Leonard

Added data filter to the BotAdversarialDialogueTeacher to remove repeating "Hey do you want to talk about something else" episodes.

2

In the data ~7k episodes end with "Hey do you want to talk about something else". With this flag we can skip these episodes.

Updated director toxic classifier task

1

Updated toxic classifier task for the director model: * now supports cpu-only evaluation (needed during multi-gpu training) * added additional flags to support truncation and limited number of examples *...

Moved generation activation during generation of TorchGeneratorAgent to dedicated method for easier adjusting. This enables easy adjusting when we want to use another activation function like sigmoid in the scones-style...

Generalized validation metric to support product or sum function of multiple criteria.

1

You can provide a comma separated list as validation metric and it will use the product or sum of those as final validation metric.

` c_tilde = torch.transpose(torch.stack( [c_l_, torch.cat([c_s[:, :self.small_cell_size], c_l_[:, self.small_cell_size:self.large_cell_size]], dim=1) ], dim=2), 1, 2)` Shouldn't it be `c_l[:, ...]` here instead of `c_l_[:, ...]`? The way it is right now...