1
00:00:11,662 --> 00:00:18,240
MODERATOR: Hello, everyone. Welcome back to the 
platypus theater and PyConline. Next up is Sam  

2
00:00:18,240 --> 00:00:32,800
Bishop. He loves Python, Django, cats, working on 
personal and professional hardware, playing games  

3
00:00:32,800 --> 00:00:46,000
of all kinds and tinkering with 3D printers. 
>> I am Sam I am presenting this talk. This talk  

4
00:00:50,747 --> 00:00:56,880
is going to be a bit of a rush in places but I 
think there should be enough time to get through  

5
00:00:56,880 --> 00:01:04,240
everything. We will start off with faker. If 
you are trying to get random data maybe you  

6
00:01:04,240 --> 00:01:09,840
came across this but testing or any of these 
other scenario. It has a Pychess plugin.  

7
00:01:13,840 --> 00:01:22,000
It let's you generate all kinds of fake 
information. Names, jobs, credit cards,  

8
00:01:23,040 --> 00:01:29,520
whole profiles even. It is very 
powerful and a lot of data in it. It is  

9
00:01:31,440 --> 00:01:37,520
more data than you think. It has a number of 
plugins that allow you to generate extra things  

10
00:01:37,520 --> 00:01:45,520
like social security numbers, ISBN books, and all 
these community of plugins. It's got localization  

11
00:01:45,520 --> 00:01:50,240
and the ability to get all kinds of extra 
character sets and things like that.  

12
00:01:51,600 --> 00:01:55,600
Interestingly, compared to the next 
one, the ability to do unique sets  

13
00:01:55,600 --> 00:01:59,120
which is handy if you want to get a whole 
bunch of things and not have overlap.  

14
00:02:00,400 --> 00:02:09,440
Then we have mimesis it is a fast data generator. 
You will be generating lots and lots and don't  

15
00:02:09,440 --> 00:02:15,200
mind about things like uniqueness you can use 
it to get a large amount of fake data cooked  

16
00:02:15,200 --> 00:02:21,840
up for whatever you are trying to build. Fake 
census. I want thousands of names or jobs or  

17
00:02:22,400 --> 00:02:30,640
cars. It can put them a lot of interesting 
information like political views and occupations  

18
00:02:31,680 --> 00:02:38,400
and addresses and geography and it can do 
food which isn't something in faker and a  

19
00:02:38,400 --> 00:02:44,800
few of the other ones. Drinks and ingredients and 
dishes. We have our trucks I mentioned earlier. We  

20
00:02:44,800 --> 00:02:50,480
also have cars and airplanes. Those of you who 
are a bit of an airplane nerd may have spotted  

21
00:02:50,480 --> 00:02:57,680
two that don't exist so it making fake airplanes. 
There is no such thing as an Airbus 50.  

22
00:02:59,280 --> 00:03:12,120
We have a couple phones. We have CPUs, GPUs, fake 
resolutions. They are real but it made them up. We  

23
00:03:12,120 --> 00:03:20,880
have programming languages. OS licenses. Operating 
system names. It's got a very comprehensive set of  

24
00:03:20,880 --> 00:03:26,160
things it can generate for you. Not all this stuff 
is stuff you will think of getting when you are  

25
00:03:26,160 --> 00:03:32,000
making creative content but if you are trying to 
think of something in the middle of writing and  

26
00:03:32,560 --> 00:03:38,400
get stuck it is a great way to unblock if you 
need a name of a person while writing a story  

27
00:03:39,200 --> 00:03:45,040
and can't think of anything. You can reach for 
a tool like this and say give me a name, job  

28
00:03:45,600 --> 00:03:53,920
or place and it can push through that kind of 
scenario. It has bunch of locale settings. You can  

29
00:03:55,840 --> 00:04:06,640
have English spelling and French and German. It 
has country-specific generators to do things like  

30
00:04:06,640 --> 00:04:17,280
I want a European ID number. And a schema-based 
tool to create specific generators for your own  

31
00:04:17,280 --> 00:04:20,480
purpose. It is definitely worth looking 
at if you haven't looked at it before.  

