add dynamic MML, sampling #17
Conversation
| # this is the only change required to sample from current log probabilities instead of beam search | ||
| if sample: | ||
| categorical_dist = Categorical(logits=curr_log_probs) | ||
| _, group_index, log_prob, action_embedding, action = batch_states[categorical_dist.sample()] |
There was a problem hiding this comment.
Looks like you're sampling only one action here. Was that intentional?
There was a problem hiding this comment.
I guess it was. So the idea is that you sample one action at each time step instead of keeping a sorted beam, correct?
There was a problem hiding this comment.
Yes, so I want a sampler function that can sample a whole decoded sequence, and then call THAT multiple times. So at the end i'll have max_num_decoded_sequences number of sequences for each batch instance
| # (group_size, num_start_type) | ||
| start_action_logits = self._start_type_predictor(hidden_state) | ||
| log_probs = torch.nn.functional.log_softmax(start_action_logits, dim=-1) | ||
| sorted_log_probs, sorted_actions = log_probs.sort(dim=-1, descending=True) |
There was a problem hiding this comment.
You'll have to make changes here too, to sample the first action from log_probs above.
There was a problem hiding this comment.
Oh i didn't see this. why is the first step handled in a different way than the rest of the steps?
There was a problem hiding this comment.
ok, so i'm going to add the same sampling logic to _take_first_step as well
…y finished states...
| all_actions = all_actions.detach().cpu().numpy().tolist() | ||
| if sample_states: | ||
| # (group_size,) one action per group element | ||
| sampler = Categorical(logits=sorted_log_probs) |
There was a problem hiding this comment.
Why do you have to use sorted log_probs here? You could have just passed log_probs, correct?
There was a problem hiding this comment.
so I wanted sampler to be able to return indices into all_actions which works consistently only if we use sorted_log_probs
merging master
The logic for sampling seems simple, but I feel I'm missing something here. Can you please take a look at line 305 in
basic_transition_function.py, @pdasigi?