INNER CODE UNIT · Python

indices

lucidrains/vector-quantize-pytorch · vector_quantize_pytorch/binary_mapper.py:100

            indices = one_hot.argmax(dim=-1)

        # allow for any number of leading dimensions

        logits, inverse_pack_lead_dims = pack_with_inverse(logits, '* bits')
        indices, _ = pack_with_inverse(indices, '*')

        # sampled bits representation

        sampled_bits = self.codes[indices]

        # calculate log probability

        log_probs_1 = F.logsigmoid(logits)
        log_probs_0 = F.logsigmoid(-logits)

        log_probs = torch.where(sampled_bits, log_probs_1, log_probs_0)

View source record →