert_tokenizer = BertTokenizer.from_pretrained(
args.tokenizer_name
if args.tokenizer_name else args.model_name_or_path,
do_lower_case=args.do_lower_case)
special_tokens_dict = {
'eos_token': '[eos]',
'additional_special_tokens': ['[soe]', '[eoe]']
}
num_added_toks = self.bert_tokenizer.add_special_tokens(
special_tokens_dict)
self.bert_model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=self.bert_config)
self.bert_model.resize_token_embeddings(len(self.bert_tokenizer))
self.bert_model = self.bert_model.cuda()
self.attn = nn.Linear(768, 768)
self.rnn = nn.GRU(input_size=768,
hidden_size=200,
num_layers=1,
batch_first=True,
bidirectional=False)
self.bilinear = nn.Bilinear(768, 768, 1)
def __init__(self,inshape,kernelsize,hiddensize):
super(CNN_LSTM_predictor,self).__init__()
# shape parameters:
self.batch_size, self.n_filters = inshape
self.lstm_input = seq_len = 20 - kernelsize + 1
# layers
self.cnnPrice = nn.Conv1d(1,self.n_filters,kernelsize)
self.cnnVolume = nn.Conv1d(1,self.n_filters,kernelsize)
self.lstmPrice = nn.LSTM(self.lstm_input,self.lstm_input)
self.lstmVolume = nn.LSTM(self.lstm_input,self.lstm_input)
bilin_size = seq_len * self.n_filters
self.Bilin = nn.Bilinear(bilin_size,bilin_size,36)
self.output_layer = nn.Sequential(
nn.Dropout(p=0.5), # explained later
nn.Linear(36, 1)
)
def __init__(self, config):
super(StructuredAttention, self).__init__()
# self.use_gpu = config.use_gpu
#self.device = torch.device("cuda" if config.use_gpu else "cpu")
self.device = config.device
self.bidirectional = config.rnn_bidir
self.sem_dim_size = None
if config.rnn_bidir:
self.sem_dim_size = 2 * config.sem_dim_size
else:
self.sem_dim_size = config.sem_dim_size
self.rnn_cell_size = config.rnn_cell_size
if self.bidirectional:
self.rnn_cell_size = self.rnn_cell_size * 2
self.str_dim_size = self.rnn_cell_size - self.sem_dim_size
self.pytorch_version = "stable"
# print("Setting pytorch "+self.pytorch_version+" version for Structured Attention")
self.tp_linear = nn.Linear(self.str_dim_size, self.str_dim_size, bias=True)
torch.nn.init.xavier_uniform_(self.tp_linear.weight)
nn.init.constant_(self.tp_linear.bias, 0)
self.tc_linear = nn.Linear(self.str_dim_size, self.str_dim_size, bias=True)
torch.nn.init.xavier_uniform_(self.tc_linear.weight)
nn.init.constant_(self.tc_linear.bias, 0)
self.fi_linear = nn.Linear(self.str_dim_size, 1, bias=False)
torch.nn.init.xavier_uniform_(self.fi_linear.weight)
self.bilinear = nn.Bilinear(self.str_dim_size, self.str_dim_size, 1, bias=False)
torch.nn.init.xavier_uniform_(self.bilinear.weight)
self.exparam = nn.Parameter(torch.Tensor(1,1,self.sem_dim_size))
torch.nn.init.xavier_uniform_(self.exparam)
self.fzlinear = nn.Linear(3*self.sem_dim_size, self.sem_dim_size, bias=True)
torch.nn.init.xavier_uniform_(self.fzlinear.weight)
nn.init.constant_(self.fzlinear.bias, 0)
self.tanh = nn.Tanh()
self.relu = nn.ReLU()
self.leak_relu = nn.LeakyReLU()
def __init__(self,
dim_embeddings,
dim_rnn=128,
num_layers=2,
dropout_rate=0,
similarity='inner',
pooling='avg'):
super(RnnAttentionNet, self).__init__()
self.dim_embeddings = dim_embeddings
self.dim_rnn = dim_rnn
self.dim_encoded = 2 * dim_rnn
self.num_layers = num_layers
self.dropout_rate = dropout_rate
self.similarity = similarity
self.pooling = pooling
self.rnn_context = nn.GRU(input_size=dim_embeddings,
hidden_size=dim_rnn,
num_layers=num_layers,
batch_first=True,
dropout=dropout_rate,
bidirectional=True)
# self.rnn_option = nn.GRU(input_size=dim_embeddings, hidden_size=dim_rnn,
# num_layers=num_layers, batch_first=True,
# dropout=dropout_rate, bidirectional=True)
self.rnn_attn_context = nn.GRU(input_size=4 * self.dim_encoded,
hidden_size=dim_rnn,
num_layers=1,
batch_first=True,
dropout=dropout_rate,
bidirectional=True)
# self.rnn_attn_option = nn.GRU(input_size=4*self.dim_encoded, hidden_size=dim_rnn,
# num_layers=num_layers, batch_first=True,
# dropout=dropout_rate, bidirectional=True)
if self.similarity == 'cosine' or self.similarity == 'inner':
