面部表情识别（FER）动态模型中特征提取器和 Transformer 编码器层的具体配置和训练过程

We should note that the feature extractor is firstly pretrained on several FER datasets, then fine-tuned on the AffWild2 dataset frames, and is finally frozen so that it does not change its weights during the training of the dynamic model. As we fixed the number of static embeddings by modifying respective static models, the feature extractor always outputs 256 features per frame. For the Transformer based layers, we set the number of heads equal to 8 and dropout equals 0.1. Additionally, the positional encoding employed in [43] is applied to visual embeddings

这段描述详细介绍了面部表情识别（FER）动态模型中特征提取器和 Transformer 编码器层的具体配置和训练过程：

1. 特征提取器的预训练和微调：

   • 预训练：特征提取器首先在多个 FER 数据集上进行了预训练。这一步是为了让特征提取器在广泛的数据上学习基本的表情特征，从而具备对表情特征的初步理解。
   • 微调：在预训练之后，特征提取器在特定的 AffWild2 数据集帧上进行微调。AffWild2 是一个包含情绪标签的大型动态面部表情数据集，其数