xiaomengyc / i2c Goto Github PK

obj_simi.obj_valid_inds(mask, obj_inds)
这个函数在哪里啊，我没有找到

Greetings!
I am trying to reproduce the experimental results in Table 1. However, after fixing the bugs in these scripts and training the inception3_i2c with hyper-parameters provided in script/train.sh, I only get 'Top1: tensor(34.6140) Top5: tensor(59.7620)', which are far away from the 26.7, 8.4 in the last line of Table 1.
I think the reason behind such phenomenon lies in the hyper-parameters of training, as 5 epochs and no pretraining usually cannot guarantee the model convergence. But I cannot find any details about parameter setting in the paper. Could you give me some clues?
Hope for reply.

Hi,
Great thanks for your great work I2C.

When I ran the code, I met many problems. I have changed the code for training I2C on CUB-200-2011 dataset, however, I found that the obtained top1 and top 5 accuracies using the code in val_fram.py are surprisingly tensor(100,) and tensor(100.). Is that correct? And the code for computing the metric of Gt-Known, Top1 Loc, and Top5 Loc is not released in this repository.

Besides, there some bugs in this repository, for example, reduce_lr function used in

I would appreciate the author can release the complete training scripts for CUB-200-2011 dataset(including the hyper-parameters) and check the code whether it can successfully be executed on the device.

It seems that there are some files not provided in your project
1、Could you release the whole project ，or train_frame_crsimg_center_feat.py
2、 In the train_frame.py , you used my_optim.reduce_lr . Is your my_optim.py the same as it in other wsol project?
3、How can you train your model in CUB dataset , In your train.sh , the setting is --train_list='../datalist/CUB/train_list.txt' , label in it is not one-hot format， but your dataset file mydataset_crising.py satisfy the one-hot format gt ， is there something wrong in my thoughts?
Need your reply~

Hello.
I'd recently read your paper and want to try the code to run and see the result of it and understand the code.
But, for the some of the reasons, I had the trouble to run your code, and couldn't understand the structure of those.

1. In train or val function, the return parameters/data from the dataloader and initialized name couldn't match for me, so, I couldn't understand how the data come out
2. couldn't understand which parts of the code are working for the SC and GC
3. result image file save issue - no saving
4. validation accuracy issue
5. validation saving issue
   ... and some mores
   even to come out to run the validation and training I tried to edit little, but, I really want to see the your architectures' work and understand.
   If you didn't update your code, could you? or can you help me to at least the
   train_frame.py train function with dataloader.

Thanks,