avse_bert/encoder.py

import torch.nn.functional as F
import math
import torch
import torch.nn as nn
import torchtext
import numpy as np
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
import logging
import random
from transformers import BertModel

logger = logging.getLogger(__name__)


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x

class Attention(nn.Module):
    def __init__(self,
                 dim,
                 num_heads,
                 out_dim=None,
                 qkv_bias=False,
                 qk_scale=None,
                 attn_drop=0.,
                 proj_drop=0.):
        super().__init__()
        if out_dim is None:
            out_dim = dim
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim**-0.5

        self.q_proj = nn.Linear(dim, dim, bias=qkv_bias)
        self.k_proj = nn.Linear(dim, dim, bias=qkv_bias)
        self.v_proj = nn.Linear(dim, dim, bias=qkv_bias)
        
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, out_dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, query, key=None, *, value=None):
        B, N, C = query.shape
        if key is None:
            key = query
        if value is None:
            value = key
        S = key.size(1)
        # [B, nh, N, C//nh]
        q = self.q_proj(query)
        q = q.reshape(B,N,1, self.num_heads, C // self.num_heads).permute(2,0,3,1,4).squeeze(0)
        
        k = self.k_proj(key)
        k = k.reshape(B,k.shape[1],1, self.num_heads, C // self.num_heads).permute(2,0,3,1,4).squeeze(0)
        
        v = self.v_proj(value)
        v = v.reshape(B,v.shape[1],1, self.num_heads, C // self.num_heads).permute(2,0,3,1,4).squeeze(0)
        
        # [B, nh, N, S]
        attn = (q @ k.transpose(-2, -1)) * self.scale
        
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        out = (attn @ v).transpose(1, 2).reshape(B, N, -1)
        out = self.proj(out)
        out = self.proj_drop(out)
        return out

class CrossAttnBlock(nn.Module):
    def __init__(self,
                 dim,
                 num_heads,
                 mlp_ratio=4.,
                 qkv_bias=False,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 act_layer=nn.GELU,
                 norm_layer=nn.LayerNorm,
                 post_norm=False):
        super().__init__()
        self.norm_q = norm_layer(dim)
        self.norm_k = norm_layer(dim)

        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, query, key, *, mask=None):
        x = query
        x = x + self.drop_path(self.attn(self.norm_q(query), self.norm_k(key)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x


class Grouping(nn.Module):
    def __init__(self, n_group, img_dim):
        super(Grouping, self).__init__()

        self.fc = nn.Linear(img_dim, 768)
        self.init_weights()
        embed_size = 768
        # GSR
        self.group = CrossAttnBlock(
            dim=embed_size,
            num_heads=6,
            mlp_ratio=4,
            qkv_bias=True,
        )
    
        self.group_token = nn.Parameter(torch.randn(n_group, embed_size))
        self.group_proj = nn.Linear(embed_size, img_dim)

    def init_weights(self):
        """Xavier initialization for the fully connected layer
        """
        r = np.sqrt(6.) / np.sqrt(self.fc.in_features +
                                  self.fc.out_features)
        self.fc.weight.data.uniform_(-r, r)
        self.fc.bias.data.fill_(0)

    def forward(self, images, n_group=4, n_region=30):
        fc_img_emd = self.fc(images)

        # GCN reasoning
        # -> B,D,N
        group_tokens = self.group_token.expand(images.shape[0], -1, -1)
        
        grouped = self.group_proj(self.group(group_tokens, fc_img_emd))
        
        grouped = l2norm(grouped, dim=-1)
        features = l2norm(images, dim=-1)

        A = (grouped @ features.transpose(-1, -2))
        att = A.softmax(dim=-1) + (1e-8)

        n = images.shape[1]
        b = images.shape[0]
      
        att_sort, sort_ind = att.sort(-1, True)
        att_mask = torch.zeros_like(sort_ind)
#         att_mask_tmp = att_mask[]
        if self.training:
            for i in range(n_group):
                j = random.randint(int(0.9*n) , n)
                att_mask[:,i,:j] = 1
        else:
            att_mask[:,:,:n_region] = 1
            
        _, sort_ind2 = sort_ind.sort()
        att_mask_ = att_mask.gather(-1, sort_ind2).bool()
#         print(att_mask.shape)
#         print(att_mask_[:,0 ].sum())
        grouped_images = []
        grouped_lengths = []
        for i in range(n_group):
            base_features = images.masked_select(att_mask_[:,i].unsqueeze(-1)).reshape(b, -1, images.shape[-1])
            
