mGPT/metrics/base.py

from torch import Tensor, nn
from os.path import join as pjoin
from .mr import MRMetrics
from .t2m import TM2TMetrics
from .mm import MMMetrics
from .m2t import M2TMetrics
from .m2m import PredMetrics


class BaseMetrics(nn.Module):
    def __init__(self, cfg, datamodule, debug, **kwargs) -> None:
        super().__init__()

        njoints = datamodule.njoints

        data_name = datamodule.name
        if data_name in ["humanml3d", "kit"]:
            self.TM2TMetrics = TM2TMetrics(
                cfg=cfg,
                dataname=data_name,
                diversity_times=30 if debug else cfg.METRIC.DIVERSITY_TIMES,
                dist_sync_on_step=cfg.METRIC.DIST_SYNC_ON_STEP,
            )
            self.M2TMetrics = M2TMetrics(
                cfg=cfg,
                w_vectorizer=datamodule.hparams.w_vectorizer,
                diversity_times=30 if debug else cfg.METRIC.DIVERSITY_TIMES,
                dist_sync_on_step=cfg.METRIC.DIST_SYNC_ON_STEP)
            self.MMMetrics = MMMetrics(
                cfg=cfg,
                mm_num_times=cfg.METRIC.MM_NUM_TIMES,
                dist_sync_on_step=cfg.METRIC.DIST_SYNC_ON_STEP,
            )

        self.MRMetrics = MRMetrics(
            njoints=njoints,
            jointstype=cfg.DATASET.JOINT_TYPE,
            dist_sync_on_step=cfg.METRIC.DIST_SYNC_ON_STEP,
        )
        self.PredMetrics = PredMetrics(
            cfg=cfg,
            njoints=njoints,
            jointstype=cfg.DATASET.JOINT_TYPE,
            dist_sync_on_step=cfg.METRIC.DIST_SYNC_ON_STEP,
            task=cfg.model.params.task,
        )