【问题标题】:Error using pre-trained model in DIGITS with different dataset. How to modify layers based on new datasets?在具有不同数据集的 DIGITS 中使用预训练模型时出错。如何根据新数据集修改图层?
【发布时间】:2017-03-01 20:36:28
【问题描述】:

我尝试将预训练模型 (VGG 16) 用于 DIGITS,但出现此错误。

错误:检查失败:错误 == cudaSuccess (2 vs. 0) 内存不足

conv2_2 does not need backward computation.
relu2_1 does not need backward computation.
conv2_1 does not need backward computation.
pool1 does not need backward computation.
relu1_2 does not need backward computation.
conv1_2 does not need backward computation.
relu1_1 does not need backward computation.
conv1_1 does not need backward computation.
data does not need backward computation.
This network produces output label
This network produces output softmax
Network initialization done.
Solver scaffolding done.
Finetuning from /home/digits/digits/jobs/20161020-095911-9d01/model.caffemodel
Attempting to upgrade input file specified using deprecated V1LayerParameter: /home/digits/digits/jobs/20161020-095911-9d01/model.caffemodel
Successfully upgraded file specified using deprecated V1LayerParameter
Attempting to upgrade input file specified using deprecated input fields: /home/digits/digits/jobs/20161020-095911-9d01/model.caffemodel
Successfully upgraded file specified using deprecated input fields.
Note that future Caffe releases will only support input layers and not input fields.
Check failed: error == cudaSuccess (2 vs. 0)  out of memory

我成功地将deploy.prototxtVGG_ILSVRC_16_layers.caffemodelsynset_words.txt 上传到DIGITS 并使用我的数据集进行了测试,该数据集有两个类。

【问题讨论】:

    标签: machine-learning computer-vision deep-learning caffe nvidia-digits


    【解决方案1】:

    有时digits-server 无法清除内存。如果您使用的是 ubuntu,请尝试使用此命令:

    sudo restart nvidia-digits-server
    

    如果这不起作用并且你再次面对同样的情况,你需要减少batch_size

    【讨论】:

    • 谢谢。内存问题已解决,但错误仍然存​​在。我必须更改最后一层,但是当我更改名称时,训练过程并不顺利。
    • 预训练的和你的数据集都一样吗?如果至少预训练模型不是从类似的数据库中准备的,您将不会得到更好的结果
    • 我有不同的数据集,只有两个类,我想将模型用于我的数据集。为了修改和更改我的新数据集,我阅读了“微调”和“迁移学习”,但我在训练中出错了。
    猜你喜欢
    • 2017-02-28
    • 2017-03-02
    • 2017-06-04
    • 2020-05-21
    • 1970-01-01
    • 2019-10-12
    • 2023-02-17
    • 1970-01-01
    • 2017-09-10
    相关资源
    最近更新 更多