github.com/HyperGAN/HyperGAN commits | Ecosyste.ms: OpenCollective

[fix] fitness_trainer working with new structure

6679cb361edebded637f75ae0da89c6f0a808380 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] create f_decay and gradient_scale once in elastic_weight_consolidation

bd326360ad971bd059b1ecc3b7e6a406daee45c1 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] weight penalty train hook with l2nn/ortho support

a9f5fc91ce6f85c85bbd5bf054efe7d9ca8fc91e authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] curl / ewc configurable params

39ce7cdc5390a3f3801ede56507fc2a3ef803109 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] configurable_component - reference looks at layers if they exist

98e86ae46eb6626f2f5a9aa7fe25bcbe9db1efce authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] classification config

ad8ed2f0af29bbacaa6775c8f71a00b74f49e5a5 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] stride as a list in configs

73a026442816275db5f331af49dad22013f1c0c3 authored almost 6 years ago by Mikkel Garcia <[email protected]>

Merge pull request #217 from HyperGAN/develop

0.10 alpha 22

aa847e0c12f854ca95ac2d86e3dfa7cb7309410e authored almost 6 years ago by mikkel <[email protected]>

Merge branch 'master' into develop

e278e86d17a07b284fcd1299752e6da6e4e0267a authored almost 6 years ago by mikkel <[email protected]>

[experiment] option.mask_with for combining layers

a5a99afa5b4e53587989f5b4964da05956ee9408 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] avg_pool=N on layer

c5da4e1bcffe882c38a2e43a8a3c718776d9e468 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] config looks in default as well as the base namespace

fc933f0c41091651f8339331446fc8e06b000632 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] gradient_penalty, config.flex can be an array.

025a050890969ab9bdc1ce622802aefefa21770d authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] add missing abs call for flex: square(relu(abs(gradient)- flex))

29636f8507845c8a944a2cc89874c104f4a9d4a0 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] gradient_penalty_train_hook config.flex lets gp vary from zero

346c8fd14fe4f5855f8c7e7651473d8775d5c0cd authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] gradient_penalty.lambda is a configurable_param

1c34374957e6ccfb05e5c33becf4fcd88c5bab86 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] gradient penalty defaults to being applied to all gan variables, not just discriminator variables

c6b5cf7e8b7ed1e96465b354d42451012c1876ad authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] default decay() configurable param to range[0]

e419248f4f7ea307dd6ea255eb9eda661360b661 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] skip discriminator concatting x/g if different sizes (for classifier)

4ebb637738376b544940b5ddfd40db9395277547 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] examples/classification. Train a generator than classifies

3b680bcaaa013e78f0f2135ca75d7cd2d481a9ce authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] specify components with defaulting to discriminator target in gradient_penalty_train_hook

b3d8d2b709f93f9ce808b44e0946b4e00d019c55 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] configurable_param allows parameters to change over training steps

6f326abbf9b943152a25fd12ec53f89514e442be authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] update stylegan config to stabler version

4021b3b5651b0440a3ae838ddf4417ac53dac41a authored almost 6 years ago by Mikkel Garcia <[email protected]>

[wip] alignment

1193d7e26ae202ff26e5fee1296552fb076dd983 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] allow layer name lookup across components

e3d701eaca495347ee4484cdc52ba9736388a122 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] alpha channel in tk_viewer would crash sampler

000bfadfb896d3432d08b6f1ba58a3f601219159 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] batch size 1 on 2d-distribution maps contours/points

54d7a5d515f1df71bbe9fe0e49fae3a09fc0db94 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] better visualization of 2d example

521579b4cc2ba7c1ec865e847ab3d4cff56ac79f authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] 2d-distribution example

8c8a0015a5b43acd992846436c3ad494c549a8cb authored almost 6 years ago by Mikkel Garcia <[email protected]>

[experiment] alignment

c46bce2ee4a2afad38abaacdb4c104c0b50eb447 authored almost 6 years ago by Martyn Garcia <[email protected]>

Merge pull request #216 from HyperGAN/develop

0.10 alpha 21

2c4df74bb308b931a3cbb63073a83679847a3b3a authored almost 6 years ago by mikkel <[email protected]>

Merge branch 'master' into develop

8912506c8aedc47f56c6b589b7f07b8e65804191 authored almost 6 years ago by mikkel <[email protected]>

[feature] stylegan config

f3dac197b8f2548b1b668ab88ff80f644805f3a9 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] add initializer= and trainable= options to 'const'/'noise learned' layers.\n[fix] Move stylegan weight multiplier to get_weights

63d5c4449077d9f0124e93bd4fff2d5ba5e2c2e8 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] initializers were not being parsed from config.default

5b01095c5a117a5577bae6459fe936e16f99f5f8 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] output number of parameters per layer during network construction

1e0015801eff833b5d00838a0e1c1413c5f0b6b5 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] interesting and small colorizer config for 1024x1024

d4814c5908f9b6f65e54ff677e8003731c8946a5 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] interesting colorizer-resizeconv config

548b8a04d5494f6451b5550d8395d48e82a198bf authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] output training command with "hypergan new -l ."

