Justin Johnson
Justin Johnson
No; if you want the images you'll need to download them. The scene graph representation is not quite enough to reconstruct the images, since it does not store the random...
This is not currently implemented, but it wouldn't be hard to do. During training the RNN learns a language model, which is able to assign a probability to an arbitrary...
Looks like a problem with the HDF5 library. Make sure you've installed the correct one: https://github.com/deepmind/torch-hdf5
Does the following command work? ``` bash th -e "require 'hdf5'" ```
To make HDF5 work you also need to install the C library; the lua package is just a wrapper. For example on Ubuntu you need to run ``` bash sudo...
@gp-huang Sorry, I'm not sure how to do that - I've never tried to install locally. @nylki Can you take a look at the file that is throwing a syntax...
@nylki We are in the territory of debugging [torch-hdf5](https://github.com/deepmind/torch-hdf5) now, which I don't know much about; you might have better luck opening an issue over there instead. But what is...
I didn't do any super controlled experiments, but deeper models with batchnorm do seem to train faster.
Those are some beautiful curves - thanks for sharing! It's interesting that the models without BN show a plateau at the beginning - that suggests to me that maybe the...
This isn't currently supported but it would be relatively easy to add if you wanted; however I'm not really sure why this would be preferable to the existing checkpointing schedule?