Not able to train Model on CPU mode

Hi we trying to train our own dataset using the pretrained model “googlenet” on digits ver -5
The CUDA ver is 8 and CUDNN ver is 6.

The host PC has NVIDIA Geforce graphic card 820M but its architecture supports Fermi.Since TX1 is based on Maxwell architecture we want to train the model on CPU only mode and port the model to JetsonTX1.

we modified the file " make config in caffee" to make digits use only CPU. but the training failed with the following error.



Thanks for your question.
The error indicates Caffe tried to use cuDNN but it’s not in GPU mode.

Architecture is not an issue for training a model.
Digits saves model with general caffe format and is nothing related to GPU architecture.

Please feel free to train a model on desktop GPU and then port it to TX1.
In TX1, tensorRT will compile the model into tensorRT format based on GPU architecture.