Training my own model - any good tutorials apart from the official Caffe ImageNet example?

[georgi angelov]

So I've been researching how to train my own model from a set of images but I cannot seem to find any pertaining to that. I found this official Caffe tutorial:

http://caffe.berkeleyvision.org/gathered/examples/imagenet.html

but it seems like it is related to ImageNet dataset, if I am getting it right.

Does anyone have any good tutorial on training your own model?

Thanks

[Antonio Paes]

I'm searching too..

[georgi angelov]

Yeah it seems like it's a pretty important part of the framework but I can't seem to find anything related to training your own model.

[Nikiforos Pittaras]

Well you'll have to follow similar steps to prepare your data as in the imagenet example.

Which step are you having difficulty in?

[Antonio Paes]

In my case, how I generate my files some lmdb?

and what i have change to use my own images?

[Daniel Seita]

Antonio,

To generate your own data, I suggest: https://github.com/BVLC/caffe/issues/747

Basically, the steps are: get all your images, label them and split them up into train/testing (e.g., using Python scripts you create), then use Caffe's tool to convert it to lmbd format. Sadly there isn't much documentation about this, but the code itself, convert_imageset.cpp has comments in it and some sample usage.

-Daniel

[Antonio Paes]

Thanks Daniel, 

I managed to make the prepraracao data, I have problems in training now, but seems to be the problem cuda, as I'm running on a server with linux arch, I asked the server administrator to update the cuda, I'm waiting now.

thank you Daniel.

[georgi angelov]

Daniel,

What I am trying to figure out is what is the validation data ? What is the difference between the images used for training and the images used for validation ?

[Antonio Paes]

So according to Article Krizhevsky the challenge IMAGEnet has two types of evaluation with images of validating and testing, I imagine that the example set by caffe follows this challenge, but why exactly are these three types actually neh, training, validating and testing. Tests even make sense but that validating and I think something specific just the IMAGEnet challenge, but I'm not sure. 

Good Question LoL.

[georgi angelov]

So how did you train your model, Antonio? Did you just use one set of images in a folder ?

[Antonio Paes]

I am using two sets of images, one for testing and one for training. But I have some problems with CUDA.

[georgi angelov]

Yeah that is what I am wondering. What are your testing images and what are your training images ? Is the test images just the original images with boxes around the objects you are training for ?

[Antonio Paes]

I am so doing man, I'm using 7000 images for training and 10,000 for testing, without bounding box only images even. And to do that I'm following the tutorial:

  http://caffe.berkeleyvision.org/gathered/examples/imagenet.html

Is giving some problems, but I believe it's because of the network architecture, as my pictures are smaller than those set by default, will have to make some changes.

[georgi angelov]

So what I am wondering, what are you testing images? Are they just the same as your training images?

[Antonio Paes]

In my case the content is the same, but in different situations (others positions, with parcial oclusion, more lighting etc.).