Bo Zhang

Could we input the training set without scale operation? The most of CNN models process the images without scale operation. If or not scale operation for input images improves the...

How many epochs do we need for training MobileNet from scratch on ImageNet?

I think you can try to solve your non-convergence problem from BN layer. Trying to set the batch_norm_param { use_global_stats: true }

Please refer to [https://github.com/BOBrown/CCDA_LGFA/blob/3ba451760fc2456928a88da3a5c9e04645d98f58/CCDA/entropy.py#L237](url)

@lifeiove The step (3) refers to that each anchor should be represented by a vector with the same dimension as the center_features (you should set this parameter in train.prototxt). So,...

@NovaBurst I have uploaded the center_loss_train.prototxt file and center_loss_test.prototxt file. And you can configure your prototxt file according to my example files. And I think center_loss is used to offer...

I often use 1.0 as the cener_loss_weight. The default value is written in the caffe.proto file. You can set this parm in the train.prototxt file.

@weiweili123 Yes, I trained SSD with center loss on our own datasets, similar to the MRI scanning image. The performance is better than the SSD.

@yja1 mbox_center_conf is the concatenation of all the center_features in all default box. And your center_features are set to 23. Therefore, the convolution output number should be 23 * anchors_numbers.

You can obtain the easy_split.txt and hard_split.txt files by performing the entropy evaluation as follows: python entropy.py --lambda1 0.5 Thank you for your interest in our work.