32
00:04:22,560 --> 00:04:29,080
On from that, there is more focused tools that I 
have come across and put in the presentation. We  

33
00:04:29,080 --> 00:04:39,680
have Sarah's entry from 2014. In case you didn't 
know nano was a contest where they were trying  

34
00:04:39,680 --> 00:04:51,840
to generate the content. Seraph was impressive.
He built a complete nanoscript generator. He took  

35
00:04:51,840 --> 00:04:59,680
a bunch of text generators and ran them through 
and produced a whole book that can be poked and  

36
00:04:59,680 --> 00:05:06,960
produced randomly variations on. His book is this 
variation. You can use the tool to generate your  

37
00:05:06,960 --> 00:05:14,000
own version of this quite easily. It is a very 
impressive output. If you are building up some  

38
00:05:14,000 --> 00:05:21,840
kind of a world and want mystery text, content, 
cryptic cipher stuff, it is pretty and impressive.  

39
00:05:23,840 --> 00:05:31,440
He is using the flicker API and image conversions 
and processing to color patch backgrounds and  

40
00:05:32,080 --> 00:05:36,960
patching dual font. It is very impressive and 
worth the read. He has put it all online in  

41
00:05:36,960 --> 00:05:42,880
the links you saw in the earlier slides. OK. 
On from that. We have the Pantheon generator.  

42
00:05:45,520 --> 00:05:54,000
I am going to do this as an interactive demo. 
This one is really fun. Here we are pulling it in,  

43
00:05:57,680 --> 00:06:06,400
get it started, loaded up, get a first God and 
a second God. The way this works is you have to  

44
00:06:06,400 --> 00:06:15,760
have ones biologically set based because this 
is a genetic program. You need an XX and an XY.  

45
00:06:16,320 --> 00:06:22,560
There is a bunch of stuff under the hood where 
there is separate between gender and sex. It is  

46
00:06:23,120 --> 00:06:29,840
fascinating when you look up how it is combining 
multiple attributes. We have our two primary  

47
00:06:29,840 --> 00:06:35,520
gods to start things off. We will make our 
Pantheon and ask to create 10 generations.  

48
00:06:37,840 --> 00:06:45,360
It will create a whole sequence of gods for us. 
We get a whole bunch of them progressing down,  

49
00:06:46,080 --> 00:06:51,600
losing their powers as they sort of get younger 
and further down in the way a lot of classic,  

50
00:06:51,600 --> 00:06:59,680
sort of Greek and Roman era stuff tends to have. 
Once that's done producing our 10 generations,  

51
00:07:00,240 --> 00:07:03,760
I will scroll up and show a few more 
of the interesting ones along the way.  

52
00:07:06,160 --> 00:07:15,840
It is a fairly large pantheon. 
These Gods have been very busy.  

53
00:07:17,200 --> 00:07:24,640
Yup. If you look further back up in the top, you 
see that all of them are very much gods to begin  

54
00:07:24,640 --> 00:07:30,160
with. As you progress, it starts smattering 
in the ones that are sort of partial. It is  

55
00:07:31,040 --> 00:07:37,840
crossbreeding and losing the powers in that 
sort of gradual sense. We have demigods and  

56
00:07:39,920 --> 00:07:48,880
further down here we have got -- there was a 
semi-divine. Right down to an ordinary women. It  

57
00:07:48,880 --> 00:07:56,720
is this progress of becoming a very Greek myth 
pantheon here. It is all sort of built off the  

58
00:07:56,720 --> 00:08:03,680
back of this taking each generation, combining 
them, and producing an output iteratively giving  

59
00:08:03,680 --> 00:08:12,800
us this rich pantheon. We end up with various 
hierarchies and just to illustrate how it built  

60
00:08:12,800 --> 00:08:25,840
this whole pantheon with connections whose parents 
and all of the relationships we will graph it out.  

