It's impossible to resume a PytorchExperiment with just
because there is no base_dir key in self._config_raw. By default the latter is constructed with
self._config_raw = self._config_raw_from_input(config, name, n_epochs, seed, append_rnd_to_name)
where config is None by default. As a consequence there is no base_dir to read. The important part here is that the PytorchExperiment SHOULD NOT REQUIRE a base_dir. Seeing that the initialization code is very long and complex, I suggest refactoring that will make it easier to understand and avoid such mistakes.