How to save a tensor to a binary file

[fireman]

Hi everyone:

I am new to Lua/Torch and hope  you guys can help me. I would like to save the weight and bias of a nn.SpatialConvolution to a raw binary file. But I just could not find any thing that can do this in Lua. Do I miss anything?

Thanks,

Ge

[alban desmaison]

Hi,

torch.save already save data in a byte format. The option to save as "ascii" can optionally be used (https://github.com/torch/torch7/blob/master/doc/serialization.md#torchsavefilename-object--format-referenced)

[fireman]

I would like to save the weights and bias in a more compact format and read it somewhere else. I do not want to use torch saving format. 

I am asking a general lua question about how to save some data in binary file. 

[fireman]

I can not believe Lua has no way to save data to a binary file????

[soumith]

you can open a torch.DiskFile and write weight:storage() and bias:storage() to the diskfile using: https://github.com/torch/torch7/blob/master/doc/file.md#write-methods

For example:

require 'nn'

a=torch.DiskFile('abc.t7', 'w'):binary()

m = nn.SpatialConvolution(3,16,5,5)

m:float()

a:writeFloat(m.weight:storage())

a:writeFloat(m.bias:storage())

a:close()

--
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+unsubscribe@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at https://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[Daniel Galvez]

Note that saving like this leaves out the associated dimensions of the tensor, since it's just the storage. The data is stored row-major.

To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.

[fireman]

I tried that but seems there is something wrong. I just want to load a previous trained model, find the Convolution layer and dump the weight and bias to a file. Here is my code

local model = //local a previous trained torch cnn mode from file

 local a = torch.DiskFile(opt.output..'.t7', 'w'):binary()

  for i = 1, #model do

    local layer = model:get(i)

    local layer_type = torch.type(layer)

    print(string.format('Layer %d: %s', i, layer_type))

    if layer_type == 'nn.SpatialConvolution' then

      layer:float()

      if opt.writedata then

        a:writeFloat(layer.weight:storage())

        a:writeFloat(layer.bias:storage())

      end

    end

It did write to a file, but the file size is very big, it is even bigger than my original model.

[soumith]

a:writeFloat(layer.weight:storage()) is wrong. you need to do:

local w = layer.weight
a:writeFloat(w:storage().new(w:storage(), w:storageOffset(), w:nElement())

--

[fireman]

Still does not seem correct. When I read the file from matlab on windows, the weights are 

   1.0e+37 *

    0.0000

       NaN

    0.0000

   -0.0000

    3.0767

    0.0000

    0.0000

    0.0000

    0.0000

   -0.0000

On Thursday, November 3, 2016 at 3:33:07 PM UTC-7, smth chntla wrote:

a:writeFloat(layer.weight:storage()) is wrong. you need to do:

local w = layer.weight
a:writeFloat(w:storage().new(w:storage(), w:storageOffset(), w:nElement())

On Thu, Nov 3, 2016 at 5:53 PM, fireman via torch7 <torch7+APn2wQeW3esQD-dFTzB55wMyBRX8xJbzqz3UhAPNuLWPpnI6ph...@googlegroups.com> wrote:

I tried that but seems there is something wrong. I just want to load a previous trained model, find the Convolution layer and dump the weight and bias to a file. Here is my code

local model = //local a previous trained torch cnn mode from file

 local a = torch.DiskFile(opt.output..'.t7', 'w'):binary()

  for i = 1, #model do

    local layer = model:get(i)

    local layer_type = torch.type(layer)

    print(string.format('Layer %d: %s', i, layer_type))

    if layer_type == 'nn.SpatialConvolution' then

      layer:float()

      if opt.writedata then

        a:writeFloat(layer.weight:storage())

        a:writeFloat(layer.bias:storage())

      end

    end

It did write to a file, but the file size is very big, it is even bigger than my original model.

--
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.