            
            feat_lengths = torch.zeros(base_features.size(0)).to(base_features.device)
            feat_lengths[:] = base_features.size(1)
            
            grouped_images.append(base_features)
            grouped_lengths.append(feat_lengths)
    
        return grouped_images, grouped_lengths



class Rs_GCN(nn.Module):

    def __init__(self, in_channels, inter_channels, bn_layer=True):
        super(Rs_GCN, self).__init__()

        self.in_channels = in_channels
        self.inter_channels = inter_channels

        if self.inter_channels is None:
            self.inter_channels = in_channels // 2
            if self.inter_channels == 0:
                self.inter_channels = 1

        conv_nd = nn.Conv1d
        max_pool = nn.MaxPool1d
        bn = nn.BatchNorm1d

        self.g = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels,
                         kernel_size=1, stride=1, padding=0)

        if bn_layer:
            self.W = nn.Sequential(
                conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels,
                        kernel_size=1, stride=1, padding=0),
                bn(self.in_channels)
            )
            nn.init.constant(self.W[1].weight, 0)
            nn.init.constant(self.W[1].bias, 0)
        else:
            self.W = conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels,
                             kernel_size=1, stride=1, padding=0)
            nn.init.constant(self.W.weight, 0)
            nn.init.constant(self.W.bias, 0)

        self.theta = None
        self.phi = None

        self.theta = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels,
                             kernel_size=1, stride=1, padding=0)
        self.phi = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels,
                           kernel_size=1, stride=1, padding=0)

    def forward(self, v):
        '''
        :param v: (B, D, N)
        :return:
        '''
        batch_size = v.size(0)

        g_v = self.g(v).view(batch_size, self.inter_channels, -1)
        g_v = g_v.permute(0, 2, 1)

        theta_v = self.theta(v).view(batch_size, self.inter_channels, -1)
        theta_v = theta_v.permute(0, 2, 1)
        phi_v = self.phi(v).view(batch_size, self.inter_channels, -1)
        R = torch.matmul(theta_v, phi_v)
        N = R.size(-1)
        R_div_C = R / N

        y = torch.matmul(R_div_C, g_v)
        y = y.permute(0, 2, 1).contiguous()
        y = y.view(batch_size, self.inter_channels, *v.size()[2:])
        W_y = self.W(y)
        v_star = W_y + v

        return v_star


class MLP(nn.Module):
    """ Very simple multi-layer perceptron (also called FFN)"""

    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
        super().__init__()
        self.output_dim = output_dim
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
        self.bns = nn.ModuleList(nn.BatchNorm1d(k) for k in h + [output_dim])

    def forward(self, x):
        B, N, D = x.size()
        x = x.reshape(B*N, D)
        for i, (bn, layer) in enumerate(zip(self.bns, self.layers)):
            x = F.relu(bn(layer(x))) if i < self.num_layers - 1 else layer(x)
        x = x.view(B, N, self.output_dim)
        return x


def positional_encoding_1d(d_model, length):
    """
    :param d_model: dimension of the model
    :param length: length of positions
    :return: length*d_model position matrix
    """
    if d_model % 2 != 0:
        raise ValueError("Cannot use sin/cos positional encoding with "
                         "odd dim (got dim={:d})".format(d_model))
    pe = torch.zeros(length, d_model)
    position = torch.arange(0, length).unsqueeze(1)
    div_term = torch.exp((torch.arange(0, d_model, 2, dtype=torch.float) *
                          -(math.log(10000.0) / d_model)))
    pe[:, 0::2] = torch.sin(position.float() * div_term)
    pe[:, 1::2] = torch.cos(position.float() * div_term)

    return pe


class GPO(nn.Module):
    def __init__(self, d_pe, d_hidden):
        super(GPO, self).__init__()
        self.d_pe = d_pe
        self.d_hidden = d_hidden

        self.pe_database = {}
        self.gru = nn.GRU(self.d_pe, d_hidden, 1, batch_first=True, bidirectional=True)
        self.linear = nn.Linear(self.d_hidden, 1, bias=False)

    def compute_pool_weights(self, lengths, features):
        max_len = int(lengths.max())
        pe_max_len = self.get_pe(max_len)
        pes = pe_max_len.unsqueeze(0).repeat(lengths.size(0), 1, 1).to(lengths.device)
        mask = torch.arange(max_len).expand(lengths.size(0), max_len).to(lengths.device)
        mask = (mask < lengths.long().unsqueeze(1)).unsqueeze(-1)
        pes = pes.masked_fill(mask == 0, 0)

