GANs
This repository implements feedforward Generative Adversarial Networks and an experimental form of recurrent GAN. It includes some of the work done at OpenAI on GANs.
Basic results
As a quick test, I trained a small GAN to produce MNIST digits. Here is the result:
I also built a deeper GAN for extracting features from face images. The results can be found in MustacheMash
Recurrent GANs?
After a few different attempts, I still have not managed to get a recurrent GAN to produce reasonably English-sounding text. I will probably pick this up again in the future. Here, I will document some of the results I have seen while trying to generate text with a GAN.
Random outputs. Obviously, when the GAN first starts up, the generator produces random outputs. These outputs include plenty of non-ASCII characters, so I will not include them here.
Homogeneous outputs. Often, the generator will output a few letters constantly, ignoring most other characters. For example:
TTTTTTTTTTeTTTTTTTTNTTTTTITTHTTTTHTTTTTTeTTTTTTeTTTTTTTTTTTTTTTTTTTTTTTTTTTTTeTTTTeTTHTTTTTTTTTTTTTTTTTTeTTTTTTTTTTTTTTTTTTTTTTTeTTTTTTTeTTTTTTATTTTTT
eTTTTTTTTTTTTTTTTTTTTTTTTTTTTTlTTTTTTTTTTTTTTTTTTBTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTMTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTeTTTTYTTTTeTTTTTTTBTTTTTTTTTTT
Lots of spaces. Occasionally, the generator learns to output way more spaces than are present in English text, resulting in something like this:
WhA IIAto tPI A EAP S d B P A A e HA IRPJP uI I P P P n IP o A AiHT TA P P n P PH t P A sY BA hP R A
AP IiP A AP I D A P P cA E A n A Sh A PIP O hN AIPI Eeei aPA s PBPe P lP P PPAA P AK h A H AeHt e Iao h o
Starting with a capital. Usually, the generator eventually learns to start its outputs without capital letters (since the samples all look that way). Outputs might look like so:
Tlnpsndywwwdpkddc.ndwnwwnywppvhwmsdalldwvpwwswlwwvwndwpddwatsnEntosswwdywInw fsswl.pwndds.pdx pllylw.fn.ob.â.swwbywwcdlwdwwvnnndwdwwph ddelnbsdwpnwpb
Iodrwwwnwdfb.o.nteodewwdstdwwevtkFpwdflnfrv\wslpnÇebddwwppbwwlvwdTpdwsnnopdbtynfBwwfdww.eywvnp.pvwOddtdklcnlwfnnwytnyP³ voppldywwwbv.ywdeldywcwwtbwlws
Lots of dots. This is a case wherein the generator collapses to output a few consonants and dots. Outputs look like so:
ht.......hh.....hhuh....hthuhh...uuhtt.ht.huht.thohhh..ehuhuht..thhthhthrhh..thhu.u.hhutt...hnhh...htuht..hthht...hhhuut..hhuu..hahhhtrt..hhht...hhhhu
tt.....hhh....hhhut.tthhhut.t.tuuuu....huhh...thhhhh....huhrrh.tthhttn.ttuhth .huhh..hhhhuht..hthht.hhtauht..hhhhhr..hhht..t.uthhu..tthh ..htthh. u.hh
Homogeneous with 1st capital. Enough said:
Aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
Aaaaaaaaaaaaaaaaaaaaaaaakaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
GAaaaaaaaaaaaaaaaaaaaaaaiaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent