需要的包:
transformers
einops
scikit-learn
h5py
imageio
train_0219:最初的预训练版本
train_0224:加入了test部分和第二个训练部分
train_0313:mse损失,正负样本冲突
2023.03.19 train_0312.py KL损失,epoch=400
2023.03.20 train_0312.py softmax(sim_)+mse,epoch=400
2023.03.21 train_0312.py transformer_image softmax(sim_)+mse,epoch=400
2023.03.22 开题
2023.03.23 train_0312.py MyModule_v7 image_linear*2 maxminsclar+mse loss:1 0.5 0.5
(dftl202303231763641)
2023.03.25 train_0312.py MyModule_v7 image_linear3 event_linear3 maxminsclar+mse loss:1 0.5 0.5
(dftl202303251009770)
2023.03.25 train_0312.py MyModule_v7
2023.03.27 train_all.py MyModule_v8 +crossattention pretrained:epoch_144_loss_0.003096(dftl202303231763641)
2023.03.28(X) !!modify output!! train_all.py MyModule_v8 +crossattention pretrained:epoch_144_loss_0.003096(dftl202303231763641)
2023.03.29 train_0312.py (linear,bn,relu) * 3 (id:dftl202303291160175)
2023.03.31 train_all.py MyModule_v8 +crossattention pretrained:epoch_168_loss_0.000003.pt
pre_eval.py
2023.04.02 train_all.py cross_attntion+res pretrained:epoch_144_loss_0.003096(dftl202303231763641)
2023.04.03 pre_eval.py
2023.04.06 train_all.py cross_attntion*2 + img linear pretrained:epoch_144_loss_0.003096(dftl202303231763641)
2023.04.07(dftl202304071655897) train_0312.py 分类头后移