self.bi_fc = nn.Bilinear(self.dim_encoded, self.dim_encoded, 1)
elif self.similarity == 'trilinear':
self.co_attn = COAttention(d_model=self.dim_encoded,
dropout=dropout_rate)
self.fc_co = nn.Linear(self.dim_encoded, 1)
else:
raise ValueError(f"Invalid Similarity: {self.similarity}")
def __init__(self, opts): self.vocab_size = opts.vocab_size # default 10000 self.r = opts.r # 1, 2 super(GRN16, self).__init__() self.emb = nn.Embedding( self.vocab_size, 50 ) self.blstm = nn.LSTM( 50, 50, batch_first=True, bidirectional=True ) self.gate = nn.Linear(200, self.r ) self.H = nn.Bilinear( 100, 100, self.r, bias=False ) self.V = nn.Linear( 200, self.r, bias=False ) self.b = Parameter(torch.zeros(1, 2)) self.v = nn.Linear(self.r, 1) self.maxpool2d = nn.MaxPool2d([3, 3]) self.linear1 = nn.Linear(17 * 17, 50) # [50, 50] --> (3, 3) maxpool --> [17, 17] self.linear2 = nn.Linear(50, 2)
def __init__(self,
dim_in: int,
dropout: float,
score_type: str = 'dot') -> None:
super(Score_Net, self).__init__()
self.dim_in = dim_in
self.dropout = dropout
self.score_type = score_type
self.Dropout = nn.Dropout(dropout)
self.head = nn.Parameter(T(1, dim_in))
self.tail = nn.Parameter(T(1, dim_in))
self.layernorm = nn.LayerNorm(dim_in)
if score_type == 'bilinear':
self.func: Callable[[T, T], T] = nn.Bilinear(dim_in, dim_in, 1)
else:
self.func: Callable[[T, T], T] = torch.bmm
self.init_para()
def __init__(self,
dim_hid: int,
score_type: str,
bidirectional: bool = False) -> None:
super(Predic_Net, self).__init__()
self.dim_hid = dim_hid
self.score_type = score_type
self.bidirectional = bidirectional
if score_type == 'bilinear':
self.func: Callable[[T, T], T] = nn.Bilinear(dim_hid, dim_hid, 1)
elif score_type == 'dot':
self.func: Callable[[T, T], T] = torch.bmm
elif score_type == 'denselinear':
self.func = nn.Linear(4 * dim_hid, 2)
elif score_type == 'linear':
self.func = nn.Linear(2 * dim_hid, 2)
self.init_para()
self.norm_factor = np.sqrt(dim_hid)
self.layernorm = nn.LayerNorm(dim_hid)
def __init__(self, config, embedding):
super(rc_cnn_dailmail, self).__init__()
self.dict_embedding = nn.Embedding(num_embeddings=config.dict_num,
embedding_dim=100,
_weight=embedding)
# self.dict_embedding.weight.requires_grad = False
self.bilinear = nn.Bilinear(config.hidden_size * 2,
config.hidden_size * 2, 1)
self.lstm1 = nn.GRU(config.input_size,
config.hidden_size,
bidirectional=True,
batch_first=True,
dropout=config.dropout)
self.lstm2 = nn.GRU(config.input_size,
config.hidden_size,
bidirectional=True,
batch_first=True,
dropout=config.dropout)
self.linear = nn.Linear(config.hidden_size * 2, config.eneity_num)
def __init__(self, config):
super(CNNBaseline, self).__init__(config)
self.bert = BertModel(config)
self.init_weights()
self._embedding = self.bert.embeddings.word_embeddings
filter_sizes = [2, 3, 4, 5]
num_filters = 36
embedding_size = 768
self._convs = nn.ModuleList([
nn.Conv2d(1, num_filters, (K, embedding_size))
for K in filter_sizes
])
self._dropout = nn.Dropout(0.1)
self._linear = nn.Bilinear(
len(filter_sizes) * num_filters,
len(filter_sizes) * num_filters, 1)
self.apply(self.init_esim_weights)
def __init__(self, vocab, vocab_size, hidden_size, dropout, slots,
gating_dict, shared_emb):
super(Generator, self).__init__()
self.slots = slots
self.vocab = vocab
self.vocab_size = vocab_size
self.nb_gate = len(gating_dict)
self.hidden_size = hidden_size
self.gating_dict = gating_dict
self.embedding = shared_emb # token embedding matrix shared with encoders
self.dropout_layer = nn.Dropout(dropout)
self.bilinear = nn.Bilinear(self.hidden_size, self.hidden_size, 1)
self.gru = nn.GRU(hidden_size, hidden_size, dropout=dropout)
self.W_ratio = nn.Linear(3 * hidden_size, 1) # W_1
self.softmax = nn.Softmax(dim=1)
self.sigmoid = nn.Sigmoid()
def __init__(self,input_dim1,input_dim2,output_dim,
tauM = 20,tauAdp_inital =100, tau_initializer = 'normal',tauM_inital_std = 5,tauAdp_inital_std = 5,
