snippyhollow

We changed RoPE's theta from 10k to 1m and fine-tuned with 16k tokens long sequences.

Thanks for the feedback, we'll fix this.

The models are here: https://github.com/facebookresearch/wav2letter/tree/master/r... https://github.com/facebookresearch/wav2letter/tree/master/r...

It is like `cp`.

Quite usable, all of TorchCraftAI uses it https://torchcraft.github.io/TorchCraftAI/ :)

Compulsory "Spikey Spheres" notebook http://nbviewer.jupyter.org/urls/gist.githubusercontent.com/...

Similar to https://arxiv.org/abs/1311.2901

paper: https://s3.amazonaws.com/fairseq/papers/convolutional-sequen...

code: https://github.com/facebookresearch/fairseq

pre-trained models: https://github.com/facebookresearch/fairseq#evaluating-pre-t...

A few years ago SpringRTS, and in particular the BalancedAnnihilation mod [0] were active and playable. There is also Planetary Annihilation [1] (that I used to play competitively online, a few of the top players are ex-StarCraft II Grand Masters).

[0] https://springrts.com/wiki/Balanced_Annihilation [1] http://www.uberent.com/pa/

I did a quick "visualization" hack a while ago http://nbviewer.jupyter.org/urls/gist.githubusercontent.com/...

[2014]

Part 2 was never finished but supposed to be text around this http://nbviewer.jupyter.org/github/syhw/DL4H/blob/master/DNN...

I wrote this quite some time ago exactly to hold the hands of neophytes who want to do such endeavour. http://nbviewer.jupyter.org/github/syhw/DL4H/blob/master/fro...

I have 300/100mbps down/up fiber in Paris. As important as the bandwidth is the latency. Basically it means I do most of my work SSHed (mosh'ed actually) to a cluster and don't notice I'm working from home or the company's network. Streaming plots or even videos for debugging is seamless.

Wow that site still exists! :) I did one when I was 17 (I'm 28 now), I first modelised everything in 1) an optics simulator 2) Solidworks, and then assembled it with a wooden case: http://forum.allinbox.com/aspectgeek/Projetsencours/9/projec...

http://www.google.com/cse?cx=partner-pub-5733989050076975%3A...

The second chapter "High-Dimensional Space" talks about the problem of spikey spheres[0] (how most of the mass is near the surface), I made an ipython notebook to illustrate it[1].

[0] http://www.penzba.co.uk/cgi-bin/PvsNP.py?SpikeySpheres

[1] http://nbviewer.ipython.org/urls/gist.github.com/SnippyHollo...

Thomas is a labmate of mine, fire away if you have specific questions, either I can answer them, or ask him!

I also did topic modeling (LDA) and other experiments on HN, code for fetching all posts and comments, and for topic modeling, can be found here: https://github.com/SnippyHolloW/HN_stats (lein run // look at hn.py)

It's PG's parameters for a naive Bayes model based on these LDA topics, learned by taking his comments on HN as upvotes for the articles content.

Yes but what does it mean?

I did that in 2012 for a pet project with a friend https://github.com/SnippyHolloW/HN_stats

Here is the trained topic model (Nov. 30, 2012) with only 40 topics (for file-size mainly) https://dl.dropboxusercontent.com/u/14035465/hn40_lemmatized...

You can load it with Python:

  from gensim.models import ldamodel
  lda = ldamodel.LdaModel.load("hn40_lemmatized.ldamodel")
  lda.alpha = [lda.alpha for _ in range(40)]  # because there was a change since 2012
  lda.show_topics()

Now if you can figure out what is this file: https://dl.dropboxusercontent.com/u/14035465/pg40.params I'll pay you a beer next time you're in Paris or I'm in the Valley. ;-)

Is Karpelès a US citizen? I don't see it anywhere, I think he's French (born and raised in France).

XBMC torrent has been out there for a while http://forum.xbmc.org/showthread.php?tid=174736

It's free software https://github.com/steeve/xbmctorrent and uses libtorrent-rasterbar through https://github.com/steeve/libtorrent-go

Exactly, if you look at one of her answers:

""" Yes they're paying for my college and I live with them. Things were going good but then marriage and US came and ruined everything. I refused to go and I was told "you have to go". And yeah it's for real, if I do something which brings dishonour to my family they'll kill me and my mother will get in trouble as well. I know it's hard for you to understand but this is how my culture is. """

Which speech recognition engine do you use? Kaldi? Sphinx? Good ol' HTK? Homemade? With deep learning acoustic models?

Only if the car was built after august of this year (page 5). We'll see a lot of good being done soon! ;-)

I feel like setting yourself a "words/period of time" ratio is only good if you do not fear deleting these words afterwards. It's like setting a "LOC/period of time" ratio, focusing on removing bugs and increasing features (as an f-score ;-)) is a better optimization objective, but you've got to start somewhere...

As great as SpaceX is, at least half of its funding comes from NASA https://en.wikipedia.org/wiki/SpaceX#Funding

For Matlab, there are several implementations, including the DeepLearningToolbox: https://github.com/rasmusbergpalm/DeepLearnToolbox

For Python there is the great Theano, and the RBM implementations of MORB https://github.com/benanne/morb