Closed atiszabo closed 3 years ago
Hi @atiszabo ,
For Problem 1), thanks for spotting this issue, I've updated the README with the instruction and file links for downloading these two json files.
For Problem 2), I suggest to check the json files you've downloaded, or delete the https://github.com/YoungXIAO13/FewShotDetection#coco
file and re-run the code to re-generated this file, this should give you at least hundreds of images in the case of 10-shot.
@YoungXIAO13 : thanks for being responsive and for maintaining your repo!
Hi,
congratulations on the nice paper and repository! I am trying to reproduce the results on COCO with @hjraad, but I've ran into two problems:
Problem 1) Seems I don't have 2 annotations files. I've tried to run train_coco.sh when got this error on two missing files:
I've downloaded COCO following the README.md under data, I obtained these files from the annotations:
How should I get the missing files? Do I miss anything? Thanks in advance
Problem 2) I tried to run finetune_coco.sh as well with the provided base checkpoint, but it also got some problems with loading the data. Somehow it only loads 2 images wheres it should do 10 shot for multiple classes, which I guess should be more. Please see below:
Namespace(TFA=False, batch_size=4, checkepoch=20, checkpoint=21985, checkpoint_interval=10000, checksession=200, class_agnostic=False, cuda=True, dataset='coco', disp_interval=100, fix_encoder=False, log_dir='checkpoint', lr=0.001, lr_decay_gamma=0.1, lr_decay_step=4, max_epochs=30, meta_loss=True, meta_train=True, meta_type=0, net='metarcnn', num_workers=8, optimizer='sgd', phase=2, resume=True, save_dir='save_models/COCO', session=1, shots=10, start_epoch=1, use_tfboard=True)
[...] Loaded dataset
coco_2014_shots
for training [...] before filtering, there are 2 images... after filtering, there are 2 images...Any clues on this? @YoungXIAO13
Thanks!