Player is loading...
Introduction to Pytorch 1
Thursday, Feb. 14, 2019 · 12:06 p.m. · 19m 21s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
my name is john yeah and i'm that tells me that taught but sorry that yeah the top c. s. r.
00:00:06
and stay i will give you some very general information uptight
00:00:10
hutch and also the basic in p. implemented by
00:00:13
touch and then tapped out the uh p. t.'s did end
00:00:17
up how much will get you more detail about
00:00:20
how to implement high touch i uh still uh for
00:00:24
the sequence molding and also this c. t. c.
00:00:29
so first i would like to send send a thomas full
00:00:33
giving me this uh take this opportunity to present
00:00:37
something and actually this is the first time ah that i make the tutorials does that was one
00:00:43
and here i am sorry i'm a english there
00:00:49
and those are things have their of a little gave me lots of
00:00:52
suggestions and also encourage me a lot yet so let's begin
00:00:58
uh like as more and more researchers are they
00:01:03
started to the uh uh like run a improve the performance of the job or run their jobs by
00:01:10
the deeper any so uh like implementing uh now i work so hard to find a
00:01:17
like reasonable or still couple of framework for us or how to use it
00:01:23
it's uh uh like now it's very easy to show for us so
00:01:28
today i will introduce the high touch ah i haven't read this revealed
00:01:34
high touch it's really interesting like i'd be real healing height yeah
00:01:40
i'd have more energy and mostly it's clear my eyesight has improved so it seems like
00:01:46
quite it it's a magic right so what is the real high to it is
00:01:55
hi touch a a it's a popular like raining framework implemented employment
00:02:01
and it's uh it can also say it's it's the
00:02:04
uh neural network had cage like released by uh
00:02:09
face broke and if you're not tied the pie touch could also be seeing at
00:02:13
the model used to replace number by any in the neural network domain and
00:02:19
and in the pie touch it can i provide
00:02:23
some some tours to crate train passed and analysts you new
00:02:29
network like you can have different types of layers
00:02:32
uh uh let me near layers for the new question and
00:02:36
combustion lenient for c. uh and also the recruit
00:02:40
right right we can't really say their currency layer four maybe as t. m. or an
00:02:46
i know so it came payment the ah activation functions
00:02:51
like realistic one exact up and also you can
00:02:55
uh compute the gradient he opts miser like adam a
00:03:00
and s. g. d. and also you can
00:03:03
again you to uh likes meet in g. p. u. so uh as i think most of
00:03:10
your maybe now the pie talk shorthand several so there are uh some of the framework
00:03:16
like fanciful high torch terrorists or uh the uh
00:03:21
s. c. sites like psyche to their
00:03:25
yeah so why way choose by touch or why we should choose by touch
00:03:31
so let's begin with non pike and uh as i mentioned before pie touch is just
00:03:36
uh i mean people like replacements of number hike in the neural network domain
00:03:42
so not hide designed for the fast computations and could be that you peel
00:03:47
exaggerated but you need more at more like backstroke code and uh it
00:03:53
can work while for the major uh like uh the m. l. problem
00:03:59
mention any problem but now the deep training problem is
00:04:02
expert financial a hard for the number high
00:04:07
and it's like for ten civil law is returned mainly in c. plus plus and
00:04:13
also in soda i think c. plus plus is like a nightmare for years
00:04:18
yeah someone and like that has a role in provide a
00:04:25
functionality in seal sits up not class and java
00:04:29
so logic is not restricted intensive role and but it's
00:04:36
has very specific logic like please hold their stations
00:04:40
so you need so or i should say you have to make some effort on this
00:04:45
kind of things and then you can get into the tens of real high but
00:04:50
so much that the high touch is the still bit tight then so but
00:04:55
that's the only here is very positive word for us because it just
00:05:00
if tyson so i'll i'll what we can do in high then and we can see what we can
00:05:05
do imply touch so that's quite easy for us to there or is it was to implement
00:05:11
anti touch has very strong deep you support and if it is
00:05:16
specifically tailored for the g. p. o. functionality implied then
00:05:21
and uh like i mentioned before is the replacement of number high to use the power of g. p. use and
00:05:28
it's also a dynamic computation went thing uh so what
00:05:33
does it mean we can see a example here
00:05:39
yeah this is not made by me
00:05:43
mm
00:05:46
yeah i like this one off can't show it again in a way that's what
00:05:51
you like when you run you gave me it's the the reliable and you
00:05:57
maybe much to hide the make checks and they you to uh
00:06:02
and some active vision functions and then you do the bat work
00:06:06
uh back word for applications it's every time you run you
00:06:11
that's something to the rough so it dynamic knots like tends to flow you made a very
