INNER CODE UNIT · Python

probs

hyperonym/basaran · basaran/model.py:261

            probs = torch.nn.functional.softmax(logits, dim=-1)

            # Select deterministic or stochastic decoding strategy.
            if (config.top_p is not None and config.top_p <= 0) or (
                config.temperature is not None and config.temperature <= 0
            ):
                tokens = torch.argmax(probs, dim=-1)[:, None]
            else:
                tokens = torch.multinomial(probs, num_samples=1)

            # Collect log probabilities of the selected tokens.
            token_logprobs = torch.gather(probs, 1, tokens)
            token_logprobs = torch.log(token_logprobs + 1e-7).squeeze(1)
            tokens = tokens.squeeze(1)

            # Collect log probabilities of the most likely tokens.
            top_logprobs, top_tokens = probs.topk(logprobs)
            top_logprobs = torch.log(top_logprobs + 1e-7)

View source record →