Reading checkpoints... Checkpoint not find train from scratch... Total train image:3106 Total validate image:152 Total epoch:3 Batch size:4 Learning rate:0.001 Data Argument: h_flip: True v_flip: True rotate: True clip size: 450 除此之外,每一个epoch后,loss都是负的,没经过一次epoch 负号后面的值也越来越大....