jcjohnson/neural-style 17620

Torch implementation of neural style algorithm

jcjohnson/pytorch-examples 3653

Simple examples to introduce PyTorch

jcjohnson/fast-neural-style 3609

Feedforward style transfer

jcjohnson/cnn-benchmarks 2285

Benchmarks for popular CNN models

jcjohnson/densecap 1389

Dense image captioning in Torch

google/sg2im 1095

Code for "Image Generation from Scene Graphs", Johnson et al, CVPR 2018

jcjohnson/cnn-vis 493

Use CNNs to generate images

gkioxari/aims2020_visualrecognition 13

AIMS 2020, class on Visual Recognition

jcjohnson/neural-animation 12

Implementing neural art on video

push eventfacebookresearch/synsin


commit sha da12397a012078482de8dd2196e5fb29c7eb8c8c

Disabling graph computation for running staitstics in custom batch norm

view details

Justin Johnson

commit sha 501ec49b11030a41207e7b923b949fab8fd6e1b5

Merge pull request #20 from holynski/master Memory leak in custom batch-norm

view details

push time in 14 days

PR merged facebookresearch/synsin

Memory leak in custom batch-norm CLA Signed

Using the ResNetEncoder and ResNetDecoder models results in a pretty hefty (~7MB/s) memory leak on CPU memory, which ultimately results in training crashing after a couple days.

I traced it down to a few lines in the custom batch-norm. The variables self.stored_mean, self.stored_var seem to only be used during inference, but are not protected from computing gradients, and thus are extending the computation graph on each call to bn.forward().

An old issue on the main PyTorch repo seems to validate these findings:

I've tested with this change, and the memory usage remains constant.

+14 -13


1 changed file


pr closed time in 14 days

pull request commentfacebookresearch/synsin

Memory leak in custom batch-norm

This indeed looks like a bug on our part -- thanks for finding this and providing a fix!

@oawiles do you mind taking a quick look before I merge this?


comment created time in 15 days

issue commentfacebookresearch/pytorch3d

Rendering of occluded objects looks strange

Hi @abhshkdz, thanks for setting up a reproducing example! Can you change the permissions on the Colab notebook to be world-readable? I can't open it at the moment.


comment created time in 21 days

issue commentfacebookresearch/pytorch3d

knn_points unexpected error using CUDA

Our current KNN implementation will be catastrophically slow for K=4000 -- in that regime you will probably be better off using FAISS ( for KNN instead of PyTorch3D.


comment created time in a month

issue commentfacebookresearch/pytorch3d

How can I do a faster rendering?

There are two easy things you can change in your RasterizationSettings that should give you a good speedup:

  1. Change bin_size; by setting bin_size=0 you are invoking the naive rasterizer, which is quite slow; the coarse-to-fine rasterizer should be much faster. The easiest fix is setting bin_size=None which will invoke the coarse-to-fine rasterizer using our built-in heuristics for bin size. You can also try manually tuning the bin size (try different powers of two) for your application, which might work better than the built-in heuristics. These changes should give you the same rendered images as your current settings, but faster.
  2. Reduce faces_per_pixel from 100 to something smaller (maybe 50 or 10). This will result in images that are different than your current settings, so this may affect the performance of your downstream task.

comment created time in a month

issue commentmadewithml/utterances

I have no affiliation with this repo and I’m not sure what exactly it is. But it looks like this is a link to some of the content from my course at Michigan; you can find everything here:


comment created time in a month


started time in 2 months


started time in 2 months

issue commentfacebookresearch/pytorch3d

Rasterizer shouldn't require cameras as positional parameters

This should now be fixed in If that solves the problem, feel free to close the issue!


comment created time in 3 months