valeoai / bf3s Goto Github PK

Hi @gidariss,

Can you please provide the config files for the other Networks.
For example, for mini-imagenet, only the wideresnet config files have been provided.

Thanks in advance

@gidariss Hi, very interesting work. I had a doubt regarding the training mechanism for novel classes. Do you freeze the feature extractor and fine tune on just classification layer (And do you add new classification head for the novel tasks)?

in the code, some codes say that load the miniImageNet_category_split_train_phase_train.pickle
miniImageNet_category_split_train_phase_val.pickle
miniImageNet_category_split_train_phase_test.pickle
miniImageNet_category_split_val.pickle
miniImageNet_category_split_test.pickle

but how can i find this files?

Hi @gidariss , right now I am reproducing the procedure on ImageNet-FS, and trying to train ResNet-10 from scratch by using the following command python scripts/train_fewshot.py --config=ImageNet/ResNet10CosineClassifierRotSelfsupervision --num_workers=8. However, according to my experiement, it seems that the loss of the base model cannot converge to a small number. More specifically, the figure for this will be about 2.0. In this case, I wonder if this result is normal. If not, there may be something wrong for my experiment.

when we use patch location prediction as auxiliary task:
feature extractor-->patch location task --> classifier
I wonder : can the gradient be backpropogated from classifier to feature extractor?(because utils.patch_location_task is not in the net graph)
thx!

The issue as shown in the title, look forward to hearing from you!

Hi!
I was trying to run the script:
python scripts/train_fewshot.py --config=miniImageNet/WRNd28w10CosineClassifierRotAugRotSelfsupervision

The 'AccuracyNovel' of the last few epochs seem to be very high, around 67%. I expected it to be around 62% as reported in the paper. I do not understand why. Could you please upload the training's log?

Here is my training's log:log.txt

Very interesting paper! Will the code be available soon?

Sorry to bother you !
I have a question: when we use patch location self supervsion, the patches pairs have different channels(6) from image(3), how can we use the same feature extractor?

@gidariss From the code it is difficult for me to understand how you are handling prototypical networks (not the cosine classifier). Config file for the prototypical network for any of the base networks (conv 4- 64, conv 4-512 ) will make this extremely clear. Please provide such a config file.

Hello, I am confused about how to get the confidence interval of the result, I run BF3S and I can only get a single accuracy rate at one time. Is the confidence interval obtained by running BF3S/scripts/test_fewshot.py many times using the same model ?How many times?

Hi,
I have cloned the code and run it as the instructions, but I find that even I set 4 visible gpu devices but it still use the first gpu in all that I set. And also I find that there isn't "DataParallel" object in the whole program. So I'm wandering if the network could be trained with multi-GPUs?
Many Thanks! And thank you for sharing the great work, I'm really interested in it.

Best,
Yifu