61
00:08:30,720 --> 00:08:38,800
Has to think why it graphs it out. As you can see, 
it is a very, very rich set of connections here  

62
00:08:38,800 --> 00:08:44,160
with all the various intergenerational steps 
and relationships between them all and whose  

63
00:08:44,160 --> 00:08:50,480
parents produced which. All of this is available 
in that library. It is quite a powerful tool to  

64
00:08:50,480 --> 00:08:56,080
sort of create this kind of basic world building 
background material for either a story or a game.  

65
00:09:05,200 --> 00:09:13,840
The earthlike procedural planet generate 
with produce really neat things.  

66
00:09:18,400 --> 00:09:25,120
I put two seed values in here to show how we 
get two distinct planets based on that input.  

67
00:09:26,080 --> 00:09:30,160
It takes a while to render these but they 
come out looking pretty nice if you ask me.  

68
00:09:32,800 --> 00:09:42,320
On that point, there is another planet generator. 
Now this one produces an actual whole 3D world.  

69
00:09:45,360 --> 00:09:50,640
This is all done in Plotly but it is producing a 
height map for the whole world based on the input.  

70
00:09:51,200 --> 00:09:54,720
You can easily convert this into 
whatever format you want since the  

71
00:09:55,280 --> 00:09:59,920
tool produces the raw grid and you 
can work with it from that point.  

72
00:10:03,280 --> 00:10:14,080
After that planet generator we have our 
crossword composer. This allows you to fire  

73
00:10:14,080 --> 00:10:18,880
up the GY and you can put them in there and 
it will crunch out a crossword puzzle based  

74
00:10:18,880 --> 00:10:25,760
on the words and the grid. Here is an example 
of it running. You can see how it is taking the  

75
00:10:25,760 --> 00:10:32,400
words it is modifying and fitting them in so 
they have to match up and you would obviously  

76
00:10:32,400 --> 00:10:36,880
take the words and make up your own questions 
if you were actually making a crossword with it.  

77
00:10:39,360 --> 00:10:45,920
We have our spaceship generator. The spaceship 
generator is a blender plugin. It is a relatively  

78
00:10:45,920 --> 00:10:52,480
easy one to use. You just pop in here and then 
you can go in and you can add a spaceship. Easy  

79
00:10:52,480 --> 00:11:00,400
as that. But what's interesting is the variety you 
can get. You can get a number of different outputs  

80
00:11:01,520 --> 00:11:08,080
based on just the input being completely random. 
It doesn't ask for any parameters. You just  

81
00:11:08,080 --> 00:11:15,360
give me a spaceship and it produced an actual 
completed object with colors, little bit on it,  

82
00:11:15,360 --> 00:11:22,160
you can see it has things like antennas and little 
arrays and guns and engines and it is a very  

83
00:11:22,160 --> 00:11:28,560
impressive amount of detail being placed into 
these and completely randomly generated. It is  

84
00:11:28,560 --> 00:11:37,200
trivial to start using. It is definitely quite 
an impressive free plugin produced by someone who  

85
00:11:37,840 --> 00:11:41,840
put a bit of time into actually making something 
come out really well with very little effort  

86
00:11:43,520 --> 00:11:52,560
and very wide variety of outputs from the long, 
sort of narrow. It is a very impressive plugin.  

87
00:11:58,080 --> 00:12:08,400
Not a heavy blend user. After the spaceship 
generator we have the tree generator which  

88
00:12:09,360 --> 00:12:15,760
is another blender plugin used for procedurally 
creating trees. There is a lot of power and  

89
00:12:18,800 --> 00:12:34,080
it was produced as someone's undergrad. I have 
a bamboo to show variation. If we customize the  

90
00:12:34,080 --> 00:12:40,560
parameters, we can end up controlling that growth. 
Do we want a shorter plant? A taller plant?  

91
00:12:40,560 --> 00:12:49,680
Or one that is not going to spread out as 
much. That makes it look less like bamboo  

92
00:12:49,680 --> 00:12:56,160
and this is important with other trees where you 
want to control their growth. It takes a while to  

93
00:12:56,160 --> 00:13:01,200
render them out and construct them. It doesn't 
texture them so they come out without colors.  

