论文阅读: （CVPR2023 SDT ）基于书写者风格和字符风格解耦的手写文字生成及源码对应

引言

• 许久不认真看论文了，这不赶紧捡起来。这也是自己看的第一篇用到Transformer结构的CV论文。
• 之所以选择这篇文章来看，是考虑到之前做过手写字体生成的项目。这个工作可以用来合成一些手写体数据集，用来辅助手写体识别模型的训练。
• 本篇文章将从论文与代码一一对应解析的方式来撰写，这样便于找到论文重点地方以及用代码如何实现的，更快地学到其中要点。这个项目的代码写得很好看，有着清晰的说明和整洁的代码规范。跟着仓库README就可以快速跑起整个项目。
• 如果读者可以阅读英文的话，建议先去直接阅读英文论文，会更直接看到整个面貌。
• PDF | Code

SDT整体结构介绍

• 整体框架：
  
• 该工作提出从个体手写中解耦作家和字符级别的风格表示，以合成逼真的风格化在线手写字符。
• 从上述框架图，可以看出整体可分为三大部分：Style encoder、Content Encoder和Transformer Decoder。
  • Style Encoder: 主要学习给定的Style的Writer和Glyph两种风格表示，用于指导合成风格化的文字。包含两部分：CNN Encoder和Transformer Encdoer。
  • Content Encoder: 主要提取输入文字的特征，同样包含两部分：CNN Encoder和Transformer Encdoer。
• ❓疑问：为什么要将CNN Encoder + Transformer Encoder结合使用呢？
  • 这个问题在论文中只说了Content Encoder使用两者的作用。CNN部分用来从content reference中学到compact feature map。Transformer encoder用来提取textual content表示。得益于Transformer强大的long-range 依赖的捕捉能力，Content Encdoer可以得到一个全局上下文的content feature。这里让我想到经典的CRNN结构，就是结合CNN + RNN两部分。
    

