Back to Search Document details
31st Meeting: Geneva, CH, July 2023 2023-07-13 08:45
EE1-1.5: Optimization for complexity-performance trade-off of HOP network
Authors: Renjie Chang Tencent Liqiang Wang Tencent Xiaozhong Xu Tencent Shan Liu Tencent
Abstract
This contribution reports the EE-1.5 result of JVET-AE0160. Based on the unified network architecture of High Operation Point (HOP), this contribution proposes further optimization including new resblock and split network design to improve the complexity-performance trade-off of HOP network. Specifically, based on the depthwise separable convolution and group convolution, two novel resblocks are proposed to reduce the computational complexity without affecting the coding performance too much. Besides, the split network design for luma and chroma components is introduced into HOP network. Based on NNVC-5.1, the test results are shown in order of RA and AI configurations as follows.
Meeting Notes
JVET-AE0160 EE1-1.5: Optimization for complexity-performance trade-off of HOP network [R. Chang, L. Wang, X. Xu, S. Liu (Tencent)]

From BoG notes (recommendation confirmed by JVET):

This contribution reports the EE-1.5 result of JVET-AE0160. Based on the unified network architecture of the High Operation Point (HOP), this contribution proposes further optimization including new resblock and split network design to improve the complexity-performance trade-off of HOP network. Specifically, based on the depthwise separable convolution and group convolution, two novel resblocks are proposed to reduce the computational complexity without affecting the coding performance too much. Besides, the split network design for luma and chroma components is introduced into HOP network. Based on NNVC-5.1, the test results are shown in order of RA and AI configurations as follows.

Compared with NNVC-5.0 anchor:

Test 1.5.1 (Test designed resblocks with depthwise separable convolution and group convolution):

  1. Subtest 1: HOP resblock (as base comparison method for Test 1.5.1)

RA: -5.11% -14.62% -14.53% EncT: 129% DecT: 502%

AI : -3.06% -12.38% -13.54% EncT: 102% DecT: 439%

  1. Subtest 2: HOP res-block with depthwise separable convolution

RA: -4.82% -14.92% -14.18% EncT: 140% DecT: 640%

AI : -2.81% -12.01% -13.02% EncT: 106% DecT: 546%

  1. Subtest 3: HOP res-block with group convolution

RA: -4.97% -14.16% -13.79% EncT: 153% DecT: 822%

AI : -2.90% -11.79% -12.99% EncT: 110% DecT: 693%

Test 1.5.2 (Test split network design):

  1. Subtest 1: HOP network (as base comparison method for Test 1.5.2)

RA: -3.63% -12.05% -13.74% EncT: 139% DecT: 623%

AI : -2.28% -11.44% -12.15% EncT: 105% DecT: 535%

  1. Subtest 2: HOP network with split architecture for luma and chroma

RA: -3.12% -11.64% -11.47% EncT: 249% DecT: 2139%

AI : -1.96% -10.56% -11.48% EncT: 145% DecT: 1769%

Test EE1-1.5.1 Two ways of Back Bone Block simplification: 1) depth-wise separable convolution (~0.3% drop, 461🡪408 kMAC/pxl); 2) group (size 4) convolutions (0.1%...0.2% drop 461🡪402 kMAC/pxl).

Test EE1-1.5.2 Split for HOP architecture: 0.5% drop (480🡪316 kMAC/pxl).

It was recommended to study both aspects of JVET-AE0160 in EE1, using single modification of HOP.1 architecture in each sub-test, usage aspects and training strategy to keep the same as HOP.1.

Cited:
JVET-AE0255 X
ARCHIVE :
EE1-1.5.1-subtest1_vs_NNVC5.0_anchor.xlsm EE1-1.5.1-subtest2_vs_NNVC5.0_anchor.xlsm EE1-1.5.1-subtest3_vs_NNVC5.0_anchor.xlsm EE1-1.5.2-subtest1_vs_NNVC5.0_anchor.xlsm EE1-1.5.2-subtest2_vs_NNVC5.0_anchor.xlsm JVET-AE0160-v4.docx JVET-AE0160.pptx
Decisions
It was recommended to study both aspects of JVET-AE0160 in EE1, using single modification of HOP.1 architecture in each sub-test, usage aspects and training strategy to keep the same as HOP.1.
Citation