deep emoji generative adversarial network
(trying to) generate new emojis with DCGAN 🤗🏭
usage
The emojis are taken from a git submodule to initialize it after cloning this repo run:
git submodule init
git submodule updateThe code itself is currently hosted in a jupyter notebook so you may run jupyter notebook to access the latest version of the GAN and run all the cells to learn the network.
development
keeping track of different network designs and hyperparameters
15817e6
generator design
convolutions: 4
features: 512 > 256 > 128 > 64 > 4
kernel size: 5
discriminator design
convolutions: 3
features: 64 > 128 > 256
kernel size: 5
hyper params
training set: 225 (people no tones)
epochs: 768
learning rate: 0.0002
batch size: 64
opt.beta: 0.4
result
fa696d4
generator design
convolutions: 4
features: 256 > 128 > 32 > 4
kernel size: 5
discriminator design
convolutions: 3
features: 32 > 128 > 256
kernel size: 5
hyper params
training set: 714 (people & activity)
epochs: 768
learning rate: 0.0003
batch size: 256
opt.beta: 0.5
result
e8285ca
generator design
convolutions: 4
features: 256 > 128 > 32 > 4
kernel size: 5
discriminator design
convolutions: 3
features: 32 > 128 > 256
kernel size: 5
hyper params
training set: 714 (people & activity)
epochs: 4096 (only 1200 run?)
learning rate: 0.0003
batch size: 256
opt.beta: 0.5
result
151284d
generator design
convolutions: 3
features: 64 > 32 > 16
kernel size: 4 > 6 > 8
discriminator design
convolutions: 3
features: 8 > 16 > 32
kernel size: 8 > 6 > 4
hyper params
training set: 714 (people & activity)
epochs: 4096 (only 1200 run?)
learning rate: 0.0003
batch size: 256
opt.beta: 0.5
result
9637353
generator design
convolutions: 4
features: 1024 > 512 > 128 > 64
kernel size: 3 > 5 > 5 > 7
discriminator design
convolutions: 3
features: 16 > 46 > 256
kernel size: 5 > 4 > 3
hyper params
training set: 714 (people & activity)
epochs: 1024
learning rate: 0.0002
batch size: 1289
opt.beta: 0.5
result
21b7da3
generator design
convolutions: 4
features: 256 > 128 > 32 > 4
kernel size: 5
discriminator design
convolutions: 3
features: 32 > 128 > 256
kernel size: 5
hyper params
training set: 1262 (no regionla, no symbols)
epochs: 1000
learning rate: 0.0003
batch size: 256
opt.beta: 0.5
result
0344c27
generator design
convolutions: 4
features: 256 > 128 > 32 > 4
kernel size: 5
discriminator design
convolutions: 3
features: 32 > 128 > 256
kernel size: 5
hyper params
training set: 5063 (multi-set, no regionla, no symbols)
epochs: 1600
learning rate: 0.0003
batch size: 256
opt.beta1: 0.4
opt.beta2: 0.7
result
Best sample
Tuning Hyperparameters
training set: 1565 (multi-set, people)
generator design
convolutions: 4
features: 256 > 128 > 32 > 4
kernel size: 5
discriminator design
convolutions: 3
features: 32 > 128 > 256
kernel size: 5
| learnrate | epochs | opt.beta1 | opt.beta2 | alpha | batchsize | result |
|---|---|---|---|---|---|---|
| 0.0001 | 256 | 0.5 | 0.7 | 0.2 | 256 | disc ~0.3, gen ~2 |
| 0.0001 | 128 | 0.4 | 0.7 | 0.2 | 128 |  |
| 0.0001 | 128 | 0.4 | 0.7 | 0.2 | 512 |  |
| 0.0001 | 128 | 0.4 | 0.4 | 0.2 | 512 |  |
| 0.0001 | 128 | 0.8 | 0.5 | 0.2 | 512 |  |
| 0.0001 | 128 | 0.9 | 0.999 | 0.2 | 512 |  |
| 0.0001 | 128 | 0.4 | 0.999 | 0.2 | 512 |  |
| 0.0001 | 256 | 0.4 | 0.999 | 0.2 | 512 |  |
| 0.00001 | 128 | 0.4 | 0.999 | 0.2 | 512 |  |
6dfe858
generator design
convolutions: 3
features: 128 > 64 > 4
kernel size: 5
discriminator design
convolutions: 2
features: 64 > 128
kernel size: 5
hyper params
training set: 1 (one round shocked face)
epochs: 1000
learning rate: 0.003
batch size: 32
opt.beta: 0.5
result
63abc3b
generator design
convolutions: 2
features: 1024 > 256 > 4
kernel size: 5
discriminator design
convolutions: 2
features: 64 > 256
kernel size: 5
hyper params
training set: 1 (one round shocked face)
epochs: 5000
learning rate: learning_rate_d=0.0003, learning_rate_g=0.001
batch size: 32
opt.beta: 0.5
result
8c948e8
generator design
convolutions: 3
features: 512 > 128 > 64
kernel size: 5
discriminator design
convolutions: 3
features: 64 > 128 > 512
kernel size: 5
hyper params
training set: 1 (one round shocked face)
epochs: 200
learning rate: learning_rate_d=0.0002, learning_rate_g=0.0002
batch size: 768
opt.beta: 0.5
result
Note: The goal of this run was to proof that a DCGAN is able to train on a single image and will end up replicating this image. This was a way of testing the overall chain and exposed a bug in the data preparation methods.
86359f9
generator design
convolutions: 3
features: 512 > 128 > 64
kernel size: 5
discriminator design
convolutions: 3
features: 64 > 128 > 512
kernel size: 5
hyper params
training set: 273 (1565 filtered for being yellow)
epochs: 800
learning rate: learning_rate_d=0.0003, learning_rate_g=0.0003
batch size: 768
opt.beta: 0.5
result
Degradation of diversity
What's interesting is that the network managed to somehow learn diverse features and put them together and at epoch ~600 got scrambled and forgot some of the features like a sticked out tongue.
The network was trained with several emojis of this type:
Emoji shape forming at epoch50
First details emerging at e150
Diversity in the generated images at e600
Something creating a lot of noise at e650
Final result at e800 with less features than e600
7e1480d
generator design
convolutions: 3
features: 256 > 128 > 64
kernel size: 5
discriminator design
convolutions: 3
features: 64 > 128 > 256
kernel size: 5
hyper params
training set: 141 (only "face")
epochs: 800
learning rate: learning_rate_d=0.0001, learning_rate_g=0.0001
batch size: 768
opt.beta: 0.5
result
Evolution over epochs
Losses
Epoch200
Epoch 450
Final sample
