Serialization fails with custom Torch class

Question

Serialization can fails with a class object created containing __pairs:

test = torch.class('test')
function test:__init()
  self.data = {}
end

function test:__pairs(...)
    return pairs(self.data, ...)
end

function test:get_data()
  print(self.data)
end

a = test.new()
a.data = {"asdasd"}
b = torch.serialize(a)
c = torch.deserialize(b)
print(torch.typename(c))
print(c:get_data())

The following returns:

test
nil

Answer 1

The engine behind the torch.serialization is located in the File-class. The File:writeObject us the key function. For the above example the action for a Torch class starts at line 201 with the:

elseif typeidx == TYPE_TORCH then

The type is identified in the File:isWritableObject.

One could probably implement the metatable function write but in the above example the problem was non-torch metatable function __pairs that should be __pairs__ (see torch.getmetatable):

test = torch.class('test')
function test:__init()
  self.data = {}
end

function test:__pairs__(...)
    return pairs(self.data, ...)
end

function test:get_data()
  print(self.data)
end

a = test.new()
a.data = {"asdasd"}
b = torch.serialize(a)
c = torch.deserialize(b)
print(torch.typename(c))
print(c:get_data())

This gives the expected:

test    
{
  1 : "asdasd"
}