Tensor Regression Networks - Space Savings Calculation 📊

Calculations of the GLOBAL space savings for Table 1 and Table 3 from the paper:

Tensor Regression Networks by Kossaifi et al. (2020)

🚀 How to Run

pytho3 main.py

This will generate the space savings tables for both ResNet101 (TRL) and VGG19 (TCL) architectures.

📈 Results

Table 1: ResNet101 with Tensor Regression Layers (TRL)

ResNet101 baseline: 44,549,160 parameters | FC head: 2,049,000 parameters (4.6% of total)

Rank	Space Savings (%)	Total Space Savings (%)
(300, 1, 1, 700)	25.60	1.18
(200, 1, 1, 200)	68.30	3.14
(120, 1, 1, 300)	71.61	3.29
(150, 1, 1, 150)	76.59	3.52
(100, 1, 1, 100)	84.64	3.89
(50, 1, 1, 50)	92.44	4.25

Table 3: VGG19 with Tensor Contraction Layers (TCL)

VGG19 baseline: 143,667,240 parameters | Classifier head: 123,642,856 parameters (86.1% of total)

Rank	Space Savings (%)	Total Space Savings (%)
(512, 7, 7)	-0.21	-0.18
(384, 5, 5)	65.87	56.69