TResNet: High Performance GPU-Dedicated Architecture

Tal RidnikHussam LawenAsaf NoyItamar Friedman

   Papers with code   Abstract  PDF

Many deep learning models, developed in recent years, reach higher ImageNet accuracy than ResNet50, with fewer or comparable FLOPS count. While FLOPs are often seen as a proxy for network efficiency, when measuring actual GPU training and inference throughput, vanilla ResNet50 is usually significantly faster than its recent competitors, offering better throughput-accuracy trade-off... (read more)

Benchmarked Models

RANK
MODEL
REPO
CODE RESULT
PAPER RESULT
ε-REPRODUCED
BUILD
1
TResNet-XL
(input=448)
84.3%
84.3%
2
TResNet-XL
(input=448)
84.3%
84.3%
3
TResNet-XL-448
(input=448)
84.3%
84.3%
4
TResNet-XL-448
84.3%
84.3%
5
TResNet-L
(input=448)
83.8%
83.8%
6
TResNet-L
(input=448)
83.8%
83.8%
7
TResNet-L-448
(input=448)
83.8%
83.8%
8
TResNet-L-448
83.8%
83.8%
9
TResNet-M
(input=448)
83.2%
83.2%
10
TResNet-M
(input=448)
83.2%
83.2%
11
TResNet-M-448
(input=448)
83.2%
83.2%
12
TResNet-M-448
83.2%
83.2%
13
TResNet-XL-288
(Mean-Max)
82.8%
82.0%
14
TResNet-XL
(288-Mean-Max)
82.8%
82.0%
15
TResNet-XL
(288-Mean-Max)
82.8%
82.0%
16
TResNet-M
(512-mean-max)
82.4%
83.2%
17
TResNet-L
(288-Mean-Max)
82.0%
81.4%
18
TResNet-L
(288-Mean-Max)
82.0%
81.4%
19
TResNet-L-288
(Mean-Max)
82.0%
81.4%
20
TResNet-XL
82.0%
82.0%
21
TResNet-XL
82.0%
82.0%
22
TResNet-M
(448-mean-max)
82.0%
83.2%
23
TResNet-M
(288-Mean-Max)
81.6%
80.7%
24
TResNet-M
(288-Mean-Max)
81.6%
80.7%
25
TResNet-M-288
(Mean-Max)
81.6%
80.7%
26
TResNet-M
(288-Mean-Max-Squish)
81.4%
80.7%
27
TResNet-L
81.4%
81.4%
28
TResNet-L
81.4%
81.4%
29
TResNet-M
(input=288)
81.4%
80.7%
30
TResNet-M
(input=288)
81.4%
80.7%
31
TResNet-M-288-valzoom
81.4%
80.7%
32
TResNet-M
80.7%
80.7%
33
TResNet-M
80.7%
80.7%
34
TResNet-M
(288-Mean-Max-448W-Squ…)
80.3%
80.7%
35
TResNet-M
(224-Mean-Max)
80.1%
80.7%
36
TResNet-M
(288-Mean-Max-448W)
80.1%
80.7%
37
TResNet-M
(224-Mean-Max-vazoom)
80.0%
80.7%
38
TResNet-M
(288-Mean-Max-448W)
79.1%
80.7%
39
TResNet-M
(448W)
76.8%
80.7%