GIS-PuppetMaster

it should be like this: self.cliped_ratio = tf.clip_by_value(self.ratio, 1. - METHOD['epsilon'], 1. + METHOD['epsilon']) self.min_temp = tf.minimum(self.ratio, self.cliped_ratio) self.aloss = -tf.reduce_mean(self.min_temp * self.tfadv)

> Why the negative value causes failure in actor loss? > You can also refer to OpenAI baselines [here](https://github.com/openai/baselines/tree/master/baselines/ppo1), which has similar process as our repo. I drawed the loss...

buffer_r.append((r+8.1)/8.1) # normalize 为什么r需要标准化？ 8.1是怎么来的？

> Hi, > It supports dict state, but you need a wrapper for your env. > Please take a look at the FlattenDictWrapper (./common/env_wrappers.py) for robotics env. Thanks, I think...

最大化卡顿，窗口化好了很多，好像越小越流畅

Is it possible? I think PPO v2 is not strictly an 'off-policy' algorithm.

This is only a demo version, it seems Graphviz always displays the attribute with center alignment since this tool uses caption to contain attributes. Better to display them with left...

same problem here, AVIF photo that exported by LightRoom with HDR information looks gray