Method forward

slowfast/models/video_model_builder.py:968–1013

(self, x)

            return {}

    def forward(self, x):
        x = x[0]
        x = self.patch_embed(x)

        T = self.cfg.DATA.NUM_FRAMES // self.patch_stride[0]
        H = self.cfg.DATA.TRAIN_CROP_SIZE // self.patch_stride[1]
        W = self.cfg.DATA.TRAIN_CROP_SIZE // self.patch_stride[2]
        B, N, C = x.shape

        if self.cls_embed_on:
            cls_tokens = self.cls_token.expand(
                B, -1, -1
            )  # stole cls_tokens impl from Phil Wang, thanks
            x = torch.cat((cls_tokens, x), dim=1)

        if self.sep_pos_embed:
            pos_embed = self.pos_embed_spatial.repeat(
                1, self.patch_dims[0], 1
            ) + torch.repeat_interleave(
                self.pos_embed_temporal,
                self.patch_dims[1] * self.patch_dims[2],
                dim=1,
            )
            pos_embed_cls = torch.cat([self.pos_embed_class, pos_embed], 1)
            x = x + pos_embed_cls
        else:
            x = x + self.pos_embed

        if self.drop_rate:
            x = self.pos_drop(x)

        if self.norm_stem:
            x = self.norm_stem(x)

        thw = [T, H, W]
        for blk in self.blocks:
            x, thw = blk(x, thw)

        x = self.norm(x)
        if self.cls_embed_on:
            x = x[:, 0]
        else:
            x = x.mean(1)

        x = self.head(x)
        return x