Kaggle dogs vs cats redux on flyodhub
ResNet50 as base model with different top layers.
Result: scored in the top 8% on public LB (ensembling in avg_subm.ipynb).
ResNet50, inceptionV3 adn Xception as base models with simple fully-connected top layers.
Result: scored within top 2% on public LB.
Average test image predictions using k-Nearest-neighbors.
Comparisons between optimizers: SGD, Adam Nadam and RMSprop inspired by this paper on arxiv. Result shows that SGD with momentum has better performnace on validation data than adaptive optimizers. See this blog post for detail.
- Keras 1.2.2
- Tensorflow 1.0