3
99%
master: 99%

Ran 29 Aug 2019 03:00PM UTC

Files 6

Run time 1s

Badge

Embed ▾

Committed 29 Aug 2019 02:50PM UTC coverage: 79.003% (-0.4%) from 79.42%

Job # KERAS_BACKEND=tensorflow TF_KERAS=1

Build Type

push

travis-ci

Committed by

CyberZHG

Commit Message

Change lr decay logic (#26)

* Update optimizers.py

Changed `lr` decay logic

Once `t > warmup_steps`, the movement toward `min_lr` is abrupt. _Ex:_ `self.total_steps = 200`, `warmup_steps = 100`, `self.lr = 1e-3`, `self.min_lr = 1e-4`:

- `t = 99 `  --> `lr = 9.9e-4`
- `t = 100` --> `lr = 1.0e-3` 
- `t = 101` --> `lr = 5.455e-4`

Doubt this is intended behavior. Plot ([code](https://puu.sh/E8HuR/a07ccd9e6f.txt)) comparison below:

<hr>

![image](https://user-images.githubusercontent.com/16495490/63619016-96bc7480-c5bb-11e9-9e91-e32a2a8ebde4.png)

* Fixed to account for decay_steps = 0

* Fixed styling to appease Travis

* Update optimizers.py

Run Details

301 of 381 relevant lines covered (79.0%)

0.79 hits per line

CyberZHG / keras-radam / 47 / 3
99%
master: 99%

README BADGES
x

Markdown

Textile

RDoc

HTML

Rst

Source Files on job 47.3 (KERAS_BACKEND=tensorflow TF_KERAS=1)

CyberZHG / keras-radam / 47 / 3 99% master: 99%

README BADGES x

Markdown

Textile

RDoc

HTML

Rst

Source Files on job 47.3 (KERAS_BACKEND=tensorflow TF_KERAS=1)

CyberZHG / keras-radam / 47 / 3
99%
master: 99%

README BADGES
x