/aig/ - Artificial Intelligence General

>>58457022
Oh and here, my contribution as OP:

I trained a wavenet implementation to try to generate human speech. The results are... halfway there.

https://files.catbox.moe/nwgc7v.wav

File: rmLKsel.jpg (112KB, 1920x1080px) Image search: [Google]

112KB, 1920x1080px

>>58457065
Right now I'm training it on https://www.youtube.com/watch?v=93RvNGbnlXo&t and all of her other albums and hopefully it'll make some funky music.

Shit why is /g/ so fast now?

>>58457022
>Teaching a computer to see:
>https://www.youtube.com/watch?v=i8D90DkCLhI&list=PLiaHhY2iBX9ihLasvE8BKnS2Xg8AhY6iV

That shitty series is just a crappy theoretical history lesson with basically no practical substance.

Other than the first episode which explains the problem you can basically skip all episodes I've watched so far. I'm at part 8 and that person is still talking about irrelevant things.

>>58458023
lol sorry about that. Do you think that I should remove it from the header?

bump, doesn't /g/ care about actual computer science?

what's a simple and fun project I could get working in less than an evening? I've watched all of Andrew Ng's course and wrote a bare bone regression neural network in numpy/python. Willing to go straight to theano and those other tools that make it really easy.

>>58457065
You got a drowning simulator man that's cool.

Or a banjo kazooie npc voice sim if you slow it down.

>>58457022
>Interesting evolution algorithm for simulated locomotion:
If you actually read the paper for this it reveals that there was actually a lot of hand-crafting that went into achieving those results and that the methods they used were in no way general, i.e. you can't apply the same methodology to spider-like robots.

Very misleading video.

>>58461438
try to make a basic classifier that can tell manga/animoo from non-manga images.

Does this make sense?

inputs = Input(shape=(sequence_length,), dtype='int32')
x = Embedding(output_dim=embeddings_size, input_dim=len(dictionary), input_length=sequence_length)(inputs)
x = Dropout(0.25)(x)

x1 = Bidirectional(GRU(32, return_sequences=True))(x)
x2 = Bidirectional(GRU(32, return_sequences=True))(x1)
x3 = Bidirectional(GRU(32, return_sequences=True))(x2)
x = merge([x1,x2,x3], mode='concat')
x = Dropout(0.25)(x)

x = TimeDistributed(Dense(output_dim=1, activation='sigmoid'))(x)
outputs = Reshape((sequence_length,))(x)

ie connecting all recurrent layers to the output instead of the last one

I'd check myself but training takes too long so I just want to hear an advice.

>>58461254
No, that's what /prog/ is for. /g/ is for desktop threads, anime, and consumer electronics.

>>58462832
But I don't wanna go back to lain! It's slow and my custom cs don't work!

File: 3mWDngD.png (25KB, 227x257px) Image search: [Google]

25KB, 227x257px

>>58461438
Train the network to create an image on the right from the image on the left.

I uploaded the training set for you (those are icons from the game Stellaris):
https://mega.nz/#!ZUlhibLK!AZZmbpAuyzLY4aH7aX-S88PoI2CxJiJnGFiwfCsqQI0

It is extremely simple and very rewarding. It's the only NN-related thing I did from scratch that works really well.

>>58462842
install gentoo

or just tell An Hiro to bring back the textboards

>>58462856
>Train the network to create an image on the right from the image on the left.

That has nothing to do with neural networks, you fucking retard. Kill yourself.

Total params: 576,769
Trainable params: 576,769
Non-trainable params: 0
____________________________________________________________________________________________________
Train on 204537 samples, validate on 10765 samples
Epoch 1/1000
204537/204537 [==============================] - 8778s - loss: 0.0165 - predicted: 527.0602 - correctly: 297.3716 - missed: 520.5564 - val_loss: 1.9946e-04 - va
l_predicted: 36.2254 - val_correctly: 30.4704 - val_missed: 12.1758
Epoch 2/1000
204537/204537 [==============================] - 8716s - loss: 0.0033 - predicted: 697.8041 - correctly: 599.9527 - missed: 217.9762 - val_loss: 1.7904e-04 - va
l_predicted: 37.1766 - val_correctly: 31.7546 - val_missed: 10.8916
Epoch 3/1000
204537/204537 [==============================] - 8753s - loss: 0.0029 - predicted: 728.5858 - correctly: 637.1651 - missed: 180.7629 - val_loss: 1.8324e-04 - va
l_predicted: 35.5833 - val_correctly: 30.8271 - val_missed: 11.8190

>that feel that validation set loss is smaller than training set loss
>that feel when an epoch takes two and a half hours

>>58462922
Please elaborate.

>>58462922
lol wut

File: 1483342988862.png (379KB, 689x752px) Image search: [Google]

379KB, 689x752px

>>58462922

robot waifu the dreams

>>58464669
You heard the voice at >>58457065 didn't you? I'm trying damnit!!

>>58464724

sounds like animal crossing npc's

>>58457065
>>58457099
What code did you use? Please share. I read about wavenet and its porous convolution thing, but I have no idea what they mean by that.

>>58462923
Minibatch and use less epochs?

damn hackerrank challenge is killing me lol
whats a good cost function for this thing

https://www.hackerrank.com/challenges/the-bidding-game

>>58465811
I'm using batch size of 512. Making it smaller almost does not decrease time per batch so I decided that there's not much sense to it.

Why do they make convolutional neural networks sound so complicated when they're literally just performing a sliding window on the input?

>>58466097
I don't think they are making convolutional neural networks sound complicated.

>>58465118
I used this

https://github.com/ibab/tensorflow-wavenet

>>58466227
They are trying to make them sound convoluted.