        self.gru.flatten_parameters()
        packed = pack_padded_sequence(pes, lengths.cpu(), batch_first=True, enforce_sorted=False)
        out, _ = self.gru(packed)
        padded = pad_packed_sequence(out, batch_first=True)
        out_emb, out_len = padded
        out_emb = (out_emb[:, :, :out_emb.size(2) // 2] + out_emb[:, :, out_emb.size(2) // 2:]) / 2
        scores = self.linear(out_emb)
        scores[torch.where(mask == 0)] = -10000

        weights = torch.softmax(scores / 0.1, 1)
        return weights, mask

    def forward(self, features, lengths):
        """
        :param features: features with shape B x K x D
        :param lengths: B x 1, specify the length of each data sample.
        :return: pooled feature with shape B x D
        """
        pool_weights, mask = self.compute_pool_weights(lengths, features)

        features = features[:, :int(lengths.max()), :]
        sorted_features = features.masked_fill(mask == 0, -10000)
        sorted_features,  sort = sorted_features.sort(dim=1, descending=True)

        sorted_features = sorted_features.masked_fill(mask == 0, 0)

        pooled_features = (sorted_features * pool_weights).sum(1)
        return pooled_features, pool_weights

    def get_pe(self, length):
        """

        :param length: the length of the sequence
        :return: the positional encoding of the given length
        """
        length = int(length)
        if length in self.pe_database:
            return self.pe_database[length]
        else:
            pe = positional_encoding_1d(self.d_pe, length)
            self.pe_database[length] = pe
            return pe


def l1norm(X, dim, eps=1e-8):
    """L1-normalize columns of X
    """
    norm = torch.abs(X).sum(dim=dim, keepdim=True) + eps
    X = torch.div(X, norm)
    return X


def l2norm(X, dim, eps=1e-8):
    """L2-normalize columns of X
    """
    norm = torch.pow(X, 2).sum(dim=dim, keepdim=True).sqrt() + eps
    X = torch.div(X, norm)
    return X


def maxk_pool1d_var(x, dim, k, lengths):
    results = list()
    lengths = list(lengths.cpu().numpy())
    lengths = [int(x) for x in lengths]
    for idx, length in enumerate(lengths):
        k = min(k, length)
        max_k_i = maxk(x[idx, :length, :], dim - 1, k).mean(dim - 1)
        results.append(max_k_i)
    results = torch.stack(results, dim=0)
    return results


def maxk_pool1d(x, dim, k):
    max_k = maxk(x, dim, k)
    return max_k.mean(dim)


def maxk(x, dim, k):
    index = x.topk(k, dim=dim)[1]
    return x.gather(dim, index)


def get_text_encoder(vocab_size, embed_size, word_dim, num_layers, opt, use_bi_gru=True, no_txtnorm=False):
    return EncoderText(vocab_size, embed_size, word_dim, num_layers, opt, use_bi_gru=use_bi_gru,
                       no_txtnorm=no_txtnorm)


def get_sim_encoder(opt):
    return EncoderSimilarityGraph(opt)


def get_image_encoder(img_dim, embed_size, precomp_enc_type='basic',
                      backbone_source=None, backbone_path=None, no_imgnorm=False):
    """A wrapper to image encoders. Chooses between an different encoders
    that uses precomputed image features.
    """
    if precomp_enc_type == 'basic':
        img_enc = EncoderImageAggr(
            img_dim, embed_size, precomp_enc_type, no_imgnorm)
    elif precomp_enc_type == 'backbone':
        backbone_cnn = ResnetFeatureExtractor(backbone_source, backbone_path, fixed_blocks=2)
        img_enc = EncoderImageFull(backbone_cnn, img_dim, embed_size, precomp_enc_type, no_imgnorm)
    else:
        raise ValueError("Unknown precomp_enc_type: {}".format(precomp_enc_type))

    return img_enc


class EncoderImageAggr(nn.Module):
    def __init__(self, img_dim, embed_size, precomp_enc_type='basic', no_imgnorm=False):
        super(EncoderImageAggr, self).__init__()
        # embed_size = embed_size * 2
        self.embed_size = embed_size
        self.no_imgnorm = no_imgnorm
        self.dropout = nn.Dropout(0.2)
        self.fc = nn.Linear(img_dim, embed_size)

        self.precomp_enc_type = precomp_enc_type
        if precomp_enc_type == 'basic':
            self.mlp = MLP(img_dim, embed_size // 2, embed_size, 2)
        self.gpool = GPO(32, 32)
        self.init_weights()

    def init_weights(self):
        """Xavier initialization for the fully connected layer
        """
        r = np.sqrt(6.) / np.sqrt(self.fc.in_features +
                                  self.fc.out_features)
        self.fc.weight.data.uniform_(-r, r)
        self.fc.bias.data.fill_(0)



    def forward(self, images, image_lengths):
        """Extract image feature vectors."""
        # if torch.cuda.is_available():
        #   img_p1 = images[0].cuda()
        #   img_p2 = images[1].cuda()