94
00:13:02,880 --> 00:13:08,880
Without a lot of work, it is possible to produce 
some very impressive results. I asked someone who  

95
00:13:08,880 --> 00:13:15,840
knows how to do the blendering and coloring and 
shaders and without a lot of work they were able  

96
00:13:15,840 --> 00:13:26,640
to convert this out within one day and they 
produced this absolutely marvelous Cambridge oak  

97
00:13:27,920 --> 00:13:34,720
from just a few hours to color that up. And most 
of that was working out what I handed them. They  

98
00:13:34,720 --> 00:13:41,600
said it was pretty easy. They did that with no 
complex, just basically the 3D equivalent of the  

99
00:13:41,600 --> 00:13:49,280
old paint bucket fill in the Photoshop or paint 
and it comes out looking pretty impressive. It is  

100
00:13:49,280 --> 00:13:55,120
a very powerful plugin. If you need to generate 
anything tree-plant like worth looking at.  

101
00:13:57,760 --> 00:14:04,720
We have got weld engine. This one is a big one. 
I keep going past this page. You may have been  

102
00:14:04,720 --> 00:14:18,560
wondering what it is. What it is is a full 
plate tectonic generation tool. Participation  

103
00:14:20,960 --> 00:14:24,800
and temperatures. It allows you to generate 
these and save them all. Here is one I have  

104
00:14:24,800 --> 00:14:30,000
already got generated. You can see it 
has a temperature banding, we have got  

105
00:14:30,000 --> 00:14:35,920
average global temperatures, we have got things 
with higher terrain. I am not actually being  

106
00:14:35,920 --> 00:14:41,600
as warm. We have rain shadowing and similar 
things we switch back to precipitation view.  

107
00:14:43,760 --> 00:14:49,680
It is pulling this in from a 100 meg 
protobuff file. You can see how these  

108
00:14:49,680 --> 00:14:56,160
land forms are producing rain shadows. 
It is an incredibly detailed simulation.  

109
00:14:56,160 --> 00:15:02,720
It is quite powerful to be able to create these 
things. It does a full biome. It allows you to  

110
00:15:03,680 --> 00:15:10,560
plot out how much of the planet is covered by 
various types. The more useful one is the biome  

111
00:15:10,560 --> 00:15:16,960
view itself which gives you a sort of layout of 
what all the different biome areas look like.  

112
00:15:18,000 --> 00:15:23,600
You can do a satellite view that sort of shows 
you what does this look like in real color.  

113
00:15:25,360 --> 00:15:31,680
Just wait for that one to pull itself up. This 
one starts out with an actual plate tectonic  

114
00:15:31,680 --> 00:15:36,320
simulation. It generates a bunch of land forms 
and does a fiction simulation of them -- physic  

115
00:15:36,320 --> 00:15:43,200
simulation of them colliding with each other to 
build up the continental and ocean crust. You put  

116
00:15:43,200 --> 00:15:48,880
in how many plates you want and it does a number 
of iterations which is the slowest part of the  

117
00:15:48,880 --> 00:15:53,840
process. It takes a lot longer to do the plate 
simulation than it does to do anything else.  

118
00:15:57,360 --> 00:16:01,840
This is one of it larger files -- the large 
files I generated and it produces more detail  

119
00:16:01,840 --> 00:16:07,280
in terms of rivers and things like that. It 
might take a while to load it up but if it  

120
00:16:07,280 --> 00:16:14,800
keeps going we will skip past it. I think that 
might be stuck there. We will skip over it.  

121
00:16:15,760 --> 00:16:20,560
But it is online. I have produced a fork of that 
one. If you want to look for that one in mind  

122
00:16:20,560 --> 00:16:26,240
it is definitely worth a look at. Yeah. Ah, 
there we go. You can see it as produced the  

123
00:16:26,240 --> 00:16:35,840
color terrain and we have deserts and river 
valleys and various places there. OK. After that,  

124
00:16:36,800 --> 00:16:41,680
how are these actually put together? What 
is under the hood of the Python tools?  