00:06:17
uh does that take worth well first and if you use it you should call it out
00:06:25
and yeah uh this is also the same thing to the also wait package
00:06:31
in the uh in high touch and we k. show you here
00:06:36
in the official documents up high touch exactly ah there's this
00:06:42
this this intro to own your network somehow touch
00:06:44
it this package the package provides the automatic the french
00:06:49
information they uh uh sorry the beef rich
00:06:53
anyway yeah uh for operations in kansas an eight is defined by right but
00:07:02
did you find my ruined framework it's just like i show and uh then i make graph which means that you
00:07:08
are maybe back propagation is defined by how you call this
00:07:12
one and that every single iteration can be different
00:07:16
and you can get more information in this dog a pipe watched the document uh i i will continue
00:07:25
yeah because it is like rate by a written by the pipeline so it's
00:07:30
very easily to use and flexible and this meeting is very high
00:07:38
so i was a start has them really basic things there uh stands there is it is a
00:07:43
little uh like no level of high touch and then we'll build up something of afterwards yeah
00:07:50
the first thing you should do is import touch someone will ask why not high touch because this is high touch and i
00:07:57
search more that because high touch is like the idea of
00:08:00
high touches from torch which is within by uh uh
00:08:04
language but that work pro program language names
00:08:09
lou well yes i'll you a a but that's like several people here that and
00:08:15
uh it's might be difficult to various so the research is it ten that's too
00:08:20
little pipe it so well here way just import touch it and we will
00:08:26
compare that taught by talks with number out or so we also imports nom tiny
00:08:33
yeah and then we can set battles will the cancer that as
00:08:37
but there is just like uh the number high is that and
00:08:41
that mentioned all angry but the uh uh only difference of
00:08:45
the them are uh like cancer can
00:08:50
uh can implement a total different she ate asian
00:08:55
and then you can do the backward propagation automatically a the not
00:08:59
as in the cancer flow you should do it manually
00:09:03
yeah and it's a bit if we can arrange a cancer it back to us
00:09:10
here by give it a value that means how many adam and seen
00:09:14
this data and uh we should we should not his is that
00:09:19
the index is non zero so if we run it's the t.
00:09:23
one should be of tens or better from zero to five
00:09:27
and we can also specifically range like from four to ten but it's
00:09:32
not going to ten so it should be forty nine and
00:09:37
yeah we can why it's now and we can see uh the t. one it's one
00:09:42
zero two five eighth of data and the t. two is from four to nine
00:09:48
and also we can said it's randomly uh by using the torch taught right and it
00:09:54
will show like six adamant uh but maybe we've learned every time it's different
00:10:01
yeah and we can matter they set the what what you want like this i use talk start stands
00:10:08
there and i wanna cancer to be just this back to and it will shows like this
00:10:14
ah we can also say the size uh it's beer out
00:10:18
no matter to show the size the first is we
00:10:21
can use a dodge shape and it can shows the
00:10:25
size is full and we can also use like
00:10:30
a slice
00:10:36
okay use use this one to show the shape
00:10:40
oh and also we can do the uh taught for that
00:10:44
that's and for example that t. one dot t. till should be
00:10:50
there's some off uh the multiplication of every page here
00:10:55
of the adamant that had the same index
00:10:59
i hear is zero by four plus one by five exact for
00:11:06
yeah and
00:11:09
it is that we can like revenge but that are we can also uh change the battles by index
00:11:17
like if we want to now oh what it is in the t. four three
00:11:25
like this or elements so to force re is i think
00:11:30
yeah i think is for no this is the not the third one is the
00:11:34
fourth one because in indexes from zero and we use that thought i'd some
00:11:40
and you had the brackets to show this that it'll any view
00:11:46
i think if if you didn't it's this one and you will you will see a ten set
00:11:55
yeah and we can also trend is that you know if you want to set this to ten
00:11:59
and i have to run we'll look at this tends to it's original is one
00:12:05
two three four and now is a bit above four elements have changed
00:12:11
we can also it to see the size uh change type
00:12:18
and we said we can change the battery or uh we can also chained uh type by use this
00:12:23
message with the type they have the floats for a thirty two of road softball sex tape
00:12:29
for float and we can find it in the official document as well
00:12:34
and also they are different for g. p. u. n. c. view
00:12:39
ah and next thing i would like to show is that a
00:12:43
a elements why's operation like that thought we had seen before
00:12:50
yeah like he wanted to ease the cancer is like this and way um
00:12:56
implement the ad addition multiplication and division here
