Download to read offline
Trial for Practical NN Using
- 1. Trial for PracticalNN Using RTSS JUN YOUNG PARK
- 2. Approach Get • Get &Preprocess dataset Select • Select the model for learning Trim • Trim the model for appropriate perception
- 3. Prepare Images my_imgs train validation food face misc … • 3Classes (Each 100 images) • Use image data divider that I made before • Train 70%, Validation 30% < Face > < Food > < Misc >
- 4. Small Change onExample ◦ Change location of image dataset. ◦ Change model and number of its classes of FC layer. Fine-Tuning (Don’t freeze other layers) Freeze other layers to train only FC layer
- 5.
- 6. Classification Result (Convas Extractor) Accuracy : 90.4% Mismatch !
- 7. Why mismatched? Food tickets(Human may predict it ‘misc’) A train ticket on beverage A cup of tea and text ./training/food
- 8. AlexNet ◦ Alexnet Get numberof input features Replace FC layer Finetune Fixed
- 9. AlexNet Result Accuracy :86% (-8.8%) Accuracy : 89.6% (-0.8%)
- 10. Try other models ◦Resnet 152 ◦ Densenet 201 <Fine Tune> <Fixed> <Fine Tune> <Fixed>
- 11. Common Problem < Densenet205 > ◦ Accuracy is hovering around between 90~95% ◦ Why ? ◦ Few training data (70 images / a class) ◦ Need hyperparameter tuning (Batch size, Learning Rate … ) ◦ Some images make the model confused
- 12. Self Test ◦ TransferLearning 을 위해 기존 모델을 적절히 수정할 수 있다. ◦ Learning Rate 와 Batch Size 등의 Hyperparameter 등을 조정할 수 있다. ◦ 다양한 모델에 사용자 데이터 셋을 효율적으로 훈련시킬 수 있다.