Efficient Image Captioning for Edge Devices: Knowledge Distillation

Table of Links

Abstract and 1 Introduction

2 Related Work

3 Methodology and 3.1 Model Architecture

3.2 Model Training

3.3 Knowledge Distillation

4 Experiments

4.1 Datasets and Metrics and 4.2 Implementation Details

4.3 Ablation Study

4.4 Inference on the Mobile Device and 4.5 State-of-the-art Comparison

5 Conclusion and References

A Implementation Details

B Visualization Results

C Results on Nocaps

D Limitations and Future Work

3.3 Knowledge Distillation

Ensemble KD. Actually, instead of adopting a single head, we construct the ensemble head with three parallel branches. We train three teacher models with different model initializations. These teachers jointly distill different branches of the ensemble head model, as shown in Figure 3 (right).