mirror of
https://github.com/gryf/coach.git
synced 2025-12-17 19:20:19 +01:00
6d40ad1650
* updating the documentation website * adding the built docs * update of api docstrings across coach and tutorials 0-2 * added some missing api documentation * New Sphinx based documentation
620 lines
72 KiB
HTML
620 lines
72 KiB
HTML
|
|
|
|
<!DOCTYPE html>
|
|
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
|
|
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
|
|
<head>
|
|
<meta charset="utf-8">
|
|
|
|
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
|
|
|
<title>rl_coach.agents.ppo_agent — Reinforcement Learning Coach 0.11.0 documentation</title>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<link rel="stylesheet" href="../../../_static/css/theme.css" type="text/css" />
|
|
<link rel="stylesheet" href="../../../_static/pygments.css" type="text/css" />
|
|
<link rel="stylesheet" href="../../../_static/css/custom.css" type="text/css" />
|
|
<link rel="index" title="Index" href="../../../genindex.html" />
|
|
<link rel="search" title="Search" href="../../../search.html" />
|
|
<link href="../../../_static/css/custom.css" rel="stylesheet" type="text/css">
|
|
|
|
|
|
|
|
<script src="../../../_static/js/modernizr.min.js"></script>
|
|
|
|
</head>
|
|
|
|
<body class="wy-body-for-nav">
|
|
|
|
|
|
<div class="wy-grid-for-nav">
|
|
|
|
|
|
<nav data-toggle="wy-nav-shift" class="wy-nav-side">
|
|
<div class="wy-side-scroll">
|
|
<div class="wy-side-nav-search">
|
|
|
|
|
|
|
|
<a href="../../../index.html" class="icon icon-home"> Reinforcement Learning Coach
|
|
|
|
|
|
|
|
|
|
<img src="../../../_static/dark_logo.png" class="logo" alt="Logo"/>
|
|
|
|
</a>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<div role="search">
|
|
<form id="rtd-search-form" class="wy-form" action="../../../search.html" method="get">
|
|
<input type="text" name="q" placeholder="Search docs" />
|
|
<input type="hidden" name="check_keywords" value="yes" />
|
|
<input type="hidden" name="area" value="default" />
|
|
</form>
|
|
</div>
|
|
|
|
|
|
</div>
|
|
|
|
<div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<p class="caption"><span class="caption-text">Intro</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../usage.html">Usage</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../features/index.html">Features</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../selecting_an_algorithm.html">Selecting an Algorithm</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../dashboard.html">Coach Dashboard</a></li>
|
|
</ul>
|
|
<p class="caption"><span class="caption-text">Design</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../design/control_flow.html">Control Flow</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../design/network.html">Network Design</a></li>
|
|
</ul>
|
|
<p class="caption"><span class="caption-text">Contributing</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../contributing/add_agent.html">Adding a New Agent</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../contributing/add_env.html">Adding a New Environment</a></li>
|
|
</ul>
|
|
<p class="caption"><span class="caption-text">Components</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/agents/index.html">Agents</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/architectures/index.html">Architectures</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/environments/index.html">Environments</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/exploration_policies/index.html">Exploration Policies</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/filters/index.html">Filters</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/memories/index.html">Memories</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/core_types.html">Core Types</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/spaces.html">Spaces</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../components/additional_parameters.html">Additional Parameters</a></li>
|
|
</ul>
|
|
|
|
|
|
|
|
</div>
|
|
</div>
|
|
</nav>
|
|
|
|
<section data-toggle="wy-nav-shift" class="wy-nav-content-wrap">
|
|
|
|
|
|
<nav class="wy-nav-top" aria-label="top navigation">
|
|
|
|
<i data-toggle="wy-nav-top" class="fa fa-bars"></i>
|
|
<a href="../../../index.html">Reinforcement Learning Coach</a>
|
|
|
|
</nav>
|
|
|
|
|
|
<div class="wy-nav-content">
|
|
|
|
<div class="rst-content">
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<div role="navigation" aria-label="breadcrumbs navigation">
|
|
|
|
<ul class="wy-breadcrumbs">
|
|
|
|
<li><a href="../../../index.html">Docs</a> »</li>
|
|
|
|
<li><a href="../../index.html">Module code</a> »</li>
|
|
|
|
<li>rl_coach.agents.ppo_agent</li>
|
|
|
|
|
|
<li class="wy-breadcrumbs-aside">
|
|
|
|
</li>
|
|
|
|
</ul>
|
|
|
|
|
|
<hr/>
|
|
</div>
|
|
<div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
|
|
<div itemprop="articleBody">
|
|
|
|
<h1>Source code for rl_coach.agents.ppo_agent</h1><div class="highlight"><pre>
|
|
<span></span><span class="c1">#</span>
|
|
<span class="c1"># Copyright (c) 2017 Intel Corporation</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># Licensed under the Apache License, Version 2.0 (the "License");</span>
|
|
