|

Explorer

Filters

234 total runs

Dataset

ADE20k val

Semantic Segmentation

BraTS

Semantic Segmentation

ImageNet

Image Classification

OpenWebText

Language Modeling

Baseline

ResNet-50, 16-bit Optimized (AWS A100)

Model

COMPARE

ResNet-50

234 runs

Recipe

COMPARE

MosaicML Mild

48 runs

MosaicML Medium

60 runs

MosaicML Hot

66 runs

MosaicML Baseline

30 runs

FFCV

6 runs

Pytorch Blog

6 runs

timm a1

6 runs

timm a2

6 runs

timm a3

6 runs

Cloud

COMPARE

Amazon Web Services

76 runs

Google Compute Platform

76 runs

MosaicML Cloud

44 runs

Oracle Cloud Infrastructure

38 runs

Hardware

COMPARE

Nvidia A100 40GB

114 runs

Nvidia A100 80GB

44 runs

Nvidia V100 16GB

76 runs

Methods

Binary Cross Entropy Loss

174 runs

Blurpool

180 runs

Channels Last

174 runs

Colout

66 runs

Exponential Moving Average

174 runs

FFCV

114 runs

FixRes

174 runs

Label Smoothing

180 runs

Mixup

126 runs

Progressive Resizing

180 runs

RandAugment

66 runs

Sharpness-Aware Minimization

126 runs

Scale Schedule

180 runs

Stochastic Depth

66 runs

Show 9 more

Top-1 accuracy

Time

Cost

Training efficiency frontier

Top-1 accuracy

Cost

Frontier
Selected
Baseline

Time

Top-1 accuracy

Cost

Top-1 accuracy

Frontier
Selected
Baseline

Recipe

MosaicML Mild

Methods

9 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Progressive Resizing
Scale Schedule(0.36)

Cost: $11.04

Top-1 accuracy: 76.64%

Time: 27m 09s

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.36

Recipe

MosaicML Mild

Methods

9 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Progressive Resizing
Scale Schedule(0.4)

Cost: $12.42

Top-1 accuracy: 76.99%

Time: 30m 33s

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.40

Recipe

MosaicML Mild

Methods

9 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Progressive Resizing
Scale Schedule(0.5)

Cost: $15.53

Top-1 accuracy: 77.51%

Time: 38m 11s

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.50

Recipe

MosaicML Mild

Methods

9 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Progressive Resizing
Scale Schedule(0.75)

Cost: $23.12

Top-1 accuracy: 78.13%

Time: 56m 51s

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.75

Recipe

MosaicML Medium

Methods

11 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Mixup
Progressive Resizing
Sharpness-Aware Minimization
Scale Schedule(0.75)

Cost: $25.34

Top-1 accuracy: 78.40%

Time: 1h 02m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.75

Recipe

MosaicML Medium

Methods

11 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Mixup
Progressive Resizing
Sharpness-Aware Minimization
Scale Schedule(1)

Cost: $34.05

Top-1 accuracy: 78.95%

Time: 1h 23m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
1.00

Recipe

MosaicML Medium

Methods

11 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Mixup
Progressive Resizing
Sharpness-Aware Minimization
Scale Schedule(1.5)

Cost: $51.07

Top-1 accuracy: 79.38%

Time: 2h 05m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
1.50

Recipe

MosaicML Medium

Methods

11 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Mixup
Progressive Resizing
Sharpness-Aware Minimization
Scale Schedule(2)

Cost: $68.09

Top-1 accuracy: 79.50%

Time: 2h 47m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
2.00

Recipe

MosaicML Medium

Methods

11 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Mixup
Progressive Resizing
Sharpness-Aware Minimization
Scale Schedule(2.65)

Cost: $90.03

Top-1 accuracy: 79.51%

Time: 3h 41m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
2.65

Recipe

MosaicML Hot

Methods

13 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Colout
Exponential Moving Average
FixRes
Label Smoothing
Mixup
Progressive Resizing
RandAugment
Sharpness-Aware Minimization
Stochastic Depth
Scale Schedule(2)

Cost: $101.36

Top-1 accuracy: 79.65%

Time: 4h 09m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
2.00

Recipe

MosaicML Hot

Methods

13 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Colout
Exponential Moving Average
FixRes
Label Smoothing
Mixup
Progressive Resizing
RandAugment
Sharpness-Aware Minimization
Stochastic Depth
Scale Schedule(2.2)

Cost: $111.49

Top-1 accuracy: 79.86%

Time: 4h 34m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
2.20

Recipe

MosaicML Hot

Methods

13 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Colout
Exponential Moving Average
FixRes
Label Smoothing
Mixup
Progressive Resizing
RandAugment
Sharpness-Aware Minimization
Stochastic Depth
Scale Schedule(3)

Cost: $152.04

Top-1 accuracy: 80.31%

Time: 6h 13m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
3.00

Recipe

MosaicML Hot

Methods

13 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Colout
Exponential Moving Average
FixRes
Label Smoothing
Mixup
Progressive Resizing
RandAugment
Sharpness-Aware Minimization
Stochastic Depth
Scale Schedule(4)

Cost: $202.71

Top-1 accuracy: 80.56%

Time: 8h 18m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
4.00

Recipe

MosaicML Hot

Methods

13 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Colout
Exponential Moving Average
FixRes
Label Smoothing
Mixup
Progressive Resizing
RandAugment
Sharpness-Aware Minimization
Stochastic Depth
Scale Schedule(5)

Cost: $253.39

Top-1 accuracy: 80.78%

Time: 10h 23m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
5.00

Recipe

MosaicML Hot

Methods

13 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Colout
Exponential Moving Average
FixRes
Label Smoothing
Mixup
Progressive Resizing
RandAugment
Sharpness-Aware Minimization
Stochastic Depth
Scale Schedule(7.5)

Cost: $380.09

Top-1 accuracy: 80.98%

Time: 15h 34m

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
7.50

Recipe

MosaicML Mild

Methods

9 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Progressive Resizing
Scale Schedule(0.32)

Cost: $9.66

Top-1 accuracy: 76.19%

Time: 23m 45s

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.32

Recipe

MosaicML Mild

Methods

9 Methods

Binary Cross Entropy Loss
Blurpool
Channels Last
Exponential Moving Average
FFCV
FixRes
Label Smoothing
Progressive Resizing
Scale Schedule(0.25)

Cost: $7.59

Top-1 accuracy: 75.17%

Time: 18m 40s

Model
ResNet-50
Cloud
Oracle Cloud Infrastructure
Instance
BM.GPU4.8
Scale Schedule Ratio
0.25

© 2022 MosaicML

Terms Privacy Policy