All the codes and scripts are originated from Adil Moujahid's post in http://adilmoujahid.com/posts/2016/06/introduction-deep-learning-python-caffe/ and his github page found in https://github.com/adilmoujahid/deeplearning-cats-dogs-tutorial and slightly modified for my own caffe tutorial.
The repository is consist of 3 folder as follows
- caffe model: Contains definitions of train and solver prototypes for use in caffe
- input: Empty folder where all the dataset used for training, validtion and classification of unseen images.
- pycode: Includes python codes for making LMDB files, plot learning curve and make predictions.
In order to run the scripts and train with caffe, your machine needs to be set up as follows.
- OS: Ubuntu 16.04
- opencv: At least v3.0.0 or later is installed
- Caffe installed either GPU or CPU support: To check this. Run python and try import caffe and see if no errors found.
- Python modules: pip, numpy, lmdb, graphviz, pandas
If you need detailed description on how to install Caffe with GPU suport, see http://kyubot.tistory.com/93?category=617700
To learn how deep learning is work for image classifications, please see tutorial in http://kyubot.tistory.com/96?category=617700 To learn how to train deep learning model and use for pediction, please see my tutorial in http://kyubot.tistory.com/97?category=617700
All the commands that need to entered in terminal are in https://github.com/kyuhyong/deep_learning_caffe/blob/master/terminal_command
For cat and dog classification, you will need to download dataset as train.zip and test1.zip which can be found in https://www.kaggle.com/c/dogs-vs-cats/data Copy the above files into input folder and unzip as follows
cd ~/deep_learning_caffe
cd input
unzip ~/deeplearning-cats-dogs-tutorial/input/train.zip
unzip ~/deeplearning-cats-dogs-tutorial/input/test1.zip
rm ~/deeplearning-cats-dogs-tutorial/input/*.zip
Go to pycode folder and try run create_train_lmdb.py
cd ../pycode
python create_train_lmdb.py ~/deep_learning_caffe/input/
Generate mean image by running an app in caffe tools
cd ~/caffe/build/tools
./compute_image_mean -backend=lmdb ~/deep_learning_caffe/input/train_lmdb/ ~/deep_learning_caffe/input/mean.binaryproto
Modify path to mean_file, source in model defintion files under /caffe_models/caffe_model_1 per the name of your home directory.
To print model architecture, run a python script contained in caffe as follows
python ~/caffe/python/draw_net.py ~/deep_learning_caffe/caffe_models/caffe_model_1/caffenet_train_val_1.prototxt ~/deep_learning_caffe/caffe_models/caffe_model_1/caffe_model_1.png
To train the model run caffe tools as follows,
cd ~/caffe/build/tools
./caffe train --solver ~/deep_learning_caffe/caffe_models/caffe_model_1/solver_1.prototxt 2>&1 | tee ~/deep_learning_caffe/caffe_models/caffe_model_1/model_1_train.log
To print learning curve to see how well our model trained, run python script in project's pycode folder
python ~/deep_learning_caffe/pycode/plot_learning_curve.py ~/deep_learning_caffe/caffe_models/caffe_model_1/model_1_train.log ~/deep_learning_caffe/caffe_models/caffe_model_1/caffe_model_1_learning_curve.png
![alt text] (http://cfile22.uf.tistory.com/original/99705D4C5A60DB7805059D#jpg "Learning curve")
To predict image classifiction from test images, run python script as follows
$ python make_predictions_1.py --mean ~/deep_learning_caffe/input/mean.binaryproto --prototxt ~/deep_learning_caffe/caffe_models/caffe_model_1/caffenet_deploy_1.prototxt --model ~/deep_learning_caffe/caffe_models/caffe_model_1/caffe_model_1_iter_10000.caffemodel --test ~/deep_learning_caffe/input/test1/This will create a folder named predict under test folder and put classified images in a seperate folders.
To run realtime image classification with onboard camera, run python script as folows
$ python rt_classification.py --mean ~/deep_learning_caffe/input/mean.binaryproto --prototxt ~/deep_learning_caffe/caffe_models/caffe_model_1/caffenet_deploy_1.prototxt --model ~/deep_learning_caffe/caffe_models/caffe_model_1/caffe_model_1_iter_10000.caffemodel
For any request or feedback, send an email to kyuhyong [at] gmail [dot] com Thank you.