125
00:16:46,000 --> 00:16:53,920
If you want to modify one or build your own, where 
do you even start? The first few I showed a fairly  

126
00:16:55,040 --> 00:17:01,920
random. Everyone sort of starting out does the 
pick one from a list basic random choice problems.  

127
00:17:01,920 --> 00:17:09,280
That's what they are built up for. Even the nano 
Jenn mode entry is heavily built are randomness.  

128
00:17:09,280 --> 00:17:16,427
It is additive to the image processing and 
matching the colors but under the hood it  

129
00:17:16,427 --> 00:17:23,760
is go to the flicker API and get random pictures 
and pick random pictures, pick some random text,  

130
00:17:24,640 --> 00:17:30,640
adjust it to match the font, not put that in. 
It is still randomness driving that process.  

131
00:17:31,280 --> 00:17:35,840
You can go quite a long way with randomness.
You feed to step it up if you want to  

132
00:17:35,840 --> 00:17:41,440
create richer things -- need. Get into 
the genetic algorithms which was in the  

133
00:17:41,440 --> 00:17:48,640
pantheon generator. That takes these things 
and combines them. There is a complex  

134
00:17:49,200 --> 00:17:52,080
set of things that are and 
aren't genetic algorithms but in  

135
00:17:52,800 --> 00:17:58,080
essence what they are is the process of 
taking the properties of one generation,  

136
00:17:58,080 --> 00:18:04,800
mixing them and producing a new generation and 
rather or not that is one kind or another really  

137
00:18:04,800 --> 00:18:11,520
isn't important at this level. It is a very 
powerful way for you to get a combination output.  

138
00:18:11,520 --> 00:18:17,440
You know, take properties of multiple things and 
produce a whole plethora of outcomes that gives  

139
00:18:17,440 --> 00:18:24,080
you things like that whole graph you saw where 
from just two inputs we ended up with hundreds of  

140
00:18:24,080 --> 00:18:30,320
different Gods all of their own various domains 
and specialties. We went from at the very top,  

141
00:18:30,320 --> 00:18:46,080
we started out with, you know, Goddesses. We 
have Curtis God of depends. It is going from  

142
00:18:46,080 --> 00:18:51,440
to the the other end is a generative process and 
a genetic algorithm is a great way to build that.  

143
00:18:52,960 --> 00:18:59,520
On top of that we have noise processes. This is 
the bread and butter. The noise processes are a  

144
00:18:59,520 --> 00:19:05,280
way you would have seen with there visual tools. 
These are heavily built on noise processors.  

145
00:19:09,040 --> 00:19:13,760
I will cover three noise processors. We have there 
diamond square. It is the simplest and one of the  

146
00:19:13,760 --> 00:19:20,400
oldest. That is the whole program right there. 
When you adjust the property, which it only has  

147
00:19:20,400 --> 00:19:28,000
one, the F property, what you end up with is 
a variety of noises. Random. If you actually  

148
00:19:28,000 --> 00:19:36,080
change them up or down we change the scale of that 
randomness. Les course. Change it down and it gets  

149
00:19:36,080 --> 00:19:40,400
less course. All of the way do you know there 
-- all of the way down there shows a very smooth  

150
00:19:41,680 --> 00:19:56,320
spread out randomness. We have perlin noise 
which is what that is. It is a fairly simple  

151
00:19:56,320 --> 00:20:02,560
noise process. I am not going to show the actual 
implementation. It takes a whole screen. Not  

152
00:20:02,560 --> 00:20:08,560
worth really showing. But Perlin noise is fairly 
understandable. You change the parameters and  

153
00:20:11,040 --> 00:20:18,160
end up with changes in the noise. It depends 
on the input. It won't generate different noise  

154
00:20:18,160 --> 00:20:23,600
without changing those inputs. If you do change 
the inputs, you can adjust what you get in terms  

