where is the prototxt file matched to imagenet models？

[liuwenran]

I've tried to test vgg16.v2.caffemodel with VGG16's test.prototxt in demo.py, but its output is incorrect. I wonder where is the corresponding prototxt file of vgg16.v2.caffemodel?

[sid027]

Hi, Were you able to find these prototxt files for the imagenet models?Did you write it yourself? I need them as well.

[liuwenran]

@sid027 I changed the num_output of "cls_score" layer to 201 and the num_output of "bboxes_pred" layer to 804 in train.prototxt. And I trained it on imagenet dataset.

[sid027]

is there a particular reason you used 201/804.....do you have 201 classes?

[liuwenran]

@sid027 because there are 200 classes in imagenet dataset ，and 201 classes with “background” ，804 = 201 * 4 are bounding boxes

[sid027]

one more question on the bounding boxes.Are they part of the training.Can I give 1000 instead of 804.Is there some specific rule that it needs to be four times?

[liuwenran]

because a bounding box is [xmin,ymin,xmax,ymax]，there are four numbers for a box. [xmin,ymin] is the left-top pixel location of this box in original image ,and [xmax,ymax] is the right-bottom.

[sid027]

ahh!thanks a lot.I will try this.

[sid027]

@liuwenran have a question of image database structure.Do I put each class in a subfolder?What about the background?

[liuwenran]

@sid027 you needn't to put each class in a subfolder because images will be shuffled before training,and background will be generated automatically.

[alilemus]

Hi @liuwenran! I've been trying to train this with the imagenet dataset, but have no idea on how to build the dataset (I have the JPEG files and Annotations, etc), but cant understand how to build the imdb.

any pointers?

[crazylyf]

I modified the corresponding numbers and run the demo, and found out that the result is very bad. what is the possible reason? Does anyone encounter such an problem? thanks!