        #   img_len1 = image_lengths[0].cuda()
        #   img_len2 = image_lengths[1].cuda()

        images = self.dropout(images)
        # img_p2 = self.dropout(img_p2)

        reg_emb = self.fc(images)
        # reg_emb_p2 = self.fc2(img_p2)
        if self.precomp_enc_type == 'basic':
            # When using pre-extracted region features, add an extra MLP for embedding transformation
            reg_emb = self.mlp(images) + reg_emb
            # reg_emb_p2 = self.mlp2(img_p2) + reg_emb_p2

        img_emb, _ = self.gpool(reg_emb, image_lengths)
        # img_emb_2, _ = self.gpool2(reg_emb_p2, img_len2)
        # features = features.mean(dim=1)

        # img_emb = torch.cat([img_emb_1, img_emb_2], dim=1)
        if not self.no_imgnorm:
            img_emb = l2norm(img_emb, dim=-1)
            # reg_emb = l2norm(reg_emb, dim=-1)

        return img_emb


class EncoderImageFull(nn.Module):
    def __init__(self, backbone_cnn, img_dim, embed_size, precomp_enc_type='basic', no_imgnorm=False):
        super(EncoderImageFull, self).__init__()
        self.backbone = backbone_cnn
        self.image_encoder = EncoderImageAggr(img_dim, embed_size, precomp_enc_type, no_imgnorm)
        self.backbone_freezed = False

    def forward(self, images):
        """Extract image feature vectors."""
        base_features = self.backbone(images)

        if self.training:
            # Size Augmentation during training, randomly drop grids
            base_length = base_features.size(1)
            features = []
            feat_lengths = []
            rand_list_1 = np.random.rand(base_features.size(0), base_features.size(1))
            rand_list_2 = np.random.rand(base_features.size(0))
            for i in range(base_features.size(0)):
                if rand_list_2[i] > 0.2:
                    feat_i = base_features[i][np.where(rand_list_1[i] > 0.20 * rand_list_2[i])]
                    len_i = len(feat_i)
                    pads_i = torch.zeros(base_length - len_i, base_features.size(-1)).to(base_features.device)
                    feat_i = torch.cat([feat_i, pads_i], dim=0)
                else:
                    feat_i = base_features[i]
                    len_i = base_length
                feat_lengths.append(len_i)
                features.append(feat_i)
            base_features = torch.stack(features, dim=0)
            base_features = base_features[:, :max(feat_lengths), :]
            feat_lengths = torch.tensor(feat_lengths).to(base_features.device)
        else:
            feat_lengths = torch.zeros(base_features.size(0)).to(base_features.device)
            feat_lengths[:] = base_features.size(1)

        features = self.image_encoder(base_features, feat_lengths)

        return features

    def freeze_backbone(self):
        for param in self.backbone.parameters():
            param.requires_grad = False
        print('Backbone freezed.')

    def unfreeze_backbone(self, fixed_blocks):
        for param in self.backbone.parameters():  # open up all params first, then adjust the base parameters
            param.requires_grad = True
        self.backbone.set_fixed_blocks(fixed_blocks)
        self.backbone.unfreeze_base()
        print('Backbone unfreezed, fixed blocks {}'.format(self.backbone.get_fixed_blocks()))


# Language Model with BiGRU
class EncoderText(nn.Module):
    def __init__(self, vocab_size, embed_size, word_dim, num_layers, opt, use_bi_gru=True, no_txtnorm=False):
        super(EncoderText, self).__init__()
        self.opt = opt
        self.embed_size = embed_size
        self.no_txtnorm = no_txtnorm
        # word embedding
#         self.embed = nn.Embedding(vocab_size, word_dim)

        self.bert = BertModel.from_pretrained('bert-base-uncased')
#         for p in self.bert.parameters():
#             p.requires_grad=False
        # self.embed = bert.em

        self.linear = nn.Linear(768, embed_size)

        self.dropout = nn.Dropout(0.2)
        self.gpool = GPO(32, 32)
        self.ln = nn.LayerNorm(embed_size)
        # caption embedding

        embed_size1 = embed_size
        self.fc = nn.Linear(embed_size, embed_size1)
        self.mlp = MLP(768, embed_size1 // 2, embed_size1, 2)

    def forward(self, x, lengths):
        """Handles variable size captions
        """
        # Embed word ids to vectors
        # x_emb = self.embed(x)
        # x_emb = self.dropout(x_emb)
        # self.rnn.flatten_parameters()
        # packed = pack_padded_sequence(x_emb, lengths.cpu(), batch_first=True)
        #
        # # Forward propagate RNN
        # out, _ = self.rnn(packed)

        bert_attention_mask = (x != 0).float()
        bert_emb = self.bert(x, bert_attention_mask)[0]  # B x N x D
        
        bert_emb = self.dropout(bert_emb)
        
        cap_len = lengths
        cap_emb = self.linear(bert_emb)