155
00:20:23,600 --> 00:20:28,880
of the output. If we make the area bigger even 
though the ratio is the same we get different  

156
00:20:28,880 --> 00:20:34,960
inputs. If we increase the area, you will 
notice it doesn't change. That's a double  

157
00:20:35,520 --> 00:20:43,280
picture. The grid is twice as big. The noise 
is the same more blended out because that's  

158
00:20:43,280 --> 00:20:50,560
one of the parameters that changes the noise 
versus the data output. And Perlin noise is  

159
00:20:51,200 --> 00:20:57,120
one of the stables. It is used in a lot of places. 
Stepping out from Perlin noise is simple simplex  

160
00:20:57,120 --> 00:21:08,640
and open simplex noise which is one of the more 
interesting ones. You can generate multiple noise.  

161
00:21:14,480 --> 00:21:22,320
I am not going to dig too much into open simplex 
because open simplex is the one I can show you and  

162
00:21:22,320 --> 00:21:27,920
get into it. That's the whole thing. I am not 
going to bother trying to even explain the  

163
00:21:27,920 --> 00:21:34,800
process. That's the most readable implementation 
I could find. It is a bigger algorithm than you  

164
00:21:34,800 --> 00:21:41,280
think. It is pretty powerful. If you do want 
to dig into it, there is a fantastic paper  

165
00:21:43,200 --> 00:21:49,760
on simple simplex noise which is applicable to 
open simplex because once you understand how it  

166
00:21:49,760 --> 00:21:55,680
works by getting to the end of the paper you will 
see it is pretty much flipping around which way  

167
00:21:55,680 --> 00:22:04,000
they work. It is a weird way but patterns, yeah. 
Let's see it in use. What we have is open simplex  

168
00:22:04,000 --> 00:22:12,000
noise. It is just chucked in the parameters there. 
Obviously setting it up there. The feature size  

169
00:22:12,000 --> 00:22:19,280
is set to 24 pixels. We get noise of this size. If 
we have that, the noise gets more grainy. It has  

170
00:22:19,280 --> 00:22:27,920
been shrunk. If we change it to a smaller 
one we get another step. We half the size  

171
00:22:27,920 --> 00:22:33,920
and set it down to 8. If we shrink it you 
can see it is getting grainy and granular but  

172
00:22:34,880 --> 00:22:41,920
at the same time it is one of these things that 
let's you control what text you get out of the  

173
00:22:41,920 --> 00:22:48,480
noise process. Tuning these parameters 
is the skill in using the noise process.  

174
00:22:48,480 --> 00:22:53,840
It is basically what you can use to control 
the how of the output being useful to you.  

175
00:22:56,720 --> 00:22:59,760
This is a version I have run through 
numpy which makes sense in a minute.  

176
00:23:00,880 --> 00:23:06,640
You are able to run this in a vectorised 
fashion and run it in parallel.  

177
00:23:10,080 --> 00:23:14,960
When you want to generate thousands of pixels 
worth of it, it starts to become relevant. Or  

178
00:23:17,200 --> 00:23:23,280
when you want to do it in 3-4 dimensions. Numpy is 
better that -- at that than a couple four loops.  

179
00:23:24,000 --> 00:23:30,560
I turned a 3D array into a GIF. You can think 
of it as panning up through the layers of  

180
00:23:30,560 --> 00:23:36,240
noise and you can see how the noise pattern is 
smoothly evolving up through the layers and the  

181
00:23:37,120 --> 00:23:47,680
black and white areas are not moving just at 
random. Moving on we have constraint solving.  

