Hot dog image classifiers
This repo is created for using on colab notebook for university project hot_dog/not_hot_dog: https://colab.research.google.com/drive/16aS-kiqgxyGxw3jPp5H9lj6eLRHgHMGa#scrollTo=BlJxzLkbALkp
Kaggle competition: https://www.kaggle.com/c/hotdogornot/
- Python3.5+
- CUDA 10.1
- tensorflow 2.2.0
- keras 2.3.0+
- numpy 1.18+
The provided table shows comparative analysys. Test accuracy (0.3 of full train data).
Name | Accuracy | Estimated time, mins | AP |
---|---|---|---|
svc | 76.53% | 13 | 0.64 |
xgboost | 79.53% | 20 | 0.69 |
cnn - 20 epoc 256 batch 224x224 input | 80.91% | 7 | - |
cnn - 100 epoc 32 batch 224x224 input | 71.34% | 33 | - |
cnn - 20 epoc 256 batch 256x256 input | 81.91% | 10 | - |
cnn - 40 epoc 256 batch 256x256 input | 82.66% | 21 | - |
svm on vgg16 features** | 93.88% | <1 | 0.97 |
xgboost_on_vgg16_features** | 93.71% | <1 | 0.95 |
VGG16 * | 62.03% | ~53 | - |
Fast AI ResNet50 | 94.11% | ~9 | - |
* VGG16 training (full; not pretrained on ImageNet) takes long time (12 epochs taken), 64 batch. Accuracy raises very slow and error rate raises on validation after 9 epochs
** All VGG16 was pre-trained on ImageNet.
Name | Private score | Public score |
---|---|---|
svc_model_vgg16 | 0.91760 | 0.94230 |
xgb_model_vgg16 | 0.90151 | 0.92822 |
fast_ai_resnet50 | 0.95716 | 0.97115 |