00:13:02
and we'll get this without is just a elements
00:13:05
wise operations i think everyone my notice
00:13:10
and then yeah i would like to show his them pens and make
00:13:13
shakes yeah the matrix multiplication is a similar to number high but
00:13:18
here are instead of use the am am it's use this one
00:13:24
starts at an a. t. and new l. just matrix multiplication
00:13:32
and like we said this a matrix to be so by
00:13:35
fine men check and uh this is five to
00:13:39
the five bytes so may take and we can multiply them because
00:13:44
uh the size and this size should be the same
00:13:50
we can also say the shape of the results it's is ah to buy to make chick
00:13:59
yeah if we don't know the shape object has or we want to
00:14:03
re ship it we can use this mess are taught to deal
00:14:07
unlike t. one is uh let's see one i
00:14:12
think is a six yes exactly meant
00:14:15
stands there and we can reshape it's to ought to buy so we may check
00:14:21
and uh we can also uh reshape it's to a three by what
00:14:25
to reach a uh may take so this is basically the results
00:14:31
and uh besides that's we can also uh like come to calculate the
00:14:36
mean and uh absolutes value of the h. elements in the matrix
00:14:46
so like uh this cancer is have some uh how
00:14:53
to say the negative battles and by uh
00:14:57
but by implements this message i need to all be positive
00:15:04
so and always pick here with cancer ships when doing calculations
00:15:08
uh we can have example uh you should yeah like hear me
00:15:12
x. tens or is about to like one by twelve
00:15:16
and the white cancer is we by finding checks so if you do the multiplication it will show the arrow
00:15:23
and so here we can use the the deal mass act and it will try
00:15:32
this should be work and there i want to show people so happy
00:15:40
oh
00:15:42
it's
00:15:47
and here i want social shake if if you're now you
00:15:52
have a magic is three by five and you want
00:15:56
so multiply them but you don't know um maybe it's
00:16:00
hard to calculates what it's should be we should
00:16:03
so we can use like minus one because away when no this is the way right
00:16:08
and if we want to multiply them we should have another sweet but here maybe it's hard to calculate
00:16:14
we just use mine is one and you it will to matt grey a complete this one
00:16:22
uh and this shows just how also converts non high to change their uh it's very easy
00:16:29
like to non non high to tense there is just taught taught from um hi
00:16:34
and from tested so irate acton um high is uh like
00:16:40
this is what other kinds areas and dots nom high
00:16:45
yeah they look similar but different so like here had the common here i mean side also had to come
00:16:52
mm
00:16:55
and this is uh this is with the that's non high and
00:17:00
the tears are they share the same memory locations though
00:17:03
you change one and you would change another like a a is that has a and b. is an um hi
00:17:10
and if you add that to me is the non python that has it
00:17:15
so if you add a and you would change me as well
00:17:20
so uh if you change the nom high and you you tense it will change as well
00:17:28
so that we should just yeah the last thing i would like to show some useful functions i want to give you
00:17:34
uh_huh uh like notation that uh if we want to paul something use the mad props late
00:17:41
which shorts convert ten says to non parish first or else it doesn't work because this
00:17:47
one i think that's k. in like you use non high every to draw something
00:17:55
so the whole plot them at what they will seize them sign function is the axes that hands are here
00:18:03
uh from being from minus ten to ten and they say this
00:18:06
that and this is the art type pop the ten cent
00:18:11
and is just use as starts sign to show this function
00:18:18
and the same as uh the this function i don't know how to say that but
00:18:26
all anti r. i. a. it has the impact that it's very easy
00:18:30
to implement these kind of things and i think mine as well
00:18:36
and it's an issue of function as well
00:18:40
and the soft max
00:18:43
as well as soft max uh what what we should not is
00:18:46
is here we should add a dimension it was zero here
00:18:52
uh and also like the active they should functions is lenient uh nodding and so
00:18:58
if we gave them a linear function like why was ten times x.
00:19:03
and we will use maybe that's the one function to get these
00:19:10
this one off
Conference Program
Raw Waveform-based Acoustic Modeling and its analysis
Mathew Magimai Doss, Idiap Research Institute
Feb. 14, 2019 · 9:12 a.m.
367 views
About Sequence Classification for Sound Event Detection and end-to-end ASR
Thomas Pellegrini, IRIT, France
Feb. 14, 2019 · 10:14 a.m.
Case study: Weakly-labeled Sound Event Detection
Thomas Pellegrini, IRIT, France
Feb. 14, 2019 · 11:05 a.m.
113 views
Introduction to Pytorch 1
Feb. 14, 2019 · 12:06 p.m.
114 views
Introduction to Pytorch 2
Feb. 14, 2019 · 12:26 p.m.
134 views
Recommended talks
Building and deploying Kibana plugins ... And should I do it?
Alexandre Masselot
May 11, 2017 · 2:04 p.m.
1999 views