keras-aichallenger-2018-plant-recognition
thoughts
I have tried out various kinds of models, and use data argumentation operations, but when I reach to 0.87993 precision, I can't improve it anymore, but according to other competitors' experience, pytorch deep learning framework can get a higher percision, since I have tried out many keras models, I decide to make my code open source, any questions, contact me, please
environments
tensorflow-gpu 1.9.0
jupyter 1.0.0
Keras 2.1.0
h5py 2.7.1
ubuntu 16.04
gtx 1080ti
models and validation accuracy
- inception_resnet_v2 0.8764
- xception 0.8740
- mobilenet_v2 0.8674
- inception_v3 0.8747
- vgg16 failed
- vgg19 failed
- resnet50 0.8762
- inception_v4 0.8700
- resnet34 0.8773
- densenet121 didn't try
- densenet161 0.8800
- shufflenet_v2 0.8676
- resnet_attention_56 0.8762
- squeezenet 0.8026
- seresnet50 0.8826
- se_resnext didn't try
- nasnet failed
- custom didn't try
- resnet18 0.8754
tutorial
datasets
- at first, you need to download aichallenger plant datasets from the official site:
https://challenger.ai/competition/pdr2018
since the official download is not available, I make the datasets public, download websites (baidu cloud):
url: https://pan.baidu.com/s/15csF8MyJaROQo5Hgt8sLYg
extraction code: 16yc
- for dataset generation, please refer to data_split.ipynb
it's easy to modify for your purpose
training
I provide various kinds of models for training, if you want to use one of these models, please refer to the train.py, I provide the example code here
python train.py --model_name=resnet34
inference
before you use the following command, you should manually add one line code in the reference.py, to load your trained model, for example:
model=load_model('./trained_model/resnet34/resnet34.54-0.8773.hdf5')
python inference.py
results combinations
please refer to the
- create_csv_results.py
- csv_combinations.py
grade
-
densenet161+inception_resnet_v2+seresnet50+resnet50+resnet34: 0.8817
-
我这里就当做baseline开源了
contact
any questions, please email me, my email address is: w5688414@outlook.com