182
00:23:48,400 --> 00:23:55,200
If that sounds unfamiliar, you probably have 
come across it and not known about it. These are  

183
00:23:55,200 --> 00:24:05,760
constraint solving games. Crossword games are an 
example of this. Human brains are good at this  

184
00:24:05,760 --> 00:24:12,880
without thinking. Scrabble, again, a constraint 
solving game. Constraint solving problems with  

185
00:24:12,880 --> 00:24:20,000
complicated. It is hard to teach a computer how 
to do. If you want to dig into understanding it,  

186
00:24:20,000 --> 00:24:25,200
I would recommend starting on the Wikipedia page 
and finding the specific topics you want to study  

187
00:24:25,200 --> 00:24:32,800
and dig into them that way. On top of that 
there is wavefunction collapse which is an  

188
00:24:32,800 --> 00:24:39,920
awesome algorithm that let's you take starting 
seed samples and it generates out stuff based on  

189
00:24:39,920 --> 00:24:45,600
those seed samples. A good example on the slide 
is we have the little plant, sample goes in,  

190
00:24:45,600 --> 00:24:51,360
and you get varieties of plants coming out after 
by tiling and moving across the algorithm. It is  

191
00:24:51,360 --> 00:24:56,400
a pretty interesting algorithm that's worth sort 
of reading through. It would honestly be a whole  

192
00:24:56,400 --> 00:25:02,400
presentation by itself. I can only really show 
you the high-level example here of how it is  

193
00:25:03,040 --> 00:25:07,280
iteratively building up the process. 
There is a fantastic explanation there by  

194
00:25:08,560 --> 00:25:17,200
Robert Hedan. I definitely recommend reading it. 
It is a really easy to follow explanation and it  

195
00:25:17,200 --> 00:25:26,880
would take more time than I have in the talk to 
walk you through it. It is really awesome. As you  

196
00:25:26,880 --> 00:25:31,760
can see from the output in that GIF you can do 
so much with this algorithm. It is fantastic.
 

197
00:25:33,600 --> 00:25:40,560
After that one we get to the big one -- 
Generative Adversarial Networks. They are  

198
00:25:40,560 --> 00:25:48,320
being used for a lot of things now. You may have 
seen a number of things out there at the moment.  

199
00:25:52,000 --> 00:25:58,400
This is something from the -- this Anime doesn't 
exist. I picked one at random there. Well,  

200
00:25:59,680 --> 00:26:07,200
they picked one at random and I picked one. They 
do solve a lot of problems and more every day  

201
00:26:07,200 --> 00:26:12,800
really. But they are complicated tools. While 
you can apply them to just about any problem,  

202
00:26:13,360 --> 00:26:19,200
many problems are not going to be useful for 
you as someone trying to get content out of it  

203
00:26:19,200 --> 00:26:25,520
unless you really understand what's going on.
You get artistic pictures like this one but you  

204
00:26:25,520 --> 00:26:33,920
can get a huge variety in output depending 
on what you put in. If you were aiming to get  

205
00:26:33,920 --> 00:26:39,360
images of people created for your world or game 
or whatever you are trying to do with a tool,  

206
00:26:40,000 --> 00:26:44,080
you would probably be looking for more like 
the left. But if you get parameters wrong  

207
00:26:44,720 --> 00:26:51,280
you will end up with abstract art like on the 
right. It might not be as straightforward as  

208
00:26:51,280 --> 00:27:01,120
creativity. If you are using a complex GAN it 
can be hours of work that needs to be redone  

209
00:27:02,000 --> 00:27:07,920
every time you make a change. If you 
are trying to dig into GANs I definitely  

210
00:27:07,920 --> 00:27:12,880
recommend studying up on it. Examples of how 
they are useful in the out of the box form  

211
00:27:13,600 --> 00:27:22,400
is this house-GAN that generates floor plans. It 
is someone's research paper. What you get out of  

212
00:27:22,400 --> 00:27:28,880
it is hundreds and hundreds of floor plans which 
can be a cool way to get inspiration for levels  

213
00:27:28,880 --> 00:27:35,280
in video games or other content where you need 
to generate a floor plan to lay things out in  

214
00:27:35,280 --> 00:27:41,760
the world. There is this library can is great 
library for actually generating content for  

215
00:27:42,400 --> 00:27:47,840
statistical samples. It will actually let you 
take a spreadsheet in. You can give it a CSV  

216
00:27:51,520 --> 00:27:56,480
and it will study it and produce a model for 
you automatically which is a lot more useful for  

217
00:27:56,480 --> 00:28:01,840
someone trying to get started. It is also pretty 
handy if you need to generate statistical samples.  

