Why does Lua/Torch run out of memory when an `self` is returned to console from a `torch.class` method?

I've modified the torch-dataframe to return self instead of void in order to achieve simple method chaining for a torch.class. Unfortunately this seems to be causing havoc with memory issues:

th> require 'Dataframe'; df = torch.load('dataset_4_torch.t7')
                                                                      [4.8434s] 
th> b = df:create_subsets() -- Works
                                                                      [0.7384s] 
th> df:create_subsets() -- Fails even if called before the b = df:create_...
/home/max/tools/torch/install/bin/luajit: not enough memory

I've tried overwriting the default print that is called on all returned objects but it didn't help.

Memory details

Here's some memory profiling:

th> collectgarbage("count")
1836.24609375   
                                                                      [0.0002s] 
th> require 'Dataframe'; df = torch.load('dataset_4_torch.t7')
                                                                      [4.6875s] 
th> collectgarbage("count")
59659.619140625 
                                                                      [0.0003s] 
th> b = df:create_subsets()
                                                                      [0.7571s] 
th> collectgarbage("count")
62303.567382812 
                                                                      [0.0001s] 
th> df:create_subsets()
/home/max/tools/torch/install/bin/luajit: not enough memory

If this problem is to hard then I would appreciate an example of how to properly apply the method chaining patterns for a torch.class.

Solution

Thanks for helping me out. It turns out that I put in an assertion and in the error message I added at some point a tostring of the entire column. This worked in our testing environment but in a real dataset a massive string was generated and hence the memory issue.