When buffers (i.e. the TensorDict) are memory mapped, those are by default mapped to tempfiles, which are cleared once everything is done (https://github.com/pytorch-labs/tensordict/blob/main/tensordict/memmap.py#L560): If we are going to resume a training from a checkpoint (for Dreamer's algos), the buffer is lost.
We need to specify where to save the memory mapped buffers upon buffer creation.
When buffers (i.e. the TensorDict) are memory mapped, those are by default mapped to tempfiles, which are cleared once everything is done (https://github.com/pytorch-labs/tensordict/blob/main/tensordict/memmap.py#L560): If we are going to resume a training from a checkpoint (for Dreamer's algos), the buffer is lost. We need to specify where to save the memory mapped buffers upon buffer creation.