sumerintership-lip6-slides (1)

10
1/10 Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion Image classification using deep learning on a medium dataset made of cook recipes Remi Cadene under the direction Nicolas Thome and Matthieu Cord University Pierre and Marie Curie [email protected] November 25, 2015

Transcript of sumerintership-lip6-slides (1)

Page 1: sumerintership-lip6-slides (1)

1/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Image classification using deep learning on amedium dataset made of cook recipes

Remi Cadeneunder the direction Nicolas Thome and Matthieu Cord

University Pierre and Marie Curie

[email protected]

November 25, 2015

Page 2: sumerintership-lip6-slides (1)

2/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Table of contents

1 Introduction

2 Overfeat, a deep convolutional network

3 Transfer Learning Experiments on UPMC Food-101

4 Conclusion

Page 3: sumerintership-lip6-slides (1)

3/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Context

Since 2012, Deep ConvNet : AlexNet, Overfeat, Very Deep,GoogLeNetBenchmark dataset : ImageNet made of 1000 classes, 1.6 MimagesANR project (visiir), build a classifier from the dataset UPMCFood-101 to (90,840 images) classify photos of meals and togive the associated recipes

Page 4: sumerintership-lip6-slides (1)

4/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

UPMC Food-101

Figure: Category examples of UPMC Food-101 dataset we used in ourstudy.

Page 5: sumerintership-lip6-slides (1)

5/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Overfeat, a deep convolutional network

Figure: Overfeat accurate (144 Millions parameters)

Page 6: sumerintership-lip6-slides (1)

6/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Transfer Learning Experiments on UPMC Food-101

1 Features ExtractionMethod : 80% train, 20% test, 5 foldsNetwork : Overfeat without FC layers + SVMResult : 31% of correct classification top 1Training time : 48 hours (5 folds)

2 Fine TuningMethod : 80% train, 20% test, 1 foldNetwork : Overfeat with clean FC layersResult : 44.6% of correct classification top 1Training time : 8 hours (7 epochs)

Page 7: sumerintership-lip6-slides (1)

7/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Loss Fine Tuning

Figure: Loss processed during training time for each images on severalepochs of Overfeat fine tuned without data augmentation.

Page 8: sumerintership-lip6-slides (1)

8/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Other Experiments

3 From ScratchMethod : 80% train, 20% test, 1 foldNetwork : Overfeat from scratchResult : 34.9% of correct classification top 1Training time : 60 hours (55 epochs)

4 Data AugmentationMethod : 80% train * 10, 20% test, 1 foldNetwork : Overfeat with clean FC layersResult : 49.5% of correct classification top 1Training time : 60 hours (4 epochs)

Page 9: sumerintership-lip6-slides (1)

9/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Loss Fine Tuning with Data Augmentation

Figure: Loss processed during training time for each images of Overfeatfine tuned with data augmentation.

Page 10: sumerintership-lip6-slides (1)

10/10

Introduction Overfeat, a deep convolutional network Transfer Learning Experiments on UPMC Food-101 Conclusion

Conclusion

Pos Type Perf Test Training Time1 Fine Tuned (data*10) 49.5% 60 hours (4 epochs)2 Fine Tuned 44.6% 8 hours (7 epochs)3 From Scratch 34.9% 60 hours (55 epochs)4 Features Extract. (+SVM) 31% 48 hours (5 folds)

1 NotesFine Tuning is possible on a medium dataset10 ∗ training timesimple data >> training timedata∗10

10 ∗ epoch timesimple data < epoch timedata∗102 Future works

Transfer learning with more efficient networks : VeryDeep,GoogLeNetBetter hyperparameters optimization, batch normalizationlayers, implementing the Spatial Transformer module(DeepMind)