<span class="c1"># you may not use this file except in compliance with the License.</span>
|
|
<span class="c1"># You may obtain a copy of the License at</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># http://www.apache.org/licenses/LICENSE-2.0</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># Unless required by applicable law or agreed to in writing, software</span>
|
|
<span class="c1"># distributed under the License is distributed on an "AS IS" BASIS,</span>
|
|
<span class="c1"># WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.</span>
|
|
<span class="c1"># See the License for the specific language governing permissions and</span>
|
|
<span class="c1"># limitations under the License.</span>
|
|
<span class="c1">#</span>
|
|
|
|
<span class="kn">import</span> <span class="nn">copy</span>
|
|
<span class="kn">from</span> <span class="nn">collections</span> <span class="k">import</span> <span class="n">OrderedDict</span>
|
|
<span class="kn">from</span> <span class="nn">typing</span> <span class="k">import</span> <span class="n">Union</span>
|
|
|
|
<span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">rl_coach.agents.actor_critic_agent</span> <span class="k">import</span> <span class="n">ActorCriticAgent</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.agents.policy_optimization_agent</span> <span class="k">import</span> <span class="n">PolicyGradientRescaler</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.architectures.embedder_parameters</span> <span class="k">import</span> <span class="n">InputEmbedderParameters</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.architectures.head_parameters</span> <span class="k">import</span> <span class="n">PPOHeadParameters</span><span class="p">,</span> <span class="n">VHeadParameters</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.architectures.middleware_parameters</span> <span class="k">import</span> <span class="n">FCMiddlewareParameters</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.base_parameters</span> <span class="k">import</span> <span class="n">AlgorithmParameters</span><span class="p">,</span> <span class="n">NetworkParameters</span><span class="p">,</span> \
|
|
<span class="n">AgentParameters</span><span class="p">,</span> <span class="n">DistributedTaskParameters</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">rl_coach.core_types</span> <span class="k">import</span> <span class="n">EnvironmentSteps</span><span class="p">,</span> <span class="n">Batch</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.exploration_policies.additive_noise</span> <span class="k">import</span> <span class="n">AdditiveNoiseParameters</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.exploration_policies.categorical</span> <span class="k">import</span> <span class="n">CategoricalParameters</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.logger</span> <span class="k">import</span> <span class="n">screen</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.memories.episodic.episodic_experience_replay</span> <span class="k">import</span> <span class="n">EpisodicExperienceReplayParameters</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.spaces</span> <span class="k">import</span> <span class="n">DiscreteActionSpace</span><span class="p">,</span> <span class="n">BoxActionSpace</span>
|
|
<span class="kn">from</span> <span class="nn">rl_coach.utils</span> <span class="k">import</span> <span class="n">force_list</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">PPOCriticNetworkParameters</span><span class="p">(</span><span class="n">NetworkParameters</span><span class="p">):</span>
|
|
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">input_embedders_parameters</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'observation'</span><span class="p">:</span> <span class="n">InputEmbedderParameters</span><span class="p">(</span><span class="n">activation_function</span><span class="o">=</span><span class="s1">'tanh'</span><span class="p">)}</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">middleware_parameters</span> <span class="o">=</span> <span class="n">FCMiddlewareParameters</span><span class="p">(</span><span class="n">activation_function</span><span class="o">=</span><span class="s1">'tanh'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">heads_parameters</span> <span class="o">=</span> <span class="p">[</span><span class="n">VHeadParameters</span><span class="p">()]</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">async_training</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">l2_regularization</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">create_target_network</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">batch_size</span> <span class="o">=</span> <span class="mi">128</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">PPOActorNetworkParameters</span><span class="p">(</span><span class="n">NetworkParameters</span><span class="p">):</span>
|
|
