Table 1 Training and test set performances.

From: Sequential transfer learning based on hierarchical clustering for improved performance in deep learning based food segmentation

 

U-Net

DeepLab

Conventional training approach

Sequential Transfer Learning

Absolute difference

Conventional training approach

Sequential Transfer Learning

Absolute difference

Training duration

3750min (62.5h)

1880min (31.2h)

1870min (50%)

3600min (60h)

2320min (38.67h)

1280min (36%)

Training performance

Loss

3.4229

4.0085

0.5856 (14%)

3.0793

3.9207

0.8414 (\( 21\% \))

Accuracy

0.9427

0.9347

0.008 (0.8%)

0.9439

0.9272

0.0167 (1.8%)

Validation performance

Loss

4.3396

5.3099

0.9703 (\(18\%\))

6.6332

7.3692

0.736 (10%)

Accuracy

0.9243

0.9184

0.0059 (\(0.6\%\))

0.8946

0.8750

0.0196 (\(2.2\%\))

Test performance

Loss

15.4667

9.4800

5.9867 \((39\%)\)

8.9102

8.1726

0.7376 \((8.3\%)\)

Accuracy

0.8269

0.8829

0.056 \((5.6\%)\)

0.8116

0.8360

0.0244 \((2.9\%)\)

  1. Network performance comparison table, holding training, validation and test set performance information evaluated for both network architectures, U-Net and DeepLab. The table depicts values of the overall training duration, loss and accuracy levels. Leading results are given in bold.
Back to article page