        # features = self.fc(cap_emb)
        features = self.mlp(bert_emb) + cap_emb

        pooled_features, pool_weights = self.gpool(features, cap_len.to(features.device))

        pooled_features = self.ln(pooled_features)

        return pooled_features


def log_sinkhorn_iterations(Z: torch.Tensor, log_mu: torch.Tensor, log_nu: torch.Tensor, iters: int) -> torch.Tensor:
    """ Perform Sinkhorn Normalization in Log-space for stability"""
    u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu)
    for _ in range(iters):
        u = log_mu - torch.logsumexp(Z + v.unsqueeze(1), dim=2)
        v = log_nu - torch.logsumexp(Z + u.unsqueeze(2), dim=1)
    return Z + u.unsqueeze(2) + v.unsqueeze(1)


def log_optimal_transport(scores: torch.Tensor, alpha: torch.Tensor, iters: int) -> torch.Tensor:
    """ Perform Differentiable Optimal Transport in Log-space for stability"""
    b, m, n = scores.shape
    one = scores.new_tensor(1)
    ms, ns = (m * one).to(scores), (n * one).to(scores)

    bins0 = alpha.expand(b, m, 1)
    bins1 = alpha.expand(b, 1, n)
    alpha = alpha.expand(b, 1, 1)

    couplings = torch.cat([torch.cat([scores, bins0], -1),
                           torch.cat([bins1, alpha], -1)], 1)

    norm = - (ms + ns).log()
    log_mu = torch.cat([norm.expand(m), ns.log()[None] + norm])
    log_nu = torch.cat([norm.expand(n), ms.log()[None] + norm])
    log_mu, log_nu = log_mu[None].expand(b, -1), log_nu[None].expand(b, -1)

    Z = log_sinkhorn_iterations(couplings, log_mu, log_nu, iters)
    Z = Z - norm  # multiply probabilities by M+N
    return Z


class EncoderSimilarityGraph(nn.Module):
    def __init__(self, opt):
        super(EncoderSimilarityGraph, self).__init__()
        self.opt = opt
        self.block_dim = opt.block_dim
        self.alpha = torch.Tensor(opt.alpha).cuda()
        bin_score = torch.nn.Parameter(torch.tensor(0.))
        self.register_parameter('bin_score', bin_score)

    def forward(self, img_emb, cap_emb):
        cap_emb = l2norm(cap_emb, -1)
        n_cap, cap_dim = cap_emb.size(0), cap_emb.size(1)
        n_img, img_dim = img_emb.size(0), img_emb.size(1)
        sims = []
        for i, block_dim in enumerate(self.block_dim):
            img_blk_num, cap_blk_num = img_emb.size(1) // block_dim, cap_emb.size(1) // block_dim
            img_emb_blocks = torch.chunk(img_emb, img_blk_num, -1)  # (bs, 2*n, block_dim)
            cap_emb_blocks = torch.chunk(cap_emb, cap_blk_num, -1)  # (bs, n, block_dim)

            img_emb_blocks = torch.stack(img_emb_blocks, dim=1)  # (bs, 2*n, block_dim)
            cap_emb_blocks = torch.stack(cap_emb_blocks, dim=1)  # (bs, n, block_dim)


            img_emb_blocks = l2norm(img_emb_blocks, -1)  # (bs, 2*n, block_dim)
            cap_emb_blocks = l2norm(cap_emb_blocks, -1)

            logits = torch.einsum("avc,btc->abvt", [img_emb_blocks, cap_emb_blocks])  # (bs, bs, 2*n, n)

            logits = log_optimal_transport(logits.reshape(-1, img_blk_num, cap_blk_num), self.bin_score, 20)[:, :-1,
                     :-1].reshape(n_img, n_cap, img_blk_num, cap_blk_num)
            t2i_logits = logits.max(dim=-2)[0]
            sims.append(t2i_logits.sum(dim=-1))

        sims = torch.stack(sims, -1).sum(-1)

        return sims