This repo provides my work on flower image classification using deep convnet. I trained serveral models on VGG 102 flower dataset.
The best top-1 accuracy achieved so far: 0.904.
Note here that I follow the official protocol for dataset split.
- training: 1020
- validation: 1020
- test: 6149
- Study the transfer learning on mulitclass, fine grained image classification task.
- Study the visualization tools and techniques for deep convnets.
- Try to explain the convnet-based classifier using visualization.
- Preparation
python init.py # download the dataset and organize- Training
python train.py --model=[model_name]python 2.7
keras 2
(to be completed...)
A simple 2-layer baseline convnet.
acc: 0.312
Fine tune VGG16 with weights pretrained on imagenet.
acc: 0.757
Fine tune VGG19 with weights pretrained on imagenet.
acc: 0.781
Fine tune inception-v3 with weights pretrained on imagenet.
acc: 0.904
