Created by: ahsdx

 Aistudio上进行恢复训练有问题： 下面是恢复训练前的total_loss 2020-02-27 18:09:13,889 - train.py[line:148] - INFO: epoch 35, step 840, total_loss: 0.03663 loss_cls: 0.01022, loss_bbox: 0.02641, fg_num: 103, lr: 0.00133 2020-02-27 18:09:18,844 - train.py[line:148] - INFO: epoch 35, step 845, total_loss: 0.03665 loss_cls: 0.01024, loss_bbox: 0.02641, fg_num: 103, lr: 0.00133 2020-02-27 18:09:22,940 - train.py[line:152] - INFO: Model saved. 2020-02-27 18:09:24,562 - train.py[line:148] - INFO: epoch 36, step 0, total_loss: 0.03443 loss_cls: 0.00772, loss_bbox: 0.02671, fg_num: 113, lr: 0.00133 2020-02-27 18:09:29,837 - train.py[line:148] - INFO: epoch 36, step 5, total_loss: 0.03353 loss_cls: 0.01012, loss_bbox: 0.02341, fg_num: 112, lr: 0.00133 2020-02-27 18:09:35,410 - train.py[line:148] - INFO: epoch 36, step 10, total_loss: 0.03135 loss_cls: 0.00918, loss_bbox: 0.02217, fg_num: 104, lr: 0.00133 2020-02-27 18:09:40,857 - train.py[line:148] - INFO: epoch 36, step 15, total_loss: 0.03251 我的加载预训练的代码：（cfg['pretrained']是我上次模型保存的路径） start_program = fluid.default_startup_program() main_program = fluid.default_main_program() exe.run(start_program) if cfg['pretrained'] is not None: fluid.io.load_persistables(exe, cfg['pretrained'], start_program) # fluid.io.load_persistables(exe, cfg['pretrained']) for epoch in range(0,num_epochs): 下面是我恢复训练后的total_loss: 2020-02-28 10:06:46,114 - train.py[line:150] - INFO: epoch 0, step 0, total_loss: 0.19910 loss_cls: 0.11695, loss_bbox: 0.08215, fg_num: 92, lr: 0.00133 2020-02-28 10:06:51,442 - train.py[line:150] - INFO: epoch 0, step 5, total_loss: 0.13470 loss_cls: 0.08358, loss_bbox: 0.05113, fg_num: 101, lr: 0.00133 2020-02-28 10:06:56,578 - train.py[line:150] - INFO: epoch 0, step 10, total_loss: 0.12935 loss_cls: 0.07687, loss_bbox: 0.05247, fg_num: 106, lr: 0.00133 2020-02-28 10:07:01,745 - train.py[line:150] - INFO: epoch 0, step 15, total_loss: 0.14660 loss_cls: 0.09456, loss_bbox: 0.05204, fg_num: 104, lr: 0.00133 2020-02-28 10:07:07,140 - train.py[line:150] - INFO: epoch 0, step 20, total_loss: 0.13445 loss_cls: 0.08643, loss_bbox: 0.04802, fg_num: 103, lr: 0.00133 损失函数上升将近5倍。。。。。。 我的保存模型的代码是： fluid.io.save_persistables(dirname=cfg['save_path'], main_program=main_program, executor=exe)