DCVC-DC multi_scale_feature_extractor

[wulayu]

def multi_scale_feature_extractor(self, dpb, index):
    if dpb["ref_feature"] is None:
        feature = self.feature_adaptor_I(dpb["ref_frame"])
    else:
        index = index % 4
        index_map = [0, 1, 0, 2]
        index = index_map[index]
        feature = self.feature_adaptor[index](dpb["ref_feature"])
    return self.feature_extractor(feature)

if use IPPP pattern, index =0 is I-frame, index =1 is self.feature_adaptor_I, index = 2 is self.feature_adaptor[0], index = 3 is self.feature_adaptor[2], then the self.feature_adaptor[1] is not used during training. I want to make sure I'm understanding this correctly. Could you please clarify?​

[yaohualibin]

You can train with longer sequence, then self.feature_adaptor[1] will be trained.