Transformer 架构图

横向 16:9 学术概念图，展示 Transformer 编码器-解码器架构，NeurIPS 定稿风格。左右两列垂直堆叠，中间用虚线分隔。 左列标题："ENCODER (×N)"。模块从下到上依次为："Input tokens" → "Input Embedding" → "+ Positional Encoding" → 虚线框 "Encoder layer"，包含 "Multi-Head Self-Attention"、"Add & Norm"、"Feed-Forward"、"Add & Norm"，每个子层周围有细弯曲的残差箭头。 右列标题："DECODER (×N)"。模块从下到上依次为："Output tokens (shifted right)" → "Output Embedding" → "+ Positional Encoding" → 虚线框 "Decoder layer"，包含 "Masked Multi-Head Self-Attention"、"Add & Norm"、"Multi-Head Cross-Attention"（有箭头从编码器顶部标注 "keys, values" 指向此处）、"Add & Norm"、"Feed-Forward"、"Add & Norm"。解码器上方有 "Linear"、"Softmax"、"Output probabilities"。 标题："Transformer: encoder–decoder with multi-head attention"。副标题："Vaswani et al., 2017"。