Datasets:
unstructuredio
/
SCORE-Bench

Dataset card Data Studio Files
xet
Community
2
text
stringlengths
0
2.18k
--------------------------------------------------- Unstructured Formula End
(11)
Where F′′ₐₜₜₙ and F′′ ffn are the corresponding FLOPs scales after selecting the largest value in {𝛼a l , 𝛼m l , 𝛼h l , 𝛼n l } in Eq. 10 respectively. For both stages, dynamic compression is trained with group-level gates, and the γ is set to 100.
--------------------------------------------------- Unstructured Page Number Block Begin
8
--------------------------------------------------- Unstructured Page Number Block End
--------------------------------------------------- Unstructured Plain Text Format 1
For group-level strategy, we trained the USDC method by splitting sub-groups on average, on random, and our recursive split method separately. The training is on the DeiT-small model on the Imagenet-1K dataset, the training batch size is 256 and all other parameter settings are the same. As shown in Tab. 5, we compare different sub-groups split method, the The top-1 accuracy on Imagenet-1K of our recursive split method separately is better than the average and random methods.
Table 5: The comparisons of different sub-groups split methods for group-level gate augmentation strategy. The first column splits sub-groups uniformly with step size 32. The second column splits sub-groups uniformly with step size 8. The third column splits the sub-groups randomly with step size ranges in [1,64]. The fourth column (Ours) splits sub-groups recursively by a logarithm of 2.
--------------------------------------------------- Unstructured Table Begin
Model B Top-1 Accuracy (%)
Avg-32 Avg-8 Random Ours
USDC (DeiT-S) 256 77.13 77.43 77.86 78.89
64 77.11 77.62 77.44 78.90
32 77.10 77.62 77.45 78.93
8 77.05 77.65 77.50 78.96
2 77.00 77.62 77.49 78.98
1 76.91 77.60 77.49 78.96
FLOPs - 3.30G 3.36G 3.30G 3.35G
--------------------------------------------------- Unstructured Table End
--------------------------------------------------- Unstructured Title Begin
E. Visualizations
--------------------------------------------------- Unstructured Title End
We illustrate the structures of the compressed DeiT-Small model by USDC at Fig. 5. We trained the model in Fig. 5 by unified static and dynamic compression described in the main text. We can notice that the head number of MHSA and the hidden dimension of FFN were reduced by static compression, and some blocks were pruned by static compression of USDC. Meanwhile, the dynamic compression of USDC skipped each block adaptively according to the input features of each transformer layer. The FLOPs of all 12 dynamic decision networks together are only 0.45M, and the FLOPs of the original DeiT-small is 4.6G. As shown in Fig. 5, the remaining FLOPs achieved by only the static compression part is 74.9%. The final remaining FLOPs achieved by joint static and dynamic compression is 64.8%.
--------------------------------------------------- Unstructured Image Begin
L=1
FC+LN +ReLU+FC
R=0.75
head=4
R=0.96
h_dim=1276
L=2
FC+LN +ReLU+FC
R=0.92
head=4
R=0.98
h_dim=1319
L=3
Conv1d
R=0.95
head=4
R=0.89
h_dim=1394
L=4
Conv1d
Pruned
R=0.84
h_dim=1306
L=5
FC+LN +ReLU+FC
R=1.0
head=6,
R=0.81
h_dim=1302
L=6
Conv1d
R=0.91
head=6,
R=0.85
h_dim=1249
L=7
FC+BN +ReLU+FC
R=0.78
head=6,
R=0.66
h_dim=1409
L=8
FC
R=1.0
head=6,
Pruned
L=9
FC
R=1.0