218
00:28:03,920 --> 00:28:08,960
Yeah, that's all the stuff I was trying 
to cover over. If you are curious about  

219
00:28:08,960 --> 00:28:16,320
digging deeper into anything here is some stuff 
I actually cut because I was very fragile or more  

220
00:28:16,320 --> 00:28:22,640
explanation required than I had time for. So if 
the remaining time has any questions, I am happy  

221
00:28:22,640 --> 00:28:27,760
to sort of go back and talk on any of them. 
>> Hello. We do have questions. Yes.  

222
00:28:28,880 --> 00:28:32,240
Feel free to drop more questions in the 
question chat. If we run out of time here,  

223
00:28:32,240 --> 00:28:37,120
we can answer them in the text channel 
afterward. The first is will slides  

224
00:28:37,120 --> 00:28:42,000
and links be shared online? 
>> Yes, I will put the links  

225
00:28:42,640 --> 00:28:51,360
in a cohesive document and share the slides 
>> The video is recorded and it will be up on  

226
00:28:51,360 --> 00:28:59,680
the PyCon AU YouTube channel I think. 
>> Yup. I see the one about the air pressure  

227
00:28:59,680 --> 00:29:06,320
and how random the data is. Yeah, that one 
definitely has some interesting details in it.  

228
00:29:06,880 --> 00:29:12,400
If you saw on the range shadow one, I don't 
know if it will switch fast enough, but if you  

229
00:29:13,120 --> 00:29:18,640
see on the precipitation one, what it does is it 
actually uses terrain height to affect things like  

230
00:29:19,760 --> 00:29:23,680
it doesn't have a pressure simulation but it 
definitely has thing thinks like rain shadowing.  

231
00:29:23,680 --> 00:29:29,360
There is a large mountain area here. You can see 
it is affecting how much rain we get in these  

232
00:29:29,360 --> 00:29:38,560
areas around it. It is definitely got a lot more 
variables in there. They do cascade. In fact, you  

233
00:29:38,560 --> 00:29:44,320
can't generate these things until after earlier 
layers are generated so it is folding these  

234
00:29:44,320 --> 00:29:51,680
different kind of data into each process. It is 
definitely taking those kind of data into account.  

235
00:29:53,520 --> 00:30:00,366
Since there is just a tiny bit of time I guess 
if I am going to answer the GAN question -- 
 

236
00:30:00,366 --> 00:30:07,840
>> The question is what's your favorite GAN? 
>> Yeah. At the moment it would be style GAN  

237
00:30:07,840 --> 00:30:14,320
because there is a lot of very interesting stuff 
coming out of it like you saw with the pictures  

238
00:30:14,320 --> 00:30:20,000
there. You know? It is very amazing you can get 
such a huge variety of output from a simple,  

239
00:30:20,000 --> 00:30:25,760
sort of tunable process. It is for the 
same data you have got everything from,  

240
00:30:25,760 --> 00:30:33,120
you know, water color artistic on the left 
to very, very abstract line art almost  

241
00:30:33,120 --> 00:30:39,920
engraving sort of style patterns on the 
right. It is fascinating to see how the same  

242
00:30:40,800 --> 00:30:48,640
core data and model can produce a huge variety of 
output. That's just the only one at the moment. 
 

243
00:30:50,080 --> 00:30:55,360
MODERATOR: Thank you so much, Sam. There were 
a couple questions left that I will paste in  

244
00:30:55,360 --> 00:30:59,600
the hallway text that version of Platypus 
afterwards if you want to answer those. 
 

245
00:30:59,600 --> 00:31:06,560
>> Happy to answer those. 
>> Next up at 3:30 AEST we have "Python  

246
00:31:06,560 --> 00:31:12,880
VS Space Junk" with Mars Buttfield-Addison. 
Have a nice break and thanks for hanging