Hi guys,
I’m using pytorch to train Tree-LSTMs. In order to reduce the bottleneck in data loading on GPUs, I’m trying to use the DataLoader which takes as input a custom Dataset class from this tutorial. My “Tree” data structure is implemented recursively as
{parent:[children]}
with parent being a torch.Tensor
and children being a list of torch.Tensor
if the node is childless or a data structure of similar format if the node has children itself. As an example
{Tensor :
[{Tensor :
[Tensor, Tensor]},
Tensor]}
When loading manually the data points one by one, I do manage to train my model. When I try to load the data via the DataLoader, however, I get a KeyError:parent_tensor
when iterating through the generator.
any idea on how to solve that ?
Thanks !