torch model trained with GPU be used in a CPU-only environment

[Xuwang Yin]

I have a torch model trained with GPU, now i want to use this model in a CPU-only environment.

Is it possible, if so, how can i do that? thanks!

[soumith]

model = model:float()

That's it, you should be all set.

On Wed, Jun 10, 2015 at 11:55 AM, Xuwang Yin via torch7 <torch7+APn2wQcrd_8OQt5Q0FoODsOgr...@googlegroups.com> wrote:

I have a torch model trained with GPU, now i want to use this model in a CPU-only environment.

Is it possible, if so, how can i do that? thanks!

--
You received this message because you are subscribed to the Google Groups "torch7" group.
To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at http://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[Yossi Biton]

will work unless you use some packages with cuda support only (ccn2 or inn for example).

On Wednesday, June 10, 2015 at 6:57:08 PM UTC+3, smth chntla wrote:

model = model:float()

That's it, you should be all set.

[Xuwang Yin]

I think there might be some issues here.

My experiments showed that to load a GPU-trained nn.Sequential model, 

you have to first "require 'cunn'", otherwise it won't load ("requre 'nn'" only): "torch/install/share/lua/5.1/torch/File.lua:249: unknown Torch class <torch.CudaTensor>". 

In the training code i found that when you trained a nn.Sequential model on GPU, you have to first "convert" the model to GPU: 

-- CUDA?

if opt.type == 'cuda' then

   model:cuda()

   criterion:cuda()

end

I think that after the model being "converted", you have to require cunn to load the model. 

So, even you want to use the model on CPU, you have to get a NVIDIA card and have CUDA installed, which, 

will makes running the model on CPU pointless.

Any suggestions?

在 2015年6月10日星期三 UTC+8下午11:57:08，smth chntla写道：

model = model:float()

That's it, you should be all set.

[Yossi Biton]

convert your model to float/double before saving it :

model:float()

now you don't have to require 'cunn' when loading it.

[Xuwang Yin]

Yes, it works, and if you didn't save the model as float during the training process, you can use 'cunn' to load the model, convert it to float and then save it. Nice work

On Sat, Jun 13, 2015 at 11:00 PM, Yossi Biton via torch7 <torch7+APn2wQezX724pTV9rC5uq5YnV...@googlegroups.com> wrote:

convert your model to float/double before saving it :

model:float()

now you don't have to require 'cunn' when loading it.

On Saturday, June 13, 2015 at 4:40:42 PM UTC+3, Xuwang Yin wrote:

I think there might be some issues here.

My experiments showed that to load a GPU-trained nn.Sequential model, 

you have to first "require 'cunn'", otherwise it won't load ("requre 'nn'" only): "torch/install/share/lua/5.1/torch/File.lua:249: unknown Torch class <torch.CudaTensor>". 

In the training code i found that when you trained a nn.Sequential model on GPU, you have to first "convert" the model to GPU: 

-- CUDA?

if opt.type == 'cuda' then

   model:cuda()

   criterion:cuda()

end

I think that after the model being "converted", you have to require cunn to load the model. 

So, even you want to use the model on CPU, you have to get a NVIDIA card and have CUDA installed, which, 

will makes running the model on CPU pointless.

Any suggestions?

在 2015年6月10日星期三 UTC+8下午11:57:08，smth chntla写道：

model = model:float()

That's it, you should be all set.

On Wed, Jun 10, 2015 at 11:55 AM, Xuwang Yin via torch7 <torch7+APn2wQcrd_8OQt5Q0FoODsOgr...@googlegroups.com> wrote:

I have a torch model trained with GPU, now i want to use this model in a CPU-only environment.

Is it possible, if so, how can i do that? thanks!

--
You received this message because you are subscribed to the Google Groups "torch7" group.
To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at http://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to a topic in the Google Groups "torch7" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/torch7/ugBCwaoXw_s/unsubscribe.
To unsubscribe from this group and all its topics, send an email to torch7+un...@googlegroups.com.

[ignorant]

Hi,

Is there a way to make this work for nets with inn module as well? For example, Alexnet from https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_caffenet

I want to deploy the trained net for prediction on a cheaper AWS instance.

Thanks,

On Saturday, June 13, 2015 at 11:35:03 PM UTC-4, Xuwang Yin wrote:

Yes, it works, and if you didn't save the model as float during the training process, you can use 'cunn' to load the model, convert it to float and then save it. Nice work

[Francisco Vitor Suzano Massa]

For the moment, no, because Alexnet
1. uses SpatialCrossResponseNormalization, which is not present in nn yet
2. uses groups=2 for convolution

There is an ongoing PR to add SpatialCrossResponseNormalization to nn https://github.com/torch/nn/pull/413 , but even after it gets merged nn doesn't support groups in convolution for the moment. This could be overcome by some network surgery though. I'll probably add support to groups in SpatialConvolution sometime after CVPR though (i.e., in ~3 weeks)

[ignorant]

Thanks for the work! 

I thought of taking care of it in loadcaffe next but saw responses in https://github.com/szagoruyko/loadcaffe/issues/8 . So I guess I will wait for now.. or maybe use caffe in cpu mode for prediction till there is a resolution.

[Francisco Vitor Suzano Massa]

Just for information, if you want to use groups in torch in CPU, I've hacked a simple function that takes care of that
https://github.com/fmassa/torch-nn/blob/master/SpatialConvolutionGroups.lua

But beware that I haven't tested it thoughtly, there might be bugs

[vasg]

I have trained the model on a GPU, I tried loading it in GPU environment, and received unknown Torch class <torch.CudaTensor> error. This error didn't occur when I added "require 'cunn'" before loading the model.
Does anyone know the reason for this behaviour?

Thanks!

[Vislab]

You have to require 'cunn' when loading models that have CudaTensors in them.

[Venu Sharma]

Hi  Francisco Vitor Suzano Massa,

    Do you have a PR for groups implementation in SpatialConvolution for nn?  

[Jon Pi]

Hey

Did anyone get a 

attempt to call method 'double' (a nil value)

or

attempt to call method 'float' (a nil value)

when trying:

model = torch.load(modelName)
model = model.float()    # model = model.double()

Thanks!

[Jon Pi]

oops! I meant when trying:

model = torch.load(modelName)
model = model:float()    # model = model:double()

[Mika S]

Can you paste result of :

torch.type(model)

[Wang Jie]

hello,i don't know how to use model = model:float(),please help.

在 2015年6月10日星期三 UTC+8下午11:57:08，smth chntla写道：

model = model:float()

That's it, you should be all set.

[Spandan Madan]

Hi Jon,

Did you end up finding a fix for this? I'm having the same issue.

[Spandan Madan]

Here's what I try:

model = torch.load(ModelName)

model = model:float()

Here's the error-

[string "model_lineback = model_lineback:float()"]:1: attempt to call method 'float' (a nil value)

stack traceback:

        [string "model_lineback = model_lineback:float()"]:1: in main chunk

        [C]: in function 'xpcall'

        /home/t-spmada/torch/install/share/lua/5.1/trepl/init.lua:679: in function 'repl'

        ...mada/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:204: in main chunk

        [C]: at 0x00405d50