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">input_embedders_parameters</span> <span class="o">=</span> <span class="p">{</span><span class="s1">'observation'</span><span class="p">:</span> <span class="n">InputEmbedderParameters</span><span class="p">(</span><span class="n">activation_function</span><span class="o">=</span><span class="s1">'tanh'</span><span class="p">)}</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">middleware_parameters</span> <span class="o">=</span> <span class="n">FCMiddlewareParameters</span><span class="p">(</span><span class="n">activation_function</span><span class="o">=</span><span class="s1">'tanh'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">heads_parameters</span> <span class="o">=</span> <span class="p">[</span><span class="n">PPOHeadParameters</span><span class="p">()]</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer_type</span> <span class="o">=</span> <span class="s1">'Adam'</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">async_training</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">l2_regularization</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">create_target_network</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">batch_size</span> <span class="o">=</span> <span class="mi">128</span>
|
|
|
|
|
|
<div class="viewcode-block" id="PPOAlgorithmParameters"><a class="viewcode-back" href="../../../components/agents/policy_optimization/ppo.html#rl_coach.agents.ppo_agent.PPOAlgorithmParameters">[docs]</a><span class="k">class</span> <span class="nc">PPOAlgorithmParameters</span><span class="p">(</span><span class="n">AlgorithmParameters</span><span class="p">):</span>
|
|
<span class="sd">"""</span>
|
|
<span class="sd"> :param policy_gradient_rescaler: (PolicyGradientRescaler)</span>
|
|
<span class="sd"> This represents how the critic will be used to update the actor. The critic value function is typically used</span>
|
|
<span class="sd"> to rescale the gradients calculated by the actor. There are several ways for doing this, such as using the</span>
|
|
<span class="sd"> advantage of the action, or the generalized advantage estimation (GAE) value.</span>
|
|
|
|
<span class="sd"> :param gae_lambda: (float)</span>
|
|
<span class="sd"> The :math:`\lambda` value is used within the GAE function in order to weight different bootstrap length</span>
|
|
<span class="sd"> estimations. Typical values are in the range 0.9-1, and define an exponential decay over the different</span>
|
|
<span class="sd"> n-step estimations.</span>
|
|
|
|
<span class="sd"> :param target_kl_divergence: (float)</span>
|
|
<span class="sd"> The target kl divergence between the current policy distribution and the new policy. PPO uses a heuristic to</span>
|
|
<span class="sd"> bring the KL divergence to this value, by adding a penalty if the kl divergence is higher.</span>
|
|
|
|
<span class="sd"> :param initial_kl_coefficient: (float)</span>
|
|
<span class="sd"> The initial weight that will be given to the KL divergence between the current and the new policy in the</span>
|
|
<span class="sd"> regularization factor.</span>
|
|
|
|
<span class="sd"> :param high_kl_penalty_coefficient: (float)</span>
|
|
<span class="sd"> The penalty that will be given for KL divergence values which are highes than what was defined as the target.</span>
|
|
|
|
<span class="sd"> :param clip_likelihood_ratio_using_epsilon: (float)</span>
|
|
<span class="sd"> If not None, the likelihood ratio between the current and new policy in the PPO loss function will be</span>
|
|
<span class="sd"> clipped to the range [1-clip_likelihood_ratio_using_epsilon, 1+clip_likelihood_ratio_using_epsilon].</span>
|
|
<span class="sd"> This is typically used in the Clipped PPO version of PPO, and should be set to None in regular PPO</span>
|
|
<span class="sd"> implementations.</span>
|
|
|
|
<span class="sd"> :param value_targets_mix_fraction: (float)</span>
|
|
<span class="sd"> The targets for the value network are an exponential weighted moving average which uses this mix fraction to</span>
|
|
<span class="sd"> define how much of the new targets will be taken into account when calculating the loss.</span>
|
|
<span class="sd"> This value should be set to the range (0,1], where 1 means that only the new targets will be taken into account.</span>
|
|
|
|
<span class="sd"> :param estimate_state_value_using_gae: (bool)</span>
|
|
<span class="sd"> If set to True, the state value will be estimated using the GAE technique.</span>
|
|
|
|
<span class="sd"> :param use_kl_regularization: (bool)</span>
|
|
<span class="sd"> If set to True, the loss function will be regularized using the KL diveregence between the current and new</span>
|
|
<span class="sd"> policy, to bound the change of the policy during the network update.</span>
|
|
|
|
<span class="sd"> :param beta_entropy: (float)</span>
|
|
<span class="sd"> An entropy regulaization term can be added to the loss function in order to control exploration. This term</span>
|
|
<span class="sd"> is weighted using the :math:`\beta` value defined by beta_entropy.</span>
|
|
|
|
<span class="sd"> """</span>
|
|