代码与论文对应

• 论文结构的最核心代码有两部分，一是搭建模型部分，二是数据集处理部分。

搭建模型部分

• 该部分代码位于仓库中models/model.py，我这里只摘其中最关键部分添加注释来解释，其余细节请小伙伴自行挖掘。

class SDT_Generator(nn.Module):def __init__(self, d_model=512, nhead=8, num_encoder_layers=2, num_head_layers= 1,wri_dec_layers=2, gly_dec_layers=2, dim_feedforward=2048, dropout=0.1,activation="relu", normalize_before=True, return_intermediate_dec=True):super(SDT_Generator, self).__init__()### style encoder with dual heads# Feat_Encoder：对应论文中的CNN Encoder，用来提取图像经过CNN之后的特征，backbone选的是ResNet18self.Feat_Encoder = nn.Sequential(*([nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)] +list(models.resnet18(pretrained=True).children())[1:-2]))# self.base_encoder：对应论文中Style Encoder的Transformer Encoderb部分encoder_layer = TransformerEncoderLayer(d_model, nhead, dim_feedforward,dropout, activation, normalize_before)self.base_encoder = TransformerEncoder(encoder_layer, num_encoder_layers, None)writer_norm = nn.LayerNorm(d_model) if normalize_before else Noneglyph_norm = nn.LayerNorm(d_model) if normalize_before else None# writer_head和glyph_head分别对应论文中的Writer Head和Glyph Head# 从这里来看，这两个分支使用的是1层的Transformer Encoder结构self.writer_head = TransformerEncoder(encoder_layer, num_head_layers, writer_norm)self.glyph_head = TransformerEncoder(encoder_layer, num_head_layers, glyph_norm)### content ecoder# content_encoder：对应论文中Content Encoder部分，# 从Content_TR源码来看，同样也是ResNet18作为CNN Encoder的backbone# Transformer Encoder部分用了3层的Transformer Encoder结构# 详情参见：https://github.com/dailenson/SDT/blob/1352b5cb779d47c5a8c87f6735e9dde94aa58f07/models/encoder.py#L8self.content_encoder = Content_TR(d_model, num_encoder_layers)### decoder for receiving writer-wise and character-wise styles# 这里对应框图中Transformer Decoder中前后两个部分decoder_layer = TransformerDecoderLayer(d_model, nhead, dim_feedforward,dropout, activation, normalize_before)wri_decoder_norm = nn.LayerNorm(d_model) if normalize_before else Noneself.wri_decoder = TransformerDecoder(decoder_layer, wri_dec_layers, wri_decoder_norm,return_intermediate=return_intermediate_dec)gly_decoder_norm = nn.LayerNorm(d_model) if normalize_before else Noneself.gly_decoder = TransformerDecoder(decoder_layer, gly_dec_layers, gly_decoder_norm,return_intermediate=return_intermediate_dec)### two mlps that project style features into the space where nce_loss is appliedself.pro_mlp_writer = nn.Sequential(nn.Linear(512, 4096), nn.GELU(), nn.Linear(4096, 256))self.pro_mlp_character = nn.Sequential(nn.Linear(512, 4096), nn.GELU(), nn.Linear(4096, 256))self.SeqtoEmb = SeqtoEmb(hid_dim=d_model)self.EmbtoSeq = EmbtoSeq(hid_dim=d_model)# 这里位置嵌入来源于论文Attention is all you need.self.add_position = PositionalEncoding(dropout=0.1, dim=d_model)        self._reset_parameters()# the shape of style_imgs is [B, 2*N, C, H, W] during trainingdef forward(self, style_imgs, seq, char_img):batch_size, num_imgs, in_planes, h, w = style_imgs.shape# style_imgs: [B, 2*N, C:1, H, W] -> FEAT_ST_ENC: [4*N, B, C:512]style_imgs = style_imgs.view(-1, in_planes, h, w)  # [B*2N, C:1, H, W]# 经过CNN Encoderstyle_embe = self.Feat_Encoder(style_imgs)  # [B*2N, C:512, 2, 2]anchor_num = num_imgs//2style_embe = style_embe.view(batch_size*num_imgs, 512, -1).permute(2, 0, 1)  # [4, B*2N, C:512]FEAT_ST_ENC = self.add_position(style_embe)memory = self.base_encoder(FEAT_ST_ENC)  # [4, B*2N, C]writer_memory = self.writer_head(memory)glyph_memory = self.glyph_head(memory)writer_memory = rearrange(writer_memory, 't (b p n) c -> t (p b) n c',b=batch_size, p=2, n=anchor_num)  # [4, 2*B, N, C]glyph_memory = rearrange(glyph_memory, 't (b p n) c -> t (p b) n c',b=batch_size, p=2, n=anchor_num)  # [4, 2*B, N, C]# writer-ncememory_fea = rearrange(writer_memory, 't b n c ->(t n) b c')  # [4*N, 2*B, C]compact_fea = torch.mean(memory_fea, 0) # [2*B, C]# compact_fea:[2*B, C:512] ->  nce_emb: [B, 2, C:128]pro_emb = self.pro_mlp_writer(compact_fea)query_emb = pro_emb[:batch_size, :]pos_emb = pro_emb[batch_size:, :]nce_emb = torch.stack((query_emb, pos_emb), 1) # [B, 2, C]nce_emb = nn.functional.normalize(nce_emb, p=2, dim=2)# glyph-ncepatch_emb = glyph_memory[:, :batch_size]  # [4, B, N, C]# sample the positive pairanc, positive = self.random_double_sampling(patch_emb)n_channels = anc.shape[-1]anc = anc.reshape(batch_size, -1, n_channels)anc_compact = torch.mean(anc, 1, keepdim=True) anc_compact = self.pro_mlp_character(anc_compact) # [B, 1, C]positive = positive.reshape(batch_size, -1, n_channels)positive_compact = torch.mean(positive, 1, keepdim=True)positive_compact = self.pro_mlp_character(positive_compact) # [B, 1, C]nce_emb_patch = torch.cat((anc_compact, positive_compact), 1) # [B, 2, C]nce_emb_patch = nn.functional.normalize(nce_emb_patch, p=2, dim=2)# input the writer-wise & character-wise styles into the decoderwriter_style = memory_fea[:, :batch_size, :]  # [4*N, B, C]glyph_style = glyph_memory[:, :batch_size]  # [4, B, N, C]glyph_style = rearrange(glyph_style, 't b n c -> (t n) b c') # [4*N, B, C]# QUERY: [char_emb, seq_emb]seq_emb = self.SeqtoEmb(seq).permute(1, 0, 2)T, N, C = seq_emb.shape# ========================Content Encoder部分=========================char_emb = self.content_encoder(char_img) # [4, N, 512]char_emb = torch.mean(char_emb, 0) #[N, 512]char_emb = repeat(char_emb, 'n c -> t n c', t = 1)tgt = torch.cat((char_emb, seq_emb), 0) # [1+T], put the content token as the first tokentgt_mask = generate_square_subsequent_mask(sz=(T+1)).to(tgt)tgt = self.add_position(tgt)# 注意这里的执行顺序，Content Encoder输出 → Writer Decoder → Glyph Decoder → Embedding to Sequence# [wri_dec_layers, T, B, C]wri_hs = self.wri_decoder(tgt, writer_style, tgt_mask=tgt_mask)# [gly_dec_layers, T, B, C]hs = self.gly_decoder(wri_hs[-1], glyph_style, tgt_mask=tgt_mask)  h = hs.transpose(1, 2)[-1]  # B T Cpred_sequence = self.EmbtoSeq(h)return pred_sequence, nce_emb, nce_emb_patch

数据集部分

• CASIA_CHINESE

  data/CASIA_CHINESE
  ├── character_dict.pkl   # 词典
  ├── Chinese_content.pkl  # Content reference
  ├── test
  ├── test_style_samples
  ├── train
  ├── train_style_samples  # 1300个pkl，每个pkl中是同一个人写的各个字，长度不一致
  └── writer_dict.pkl
  

• 训练集中单个数据格式解析

  {'coords': torch.Tensor(coords),                # 写这个字，每一划的点阵'character_id': torch.Tensor([character_id]),  # content字的索引'writer_id': torch.Tensor([writer_id]),        # 某个人的style'img_list': torch.Tensor(img_list),            # 随机选中style的img_list'char_img': torch.Tensor(char_img),            # content字的图像'img_label': torch.Tensor([label_id]),         # style中图像的label
  }
  

• 推理时：
  • 输入：
    • 一种style15个字符的图像
    • 原始输入字符
  • 输出：属于该style的原始字符

总结

1. 感觉对于Transformer的用法，比较粗暴。当然，Transformer本来就很粗暴
2. 模型69M (position_layer2_dim512_iter138k_test_acc0.9443.pth) 比较容易接受，这和我之前以为的Transformer系列都很大，有些出入。这也算是纠正自己的盲目认知了
3. 学到了einops库的用法，语义化操作，很有意思，值得学习。
4. 第一次了解到NCE(Noise Contrastive Estimation)这个Loss，主要解决了class很多时，将其转换为二分类问题。