The source code is running, the experimental process is recorded as follows, for beginners to get started.
Today and elder sister to run through, to share the next experience. (Pre-Training network: ImageNet, Training set: PASCAL VOC2007, GPU)
First, the entire train and test process is not unique, and the deeper you understand it, the more skilled you are.
Come down and get to the point:
1.git Clone source code. Be sure to choose recursive mode. (No Caffe the package is not in the source code, the compilation will be error)
2. Go to the Lib folder and make a click.
3. Down in the Caffe directory, CP Makefile.config.example Makefile.config
It is to be seen that a new folder appears, called Makefile.config
4. Modify the contents of the Makefile.config file (this step is very important, there are many tutorials on the web, be sure to rely on their own dependent package path of the relevant content, must not directly tiger)
5. This step will be downloaded from the Web to the Pascal VOC2007 DataSet extracted from the folder, into the root directory of the data folder.
6. Then download the Web to the Imagenet model for pre-training, and unzip the folder into the Data folder.
7. It's time to compile Caffe, go to Caffe directory, make all, make test, make runtest, make Pycaffe.
This step, if the error is to check whether the pre-preparation work is not done, if it is strange, you can skip make next (I skipped the middle of two steps, the elder sister skipped the second one, why not understand)
8. After the first 7 steps are successful, enter the Linux command for training, I did not change the number of iterations, "80000,4000,80000,4000" and then about 8 hours of training (workstations have a GPU), you can first use the "100,100,100,100" experiment.
9. After successful training, the Caffemodel suffix file, together with the test picture as input, run the demo command, it is successful.
Summary:
1. The entire demo run down is not easy, but you have experimented with fast rcnn, should be very quick to get started.
2. We ran a family use case, our own data set can be made into its format, and then apply the code on it. Standing on the shoulders of giants.
3. Well, we are not high level, the relevant fields of friends, welcome to discuss the discussion.
Caffe Study Notes (i), Ubuntu14.04+gpu (using Pascal VOC2007 training data, and testing)