精度问题

[xxx2974]

有一个问题请教，直接训练和测试时结果是对的，但是用自己保存的训练的模型拿来测试，会提示一些问题，精度也非常低？
细节如下：

Keys not found in source state_dict:
base.layer1.2.bn2.weight
base.layer3.0.conv2.weight
base.layer3.0.downsample.1.weight
base.layer1.1.conv2.weight
base.layer2.1.conv2.weight
base.layer1.0.conv1.weight
base.layer1.0.bn2.running_mean
base.layer3.4.conv1.weight
base.layer2.0.conv3.weight
base.layer4.2.bn3.running_mean
base.layer3.4.bn1.bias
base.layer4.1.conv2.weight
base.layer4.0.conv3.weight
base.layer3.4.bn3.bias
base.layer1.0.bn1.running_var
base.layer4.0.bn3.weight
base.layer3.5.bn3.running_mean
base.layer1.2.bn1.bias
base.layer2.2.bn2.running_var
base.layer2.2.bn3.bias
base.layer2.1.bn2.bias
base.layer2.0.bn3.running_var
base.layer2.3.bn1.running_mean
local_bn.running_var
base.layer4.2.bn1.bias
base.layer2.2.conv2.weight
base.bn1.bias
base.layer3.3.bn1.bias
fc.weight
base.layer3.1.bn2.running_mean
base.layer3.1.bn2.weight
base.layer4.1.bn1.bias
base.layer2.2.bn2.bias
base.layer3.4.bn1.running_mean
base.layer1.0.downsample.1.bias
base.layer2.2.bn1.running_var
base.layer1.1.bn3.running_var
base.layer1.0.downsample.0.weight
base.layer4.0.bn3.running_mean
base.layer3.5.bn1.running_mean
base.layer2.2.bn1.running_mean
base.layer1.2.bn1.running_mean
base.layer1.1.bn3.running_mean
base.layer1.2.bn2.running_mean
base.layer2.2.conv1.weight
base.layer4.0.bn2.running_var
base.layer1.2.bn3.bias
base.layer4.0.downsample.0.weight
base.layer1.1.conv3.weight
base.layer1.0.conv3.weight
base.layer1.0.bn3.bias
base.layer1.2.bn1.running_var
base.layer1.1.bn2.bias
base.layer3.2.bn3.running_var
base.layer2.0.bn2.weight
base.layer3.0.downsample.1.running_var
base.layer3.4.conv2.weight
base.layer2.3.conv1.weight
base.layer3.3.conv3.weight
base.layer3.0.bn1.bias
base.layer3.0.downsample.1.running_mean
base.layer2.2.bn3.weight
base.bn1.weight
base.layer2.1.conv3.weight
base.layer3.5.bn2.bias
base.layer2.2.bn2.running_mean
base.layer1.0.bn2.weight
base.layer3.4.bn2.running_mean
base.layer1.0.bn1.running_mean
base.layer2.0.downsample.0.weight
base.layer3.4.bn1.running_var
base.layer1.2.conv1.weight
base.layer3.4.bn2.running_var
base.layer2.0.conv1.weight
base.layer1.1.bn2.running_mean
base.layer2.0.bn1.weight
base.layer3.0.downsample.1.bias
base.layer1.1.bn2.running_var
base.layer3.2.conv1.weight
base.layer1.0.bn2.running_var
base.layer3.5.bn1.bias
base.layer2.0.downsample.1.running_var
base.layer1.1.conv1.weight
base.layer2.1.bn1.running_var
base.layer1.2.conv3.weight
base.layer2.2.bn1.bias
base.layer3.3.bn2.running_var
base.layer4.0.conv2.weight
base.layer4.2.bn2.bias
base.layer2.3.conv2.weight
base.layer3.5.bn3.bias
base.layer3.0.bn2.weight
base.layer3.4.conv3.weight
base.layer2.1.bn3.running_mean
fc.bias
。。。
Keys not found in destination state_dict:
state_dicts
ep
scores
Loaded model weights from /home/xiaozhenzhen/data/AlignedReID-Re-Production-Pytorch/exp/train/market1501/not_nf_not_ohs_ gm_0.3_lm_0.3_glw_1_llw_0_idlw_0_lr_0.0002_exp_decay_at_151_total_300/run1/ckpt.pth

=========> Test on dataset: market1501 <=========

Extracting feature...
1000/1000 batches done, +1.88s, total 99.11s
Done, 99.32s
Computing global distance...
Done, 0.91s
Computing scores for Global Distance...
[mAP: 2.09%], [cmc1: 8.31%], [cmc5: 16.83%], [cmc10: 22.57%]
Done, 13.64s
Re-ranking...
Done, 64.49s
Computing scores for re-ranked Global Distance...
[mAP: 3.24%], [cmc1: 9.98%], [cmc5: 17.87%], [cmc10: 22.09%]
Done, 14.04s

[huanghoujing]

你好，看起来是模型没有load正确，那些权重都没有复制进去。你ckpt.pth这个文件应该直接放在exp_dir底下就行，不要传递给model_weight_file这个参数。关于ckpt.pth和model_weight_file的区别，可以参考我这个回答。

[xxx2974]

好的，谢谢，已经明白两种不同的load方式，非常感谢！

[jszgz]

你好，为什么我保存的ckpt.pth有500多MB，而你提供的才几十MB。此外，我使用测试的参数按你说的方法进行测试，还会报错。

Traceback (most recent call last):
File "script/experiment/train.py", line 632, in
main()
File "script/experiment/train.py", line 389, in main
test(load_model_weight=True)
File "script/experiment/train.py", line 376, in test
load_ckpt(modules_optims, cfg.ckpt_file)
File "./aligned_reid/utils/utils.py", line 239, in load_ckpt
m.load_state_dict(sd)
File "/usr/local/anaconda3/envs/wtalignedreid/lib/python2.7/site-packages/torch/optim/optimizer.py", line 84, in load_state_dict
saved_groups = state_dict['param_groups']
KeyError: 'param_groups'