Hi,

Thanks a lot for sharing the nice code.

I am trying to reproduce the results. I notice that the provided Cifar-10 data has 4096 dimensions for each image, however the original Cifar-10 dataset downloaded here contains 3072 (32 * 32 * 3) dimensions per image. So I am wondering is there some additional steps applied on the original dataset in order to have 4096 dimensions? Is it possible to share such steps or maybe the ideas of such conversion?

Best wishes!

hello，Thank you for your excellent work！
return tf.pow(1 - normalized_dist, 1.4)
Excuse me, why should the formula for calculating the adjacency matrix be calculated this way? Does 1.4 have a special meaning? Thank you very much.

Since this approach is unsupervised I was a bit confused as to why do we need labels to make tfrecords file in make_data.py.
set_processor.py - line 7, 12, 19

I have an unlabelled dataset that I want to train on this model and I don't think it would be possible without labels

The model layers are shown in the following:

    self.encoder = encodec.Encoder(middle_dim, bbn_dim, cbn_dim)
    self.decoder = encodec.Decoder(middle_dim, self.feat_dim)
    self.tbn = twin_bottleneck.TwinBottleneck(bbn_dim, cbn_dim)
    self.dis_1 = tf.keras.layers.Dense(1, activation='sigmoid')
    self.dis_2 = tf.keras.layers.Dense(1, activation='sigmoid')

The WAE Adversarial Block Layer: cbn->1024->1 & bbn->1024->1
Maybe you lack a FC layer with 1024 nodes in tbh.py?

the file'dataset.py' in util.data, in the main function, data = Dataset(set_name='cifar10', batch_size=256) it = iter(data.test_data) print(it) x = next(it) print(x) , it can error, the error is 'UnicodeDecodeError: 'utf-8' codec can't decode byte 0xd5 in position 103: invalid continuation byte', can you provide the origin file 'cifar10_fc7_train.mat' and 'cifar10_fc7_test.mat'

Hi!
Thanks for your excellent works!

I just wonder that could you provide the preprocessing program? Since I cannot re-implement the MAP score at MSCOCO and NUS-WIDE.

Thanks a lot!