is_adaptive=1,device='cpu'):
super(spike_Bidense, self).__init__()
self.input_dim1 = input_dim1
self.input_dim2 = input_dim2
self.output_dim = output_dim
self.is_adaptive = is_adaptive
self.device = device
self.dense = nn.Bilinear(input_dim1,input_dim2,output_dim)
self.tau_m = nn.Parameter(torch.Tensor(self.output_dim))
self.tau_adp = nn.Parameter(torch.Tensor(self.output_dim))
if tau_initializer == 'normal':
nn.init.normal_(self.tau_m,tauM,tauM_inital_std)
nn.init.normal_(self.tau_adp,tauAdp_inital,tauAdp_inital_std)
elif tau_initializer == 'multi_normal':
self.tau_m = multi_normal_initilization(self.tau_m,tauM,tauM_inital_std)
self.tau_adp = multi_normal_initilization(self.tau_adp,tauAdp_inital,tauAdp_inital_std)
def __init__(self,
n_in,
n_hid,
n_out,
do_prob=0.,
bilinear=False,
bnorm=True):
super(MLP, self).__init__()
self.bilinear = bilinear
self.bnorm = bnorm
if bilinear:
self.fc1 = nn.Bilinear(n_in, n_in, n_hid)
else:
self.fc1 = nn.Linear(n_in, n_hid)
self.fc2 = nn.Linear(n_hid, n_out)
if bnorm:
self.bn = nn.BatchNorm1d(n_out)
self.dropout_prob = do_prob
self.init_weights()
def test_metabilinear_params(bias):
meta_model = MetaBilinear(2, 3, 5, bias=bias)
model = nn.Bilinear(2, 3, 5, bias=bias)
params = OrderedDict()
params['weight'] = torch.randn(5, 2, 3)
model.weight.data.copy_(params['weight'])
if bias:
params['bias'] = torch.randn(5)
model.bias.data.copy_(params['bias'])
inputs1 = torch.randn(7, 2)
inputs2 = torch.randn(7, 3)
outputs_torchmeta = meta_model(inputs1, inputs2, params=params)
outputs_nn = model(inputs1, inputs2)
np.testing.assert_equal(outputs_torchmeta.detach().numpy(),
outputs_nn.detach().numpy())
def __init__(self,
latent_dim,
hidden_dim,
n_out=1,
num_layers=1,
activation=nn.Tanh,
softplus=False,
resid=False,
expand_coords=False,
bilinear=False):
super(SpatialGenerator, self).__init__()
self.softplus = softplus
self.expand_coords = expand_coords
in_dim = 2
if expand_coords:
in_dim = 5 # include squares of coordinates as inputs
self.coord_linear = nn.Linear(in_dim, hidden_dim)
self.latent_dim = latent_dim
if latent_dim > 0:
self.latent_linear = nn.Linear(latent_dim, hidden_dim, bias=False)
if latent_dim > 0 and bilinear: # include bilinear layer on latent and coordinates
self.bilinear = nn.Bilinear(in_dim,
latent_dim,
hidden_dim,
bias=False)
layers = [activation()]
for _ in range(1, num_layers):
if resid:
layers.append(
ResidLinear(hidden_dim, hidden_dim, activation=activation))
else:
layers.append(nn.Linear(hidden_dim, hidden_dim))
layers.append(activation())
layers.append(nn.Linear(hidden_dim, n_out))
self.layers = nn.Sequential(*layers)
def __init__(self, input_dims, action_dims):
super(M1pM1, self).__init__()
self.input_dims = input_dims
self.action_dims = action_dims
self.seq = nn.Sequential(
nn.Linear(input_dims * 2 + action_dims, input_dims * 4),
cheese_layer(), nn.Linear(input_dims * 4, input_dims * 4),
cheese_layer(), nn.Linear(input_dims * 4, input_dims * 4),
cheese_layer(), nn.Linear(input_dims * 4, input_dims))
self.seq2 = nn.Sequential(nn.Linear(input_dims, input_dims * 4),
cheese_layer(),
nn.Linear(input_dims * 4, input_dims * 4),
cheese_layer(),
nn.Linear(input_dims * 4, input_dims * 4),
cheese_layer(),
nn.Linear(input_dims * 4, input_dims))
self.bi2 = nn.Bilinear(input_dims, input_dims, input_dims)
def __init__(self,
vocab_size,
emb_size,
hidden_size,
dropout_rate,
tied_embedding=None,
enc_attention=False):
super(DecoderRNN, self).__init__()
self.GRU = nn.GRU(emb_size, hidden_size, dropout=dropout_rate)
output_size = hidden_size
self.enc_attention = enc_attention
if enc_attention:
self.bilinear = nn.Bilinear(hidden_size, hidden_size, 1)
output_size += hidden_size
self.pre_out = nn.Linear(output_size, emb_size)
self.out_layer = nn.Linear(emb_size, vocab_size)
# Tied embedding means the word embedding layer weights are shared in encoder and decoder
if tied_embedding is not None:
self.out_layer.weight = tied_embedding.weight