Unified personal research repository
Some models in models.archive is not working in this framework

• m2m Flow variance 가 낮아지게 학습해보기
• distillation 할 때 epe기준 말고 variance 기준 weighting (modify get robust weight)
• 좀 흐려보이는데 cosine positional encoding 추가해보기
• RAFT 처럼 attention 할 수 있는 방법 찾아보기

IFRNet (4,959,692) Time: 0.006s Parameters: 4.96M

EMA-VFI small Time: 0.014s Parameters: 14.49M

EMA-VFI Time: 0.034s Parameters: 65.66M

• Deformable Conv로 Query building
• Source Target에 대해 다 attention하고 mixing
• Deformable attention 아니고 Swin attention임

DCNTransv1 (2,715,457)

• 원래 IFRNet은 Decoder4에서 바로 feature t 생성해버림
• 여기서는 DCN을 사용한 Query builder로 feat_t_3 생성
• GMTrans에 있는 decoder2 로 decoding

DCNTransv1_decRes10_GeoF3_noDistill_halfTonly (5,509,399)

• encoder residual layer 5개, decoder residual 10개 DCNTransv1_sepDCN_E5D10_dim64_Geo32_distill_bwarp (4,255,319) DCNTransv2_sepDCN_E5D10_dim64_Geo32_distill_featFwarp (4,255,319)

DATv1_sepDCNBwarpEmbT_shareAttBothDAT_noPE_E0D5_dim72_p256_bwarp (4,042,351) DATv1_sepDCNBwarp_shareDAT_noPE_E5D10_dim72_bwarp (5,335,111) DATv1_sepDCNBwarpEmbT_shareAttBothDAT_noPE_E5D10_dim72_bwarp(4,977,631)

DCNDATv1_shareDCNBwarpEmbT_QDCNAttnBothDAT_noPE_E5D10_distill_dim64_p256_bwarp (3,751,637) Time: 0.048s
Parameters: 3.75M