shawwn

Twitter thread: https://twitter.com/theshawwn/status/1456925974919004165
Hacker News thread: https://news.ycombinator.com/item?id=29128998

November 6, 2021

How does JAX allocate memory on a TPU?

jnp.device_put(1) is deceptively simple to write in JAX. But on a TPU, what actually happens? How does a tensor containing the value 1 actually get onto a TPU?

Turns out, the answer is "C++", and a lot of it.

zhuowei

record_action_trails
start_phone_number_auth
call_phone_number_auth
resend_phone_number_auth
complete_phone_number_auth
check_waitlist_status
get_release_notes
get_all_topics
get_topic
get_clubs_for_topic

lisawolderiksen

Using Git Commit Message Templates to Write Better Commit Messages

The always enthusiastic and knowledgeable mr. @jasaltvik shared with our team an article on writing (good) Git commit messages: How to Write a Git Commit Message. This excellent article explains why good Git commit messages are important, and explains what constitutes a good commit message. I wholeheartedly agree with what @cbeams writes in his article. (Have you read it yet? If not, go read it now. I'll wait.) It's sensible stuff. So I decided to start following the

foolishflyfox

import matplotlib.pyplot as plt
from matplotlib import animation
from IPython.display import display, HTML
import numpy as np

def plot_sequence_images(image_array):
    ''' Display images sequence as an animation in jupyter notebook
    
    Args:
        image_array(numpy.ndarray): image_array.shape equal to (num_images, height, width, num_channels)

endolith

"""
Train a neural network to implement the discrete Fourier transform
"""
import matplotlib.pyplot as plt
import numpy as np
from tensorflow.keras.layers import Dense
from tensorflow.keras.models import Sequential

N = 32
batch = 10000

systemed

(Dijkstra and plain A* are generally not included here as there are thousands of 
implementations, though I've made an exception for rare Ruby and Crystal versions,
and for Thor, Mapzen's enhanced A*. )

A*                                      Ruby    https://github.com/georgian-se/shortest-path
A*                                      Crystal https://github.com/petoem/a-star.cr
A* (bidirectional with shortcuts)       C++     https://github.com/valhalla/valhalla
NBA*                                    JS      https://github.com/anvaka/ngraph.path
NBA*                                    Java    https://github.com/coderodde/GraphSearchPal
NBA*                                    Java    https://github.com/coderodde/FunkyPathfinding

shagunsodhani

The Batch Normalization paper describes a method to address the various issues related to training of Deep Neural Networks. It makes normalization a part of the architecture itself and reports significant improvements in terms of the number of iterations required to train the network.

Issues With Training Deep Neural Networks

Internal Covariate shift

Covariate shift refers to the change in the input distribution to a learning system. In the case of deep networks, the input to each layer is affected by parameters in all the input layers. So even small changes to the network get amplified down the network. This leads to change in the input distribution to internal layers of the deep network and is known as internal covariate shift.

It is well established that networks converge faster if the inputs have been whitened (ie zero mean, unit variances) and are uncorrelated and internal covariate shift leads to just the opposite.

tencia

from PIL import Image
import sys
import os
import math
import numpy as np

###########################################################################################
# script to generate moving mnist video dataset (frame by frame) as described in
# [1] arXiv:1502.04681 - Unsupervised Learning of Video Representations Using LSTMs
#     Srivastava et al

cybear

Raspberry Pi general optimization

• Use a class 10 SD card for best speed. The USB bus can't come much higher than 30MB/s so you don't have to buy any extremely fast ones though. Not all cards are compatible, check the compatibility list: http://elinux.org/RPi_SD_cards
• Use the HardFloat version of Raspbian instead of the SoftFloat. HF has much faster floating point operations - however SF is required for running Java. So it's either Java or performance, like normal.
• The official Raspbian image gives low network speeds: http://elinux.org/RPi_Performance#NIC
• A graphics driver by Simon / teh_orph is using hardware acceleration for some instructions: http://www.raspberrypi.org/phpBB3/viewtopic.php?f=63&t=28294 installation instructions: http://elinux.org/RPi_Xorg_rpi_Driver
• The firmware can be upgraded which gives, among other things, better GPU performance.

jboner

Latency Comparison Numbers (~2012)
----------------------------------
L1 cache reference                           0.5 ns
Branch mispredict                            5   ns
L2 cache reference                           7   ns                      14x L1 cache
Mutex lock/unlock                           25   ns
Main memory reference                      100   ns                      20x L2 cache, 200x L1 cache
Compress 1K bytes with Zippy             3,000   ns        3 us
Send 1K bytes over 1 Gbps network       10,000   ns       10 us
Read 4K randomly from SSD*             150,000   ns      150 us          ~1GB/sec SSD