81a05e803fc468ccaccaa1bdd3537a9198c94e6f authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] remove resolution from config filenames.

02064c4aed4ac1014c49e2be424adfd55cc5c7a2 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] configs: fixed spacing, added runtime section, removed experimental

f03e03d46d4512869d047e90b03f0b09994abf8a authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] static batch + imle sampler

5a2d3754edad8c79644befc008cbd559618c5aab authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] runtime.train command for colorizer-64x64

b7a337a57979c8b3438cc56321d5865a15e4bd1b authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] better colorizer-256x256 config

6ce7d4f4db6b85a5391bc5d15a90e2831d5e74a1 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] colorizer-64x64 layer filter added to discriminator [refactor] sampler moved to basegan

08540b26caf5a1c777602c7643008960a8975a89 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] wgan-gp gradient penalty scaled up and bias removed on last layer of D. This more closely matches the r1 regularizer from stylegan

b68ddfc0a9d9ebb99731d74035349cd0c8b7abfe authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] remove tf.norm from gradient_penalty_train_hook. Add options for which loss and which components gradient penalty should be applied to.

5ee0f6c5fc31124eddffedddb7c2ce816dd329c8 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] bias=false option for configurable layers

fe8a6186e9c10fd91bf2e802741d8d307c58dd78 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] gitignore entries

d0af8a0cfaffc0f77cca3a4d34252db8352b2b23 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] amsgrad optimizer

1515a2aeb8141a19bf3b48e74f22912f2767992c authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] aligan example config

e479a84dc311792117a6aafc2b9aec17a9f1d290 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] link to stylegan as state of the art

6dcb8f2a24a01981ade3c91f333f12a4ac0c0022 authored almost 6 years ago by mikkel <[email protected]>

[feature] layers concat and add expanded

6adad567b5851710e765208488a16815fb35d57c authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] layer_noise mask and concat for resize_conv

692a9009351bac1efd0cfd63148699b02fa84bca authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] imports

407994fbbefabf1689424eddc8ddf0d4a70964b4 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] routes after refactoring /experimental components

1d159fbd6c02a5943fd94614d255c062935d1d7b authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] additional loss terms moved from base_loss to individual train_hooks

855b745b5fe33decf648f80d298d916601999eaa authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] options without example configuration go to /experimental

013ccad493230d9ae1d476302a415279665dae5e authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] options without example configuration go to /experimental

cc482e14d68ae894d4a9f697d4a75f8401fa7da5 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] more stable progressive growing (still wip)

4c2edba49a3e56318b7486f74c54115fcf7955cb authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] ewc_optimizer global options initial_constraint, gradient_scale. layer options ewc_gradient_scale, ewc_f_decay

5637b31beefdee3d9bc557666ffeafaaad6db2b2 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] default graph and current gan graph get out of sync with curriculum trainer

8599cc04c7f37d6c7726d30d07533ad9e5dc5c6d authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] function reference

bbe34c3fd2df6568498df8502058af42d385296a authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] syntax error

f8ffb90072110f5407e3cd4f5ba5109988008263 authored almost 6 years ago by Martyn Garcia <[email protected]>

[feature] gelu activation https://arxiv.org/pdf/1606.08415.pdf

8838125d6a6b0b79991040c9475f66fa46a42ad9 authored almost 6 years ago by Martyn Garcia <[email protected]>

Merge pull request #215 from HyperGAN/develop

0.10 alpha 20

0513e0771e5c810d95d5291f0c22a4900febf728 authored almost 6 years ago by martyn <[email protected]>

Merge branch 'master' into develop

2237994c2f96b882b2d12bebdd2e8e59e8abed44 authored almost 6 years ago by martyn <[email protected]>

[fix] hypergan sample

317ff7c072d10d864b2297fcead866925870caf7 authored almost 6 years ago by Martyn Garcia <[email protected]>

[feature] initial Sampler submenu

85adace557eee24bbaf962c1d53e7f3501e17ec6 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] refresh sample shortcut

