evelinehong / slot-attention-pytorch Goto Github PK

Thanks for your code!

The dataset.py seems to be missing. In the meanwhile I'll try to adapt https://github.com/vahidk/tfrecord/ to read directly DeepMind's https://github.com/deepmind/multi_object_datasets#clevr-with-masks

The gradient with respect to the slots_mu and slots_sigmavariables is zero. To learn the initialization of slots, you could change your model.pyin line 40 to slots = torch.distributions.Normal(mu, sigma).rsample()... with rsample()the gradients will flow for these variables

Thanks for your implementation of Slot Attention module. However, I found that the sampling operation (in Line 40 at model.py) prevents gradients from the back-propagation. During training, the gradients of slot_mu and slot_sigma will be zero, which means the two variable will not change. I think the reparameterization trick is needed to make the sampling operation differentiable.

Hi,
Just wanted to confirm if shape of image is
[batch_size, num_channels, width, height],

usually it's
[batch_size, num_channels, height, width].

Thanks and regards

Hello and thank you for the great work.
I think the eval file misses the pre-trained model ('./tmp/model3.ckpt') as indicated in the error I received. Would you please show me how I can have it?
Thank you in advance