“出现注意力模型”是什么意思？指图像出现线条吗？

我自己准备的数据量大概1GB，步长调到96跑了一天，现在15k了，Loss在几个小时之后就降到0.2之下了，现在是0.12
出现的所有图谱都只是在最顶端有一条紧贴顶边的彩色，除此之外连个线头都没有OTZ
是步长设置的太大了吗？

如何接着别人的模型用自己的数据进行训练？
是在预处理阶段把pretrained.pt换成作者提供的pretrained-11-7-21_75k.pt，还是在训练合成器开始之后生成的MockingBird\synthesizer\saved_models\run_id\run_id.pt中，把这个.pt文件换成pretrained-11-7-21_75k.pt之后继续训练合成器？

Start by reproducing the reported training setup with the roughly 1GB dataset, step size 96, and the graph output described in the issue. Inspect the preprocessing use of pretrained.pt and the saved model path MockingBird/synthesizer/saved_models/run_id/run_id.pt, then determine which checkpoint is intended for continued training. Done means the graph behavior and checkpoint procedure are clearly explained.

Written by the indexing model from the issue text.