crossmodalgroup / hrem Goto Github PK

Hi, I would like to enquire about the part of Neighbor Batch Sampling mentioned in the paper which seems to be missing from the code, can you tell why?

FileNotFoundError: [Errno 2] No such file or directory: 'data/f30k_precomp\train_caps.txt'

RuntimeError: mat1 and mat2 shapes cannot be multiplied (4480x4 and 2048x1024)

Hi @darkpromise98 , I notice this part in your paper is presented as follows:

I think your code I pinned on the top does not match with the paper. Should it correctly be as follows?

batch_sim_t2t = torch.matmul(cap_emb, cap_emb.t())
batch_sim_i2i = torch.matmul(img_emb, img_emb.t())

It seems like in this code repository, you've only implemented the Fusion mechanism architecture and there's no standalone mechanism, right?

你好，我在用mscoco数据集上运行train.py时，出现下面这个报错，请问是下载的数据格式不一样吗？难道网上公开的数据集格式不一样吗
Traceback (most recent call last):
File "/home/ubuntu/mydata2/ZFolder/HREM-main/train.py", line 282, in
main()
File "/home/ubuntu/mydata2/ZFolder/HREM-main/train.py", line 60, in main
train_loader = image_caption.get_train_loader(opt.data_path, tokenizer, opt.batch_size, opt.workers, opt)
File "/home/ubuntu/mydata2/ZFolder/HREM-main/lib/image_caption.py", line 184, in get_train_loader
train_loader = get_loader(data_path, 'train', tokenizer, opt, batch_size, True, workers, train=True)
File "/home/ubuntu/mydata2/ZFolder/HREM-main/lib/image_caption.py", line 169, in get_loader
dset = PrecompRegionDataset(data_path, data_split, tokenizer, opt, train)
File "/home/ubuntu/mydata2/ZFolder/HREM-main/lib/image_caption.py", line 27, in init
self.images = np.load(os.path.join(loc, '%s_ims.npy' % data_split))
File "/home/ubuntu/.local/lib/python3.9/site-packages/numpy/lib/npyio.py", line 413, in load
return format.read_array(fid, allow_pickle=allow_pickle,
File "/home/ubuntu/.local/lib/python3.9/site-packages/numpy/lib/format.py", line 785, in read_array
array.shape = shape
ValueError: cannot reshape array of size 699775200 into shape (113287,36,2048)

你好，请问这个项目训练和评估大概需要占用多大的空间？

Hello, I only found coco's id_mapping.json in the downloaded dataset, do you have flickr's id_mapping.json?

I don't see you using ResNet 101 and Fast RCNN as mentioned. Could you please provide me with clear information? I would really appreciate it.