A project built for kaggle competition distracted drivers detection.
- python 2.7
- python library:
in the requirements.md
to install pip packages use :
pip2.7 install -r pip_packages.txtinstall tenserflow(tensorflow==0.8.0) from here - Linux or unix System, may be windows
- move
config.tmpl.pytoconfig.py, you can use *unix command linemv config.tmpl.py config.py, and update the variable in this file - run the
main.pybypython2.7 main.py
- move
config.tmpl.pytoconfig.py, you can use *unix command linemv config.tmpl.py config.py, and update the variable in this file - change your dir to
CNN/vgg_try_karea, then train model by runningpython2.7 train_model.py, evaluate model by runningpython2.7 evaluate_model.py, predict test set by runningpython2.7 predict_model.py
- in the result folder, you can see some file end with
.csv - the cache folder, you can see cache file
#Idea
#实验结果记录
| submit date | name | of-los | on-los | feature | model | other trick | comments |
|---|---|---|---|---|---|---|---|
| 2016-05-13 | liu zheng | - | 4.4707 | 6 conv layers cnn | cnn | mirror,rotate,resize 64x64 | what a shame.... |
| 2016-05-15 | chenqiang | - | 14.* | 9600 hog feature | forest | no | it must be over-fitting |
| 2016-05-15 | chenqiang | - | 2.3025 | all 0.1 | no | base line | |
| 2016-05-19 | chenqiang | - | 1.6647 | 9600 hog feature | forest | forest with probability | still have a huge space to improve |
| 2016-06-01 | chenqiang | 1.37 | 1.59 | vgg fine-tuning | cnn | replace last layer 1000 node, to 10 node. | still have a huge space to improve |
| 2016-06-01 | chenqiang | 1.09 | 1.34 | vgg fine-tuning | cnn | replace last layer 1000 node, to 10 node. | still have a huge space to improve |
| 2016-06-01 | chenqiang | 0.10 | 1.26 | vgg fine-tuning | cnn | strange, only 2 epoch | still have a huge space to improve |
| 2016-06-01 | chenqiang | 0.03 | 1.41 | vgg fine-tuning | cnn | strange, 12 epoch | over-fitting |
| 2016-06-01 | chenqiang | 0.9 | 1.28 | vgg fine-tuning | cnn-vgg | add one softmax layer to vgg |