oucgc1996/CreoPep_conditional_generation

1import torch.nn as nn2import copy, math3import torch4import numpy as np5import torch.nn.functional as F6 7class Bert(nn.Module):8    9    def __init__(self, encoder, src_embed):10        super(Bert, self).__init__()11 12        self.encoder = encoder        13        self.src_embed = src_embed  14    15    def forward(self, src, src_mask):16    17        return self.encoder(self.src_embed(src), src_mask)18    19    20class Encoder(nn.Module):21    def __init__(self, layer, N):22        super(Encoder, self).__init__()23        self.layers = clones(layer, N)24        self.norm = LayerNorm(layer.size)25        26    def forward(self, x, mask):27        for layer in self.layers:28            x = layer(x, mask)29        return self.norm(x)30    31class LayerNorm(nn.Module):32    def __init__(self, features, eps=1e-6):33        super(LayerNorm, self).__init__()34        self.a_2 = nn.Parameter(torch.ones(features))35        self.b_2 = nn.Parameter(torch.zeros(features))36        self.eps = eps37 38    def forward(self, x):39        mean = x.mean(-1, keepdim=True)40        std = x.std(-1, keepdim=True)41        return self.a_2 * (x - mean) / (std + self.eps) + self.b_242 43class SublayerConnection(nn.Module):44    def __init__(self, size, dropout):45        super(SublayerConnection, self).__init__()46        self.norm = LayerNorm(size)47        self.dropout = nn.Dropout(dropout)48 49    def forward(self, x, sublayer):50        return x + self.dropout(sublayer(self.norm(x)))51    52class EncoderLayer(nn.Module):53    def __init__(self, size, self_attn, feed_forward, dropout):54        super(EncoderLayer, self).__init__()55        self.self_attn = self_attn56        self.feed_forward = feed_forward57        self.sublayer = clones(SublayerConnection(size, dropout), 2)58        self.size = size59 60    def forward(self, x, mask):61        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))62        return self.sublayer[1](x, self.feed_forward)63    64class PositionwiseFeedForward(nn.Module):65    def __init__(self, d_model, d_ff, dropout=0.1):66        super(PositionwiseFeedForward, self).__init__()67        self.w_1 = nn.Linear(d_model, d_ff)68        self.w_2 = nn.Linear(d_ff, d_model)69        self.dropout = nn.Dropout(dropout)70 71    def forward(self, x):72        return self.w_2(self.dropout(F.relu(self.w_1(x))))73    74def make_bert(src_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1):75    c = copy.deepcopy76    attn = MultiHeadedAttention(h, d_model)77    ff = PositionwiseFeedForward(d_model, d_ff, dropout)78    position = PositionalEncoding(d_model, dropout)79    model = Bert(80        Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),81        82        nn.Sequential(Embeddings(d_model, src_vocab), c(position)),83    )84    85    for p in model.parameters():86        if p.dim() > 1:87            nn.init.xavier_uniform_(p)88    return model89 90def make_bert_without_emb(d_model=128, N=2, d_ff=512, h=8, dropout=0.1):91    c = copy.deepcopy92    attn = MultiHeadedAttention(h, d_model)93    ff = PositionwiseFeedForward(d_model, d_ff, dropout)94    trainable_encoder = Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N)95 96    return trainable_encoder97 98 99 100def clones(module, N):101    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])102 103def subsequent_mask(size):104    attn_shape = (1, size, size)105    subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')106    return torch.from_numpy(subsequent_mask) == 0107 108def attention(query, key, value, mask=None, dropout=None):109    d_k = query.size(-1)110    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)111    if mask is not None:112        mask = mask.unsqueeze(-2)113        scores = scores.masked_fill(mask == 0, -1e9)114    p_attn = F.softmax(scores, dim = -1)115    if dropout is not None:116        p_attn = dropout(p_attn)117    return torch.matmul(p_attn, value), p_attn118 119class MultiHeadedAttention(nn.Module):120    def __init__(self, h, d_model, dropout=0.1):121        super(MultiHeadedAttention, self).__init__()122        assert d_model % h == 0123        self.d_k = d_model // h124        self.h = h125        self.linears = clones(nn.Linear(d_model, d_model), 4)126        self.attn = None127        self.dropout = nn.Dropout(p=dropout)128        129    def forward(self, query, key, value, mask=None):130        if mask is not None:131            mask = mask.unsqueeze(1)132        nbatches = query.size(0)133          134        query, key, value = \135            [l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)136             for l, x in zip(self.linears, (query, key, value))]137        138        x, self.attn = attention(query, key, value, mask=mask, 139                                 dropout=self.dropout)140        141        x = x.transpose(1, 2).contiguous() \142             .view(nbatches, -1, self.h * self.d_k)143        return self.linears[-1](x)144    145class Embeddings(nn.Module):146    def __init__(self, d_model, vocab):147        super(Embeddings, self).__init__()148        self.lut = nn.Embedding(vocab, d_model)149        self.d_model = d_model150 151    def forward(self, x):152        return self.lut(x) * math.sqrt(self.d_model)153    154class PositionalEncoding(nn.Module):155    def __init__(self, d_model, dropout, max_len=5000):156        super(PositionalEncoding, self).__init__()157        self.dropout = nn.Dropout(p=dropout)158        159        pe = torch.zeros(max_len, d_model)160        position = torch.arange(0, max_len).unsqueeze(1)161        div_term = torch.exp(torch.arange(0, d_model, 2) *162                             -(math.log(10000.0) / d_model))163        pe[:, 0::2] = torch.sin(position * div_term)164        pe[:, 1::2] = torch.cos(position * div_term)165        pe = pe.unsqueeze(0)166        self.register_buffer('pe', pe)167        168    def forward(self, x):169        x = x + self.pe[:, :x.size(1)].clone().detach()170        return self.dropout(x)171    172