andreasveit / densenet-pytorch Goto Github PK

Hello there! Very happy to use your code, I have a question looking forward to your reply!
In ``train.py"

losses.update(loss.data[0], input.size(0))
top1.update(prec1[0], input.size(0))

loss.data seems just a number (torch.Size([])) and does no have index 0.

when I runed the same code with you ?How to solve this error?Thank you very much!
I have solved the last question,Thank you very much

Hello there! Very happy to use your code, I have a question looking forward to your reply!
The same structure of DenseNet, the parameters of the pytorch version are much larger than the model parameters of the original paper? ? ? DenseNet (k=12) 40 layers, only 1.0M in the paper, 4M in the pytorch version, why is it so much worse? look forward to your reply

Thank you for the great repository.

I wonder if you got the CIFAR-100 results as stated in the paper.

For example, DenseNet-BC(100, k=24) with 17.60% error on C100+
or DenseNet-BC(100, k=40) with 17.18% error on C100+

When I use your code to train on ImageNet, the commonly used batch size 256 or 128 does not work due to short of gpu memory, how can I solve it?

I used the DenseNet-BC architecture in Table 2 of the original paper. When I tried with {L=190, k=40} and {L=250, k=24}, the number of parameters seems to be correct (15.3M and 25.6M), but it doesn't fitin to a 12GB Titan X memory. The batchsize is 64.

Do you have the same problem? If not, could you give some possible reasons for causing this problem?
Thanks!

Hey Andreas,

Do you have an implementation of FCN DenseNet https://arxiv.org/pdf/1611.09326.pdf paper?

Thanks!

Hey Andreas,

when I run your code,RuntimeError: CUDNN_STATUS_ARCH_MISMATCH,why?
Thanks!