2b7ee178fe8982cee63b905659f23feaf43e2abd authored almost 6 years ago by Martyn Garcia <[email protected]>

[feature] initial Sample menu

348347f5dadf1672f081cff1f8ffdd230cc04ef8 authored almost 6 years ago by Martyn Garcia <[email protected]>

[feature] aligan config.type 'logistic'

6ef68158d9cc23cd08be23925226c3408cf483ea authored almost 6 years ago by Martyn Garcia <[email protected]>

[feature] logistic loss

9ba71caec5e2687732c950ba00e5e71d9b75df35 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] gang trainer and sampler

af59785923c30f2baf028b55fe74bf9c95aa2f98 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] gang configuration https://arxiv.org/abs/1712.00679

df36f637220ce68a857754cc33b51c4e164fc13a authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] progress-compress configuration https://arxiv.org/abs/1805.06370

52a408bab9ad2aed8fcafe32abdc878240bce6b8 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[refactor] remove untrue comment

f4549825a6160c5c71ddc61900119695cf514535 authored almost 6 years ago by mikkel <[email protected]>

Merge pull request #210 from HyperGAN/develop

0.10 alpha 19

ed55ac68a3e812a69b8d0d2df1e8289bb880903c authored almost 6 years ago by mikkel <[email protected]>

Merge branch 'master' into develop

5e1b42c9df71153201aa4eaf98b772439446f132 authored almost 6 years ago by mikkel <[email protected]>

[fix] progressive config transitions

67fc14bd24cd5d5d0a15c1937fbf430ef61caef8 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] scale large icons to avoid out of memory

52dfb3b81b0d7fa27db9acf8033f28f31f8290ad authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] example of progressive growing type config (wip).

5096cd38f2fcca196bddf2592ba37a7830e4c050 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] norm: false for gradient magnitude optimizer, for controlling the gradient lambdas without tf.norm

21ba64145ebdb1fd77dea3fec2f1abf679b487de authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] elastic weight consolidation variable list lookup

6e03ea9d84923fe398aa8e43ef37b4ec2f424404 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] dont redefine reference layers in named_layers for configurable_component dsl

b843dc7a976b5ce616ececd4030a8fb5605a28ca authored almost 6 years ago by Mikkel Garcia <[email protected]>

Merge pull request #211 from MaffooBristol/patch-1

Fixed imagemagick command in README

76c779162db9e52ce3f8301d7325ed5a1527cacd authored almost 6 years ago by martyn <[email protected]>

Fixed imagemagick command in README

It should be `; do convert` not `; do; convert`. The latter gives a syntax error.

5f7f76104ace30b30fe8c0738b204f72fb295bad authored almost 6 years ago by Matt Fletcher <[email protected]>

[fix] large samples in tkviewer icon(oom)

b83a0fe52d116f3f510b2feacd2c451e6c7796af authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] self-supervised-64x64 config example

36dbb3d4170d556f87ae0bbc627e7b3ff7545f3b authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] wgan-gpsn example

90c04ad63affc7de77b11529b6fecbf10fa9ddbc authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] custom names for tensors. Breaks any configurable configurations with name= layers

2618233a6adbe197c413adecee5908c80c8f3cd0 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[feature] configurable_component layer_options(net) tracks options

d7d704b9929e8f9efb250e17e978eebdca026f48 authored almost 6 years ago by Mikkel Garcia <[email protected]>

Merge branch 'master' into develop

fda500299ec9b2c3f39f9df467e85358663934f0 authored almost 6 years ago by mikkel <[email protected]>

[fix] tk viewer icon and title

ffe41a7b7d0b3d24672b0dcacec8eb65fcd0574c authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] tests

f19404197fa34b80540550c248ee276947169a32 authored almost 6 years ago by Mikkel Garcia <[email protected]>

[fix] circleci tests

03a35637f9940b5d718d65feaf7c49dfffbd8407 authored almost 6 years ago by Mikkel Garcia <[email protected]>

Merge pull request #209 from HyperGAN/develop

0.10 alpha 18

282485e272e93663edbf36e1072d0eb2e1f76e33 authored almost 6 years ago by martyn <[email protected]>

Merge branch 'master' into develop

0a979d4c8d818521c8787887f4dc941445a585ca authored almost 6 years ago by martyn <[email protected]>

[fix] prevent training if save is never triggered

24fab23ae28c1d544115025390594d495dd4b240 authored almost 6 years ago by Martyn Garcia <[email protected]>

[fix] more tests

ac8154eb30b835ee99e131f08eed95dbefb53d74 authored almost 6 years ago by Martyn Garcia <[email protected]>