ProGamerGov

@htoyryla On the subject of grey spots/haze, I collect a bunch of examples caused by the `adam` optimizer. These did not appear until after I switched from using `lbfgs` to...

In the `optim` library. [adam.lua](https://github.com/torch/optim/blob/master/adam.lua) contains all the usable parameters for the Adam optimizer: ``` ARGS: - 'opfunc' : a function that takes a single input (X), the point of...

@htoyryla Adam seems to act sort of like a higher content weight with some style/content combos, and with others it seems to create less details, or less "refined" details. The...

Here's a better comparison of different epsilon values for the Adam optimizer:  Direct link to the full image: https://i.imgur.com/tQDF2Qd.jpg The full album: https://imgur.com/a/PU004

> Adam does not really create anything, it looks for a minimum in a HxWx3 dimensioned space. I should have clarified that I meant "Neural-Style with Adam, creates" instead of...

I discovered that the `beta1` parameter cannot be equal to `1` or else it results in `nan` for all the loss values. The `beta2` parameter seems to have significantly less...

After some more experimentation, I think that using this combination settings for Adam, creates better results, or at least results that are equal in quality to the default Adam parameter...

@htoyryla Have you given any thought into ways that we could optimize neural_style.lua, even slightly more than it is currently? I'd like to try and counteract the newer Cuda/cuDNN/Torch7 memory...

I have identified a new issue in Neural-Style than I haven't seen described anywhere else: This issue appears to related to the FCN-32s PASCAL model, which can be found here:...

Using variations (only changing the `-image_size` value) of these two commands, I have noticed that VGG models with their FC layers removed, use less memory: ``` th neural_style.lua -backend cudnn...