Table 5 Hyperparameter configurations of compared models.

From: Deep multi-task learning framework for gastrointestinal lesion-aided diagnosis and severity estimation

Model

Batch size

Optimizer

LR

Epochs

Activation

Dropout rates

Layers

Number of Heads

Embedding dimension

Patch size

Baseline

16

Adam

1e − 4

100

ReLU

0.1

12

8

384

16 × 16

CrossViT19

16

AdamW

3e − 4

100

GeLU

0.1

16

8

192 and 384

Multi-scale

ViTfSCD30

 

Adam

1e − 4

 

GeLU

0.1

24

16

512

16 × 16

MedViT20

32

Adam

1e − 4

100

ReLU

0.1

12

–

768

16 × 16

FLATer16

16

Adam

1e − 3

300

ReLU

0.1

12

–

512

16 × 16

CST31

8

SGD

1e − 4

80

ReLU

0.1

12

–

512

16 × 16

AG-CNN18

32

SGD

1e − 3

50

ReLU

0.1

–

–

–

–

GTCAD

16

AdamW

1e − 4

100

ReLU

0.1

12

8

384 and 512

16 × 16