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">policy_gradient_rescaler</span> <span class="o">=</span> <span class="n">PolicyGradientRescaler</span><span class="o">.</span><span class="n">GAE</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">gae_lambda</span> <span class="o">=</span> <span class="mf">0.96</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">target_kl_divergence</span> <span class="o">=</span> <span class="mf">0.01</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">initial_kl_coefficient</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">high_kl_penalty_coefficient</span> <span class="o">=</span> <span class="mi">1000</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">clip_likelihood_ratio_using_epsilon</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">value_targets_mix_fraction</span> <span class="o">=</span> <span class="mf">0.1</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">estimate_state_value_using_gae</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">use_kl_regularization</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">beta_entropy</span> <span class="o">=</span> <span class="mf">0.01</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_consecutive_playing_steps</span> <span class="o">=</span> <span class="n">EnvironmentSteps</span><span class="p">(</span><span class="mi">5000</span><span class="p">)</span></div>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">PPOAgentParameters</span><span class="p">(</span><span class="n">AgentParameters</span><span class="p">):</span>
|
|
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">(</span><span class="n">algorithm</span><span class="o">=</span><span class="n">PPOAlgorithmParameters</span><span class="p">(),</span>
|
|
<span class="n">exploration</span><span class="o">=</span><span class="p">{</span><span class="n">DiscreteActionSpace</span><span class="p">:</span> <span class="n">CategoricalParameters</span><span class="p">(),</span>
|
|
<span class="n">BoxActionSpace</span><span class="p">:</span> <span class="n">AdditiveNoiseParameters</span><span class="p">()},</span>
|
|
<span class="n">memory</span><span class="o">=</span><span class="n">EpisodicExperienceReplayParameters</span><span class="p">(),</span>
|
|
<span class="n">networks</span><span class="o">=</span><span class="p">{</span><span class="s2">"critic"</span><span class="p">:</span> <span class="n">PPOCriticNetworkParameters</span><span class="p">(),</span> <span class="s2">"actor"</span><span class="p">:</span> <span class="n">PPOActorNetworkParameters</span><span class="p">()})</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">path</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="s1">'rl_coach.agents.ppo_agent:PPOAgent'</span>
|
|
|
|
|
|
<span class="c1"># Proximal Policy Optimization - https://arxiv.org/pdf/1707.06347.pdf</span>
|
|
<span class="k">class</span> <span class="nc">PPOAgent</span><span class="p">(</span><span class="n">ActorCriticAgent</span><span class="p">):</span>
|
|
<span class="k">def</span> <span class="nf">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">agent_parameters</span><span class="p">,</span> <span class="n">parent</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span><span class="s1">'LevelManager'</span><span class="p">,</span> <span class="s1">'CompositeAgent'</span><span class="p">]</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">(</span><span class="n">agent_parameters</span><span class="p">,</span> <span class="n">parent</span><span class="p">)</span>
|
|
|
|
<span class="c1"># signals definition</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">value_loss</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">register_signal</span><span class="p">(</span><span class="s1">'Value Loss'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">policy_loss</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">register_signal</span><span class="p">(</span><span class="s1">'Policy Loss'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kl_divergence</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">register_signal</span><span class="p">(</span><span class="s1">'KL Divergence'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">total_kl_divergence_during_training_process</span> <span class="o">=</span> <span class="mf">0.0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">unclipped_grads</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">register_signal</span><span class="p">(</span><span class="s1">'Grads (unclipped)'</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">fill_advantages</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">batch</span><span class="p">):</span>
|
|
<span class="n">batch</span> <span class="o">=</span> <span class="n">Batch</span><span class="p">(</span><span class="n">batch</span><span class="p">)</span>
|
|
<span class="n">network_keys</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">input_embedders_parameters</span><span class="o">.</span><span class="n">keys</span><span class="p">()</span>
|
|
|
|
<span class="c1"># * Found not to have any impact *</span>
|
|
<span class="c1"># current_states_with_timestep = self.concat_state_and_timestep(batch)</span>
|
|
|
|
<span class="n">current_state_values</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">states</span><span class="p">(</span><span class="n">network_keys</span><span class="p">))</span><span class="o">.</span><span class="n">squeeze</span><span class="p">()</span>
|
|
<span class="n">total_returns</span> <span class="o">=</span> <span class="n">batch</span><span class="o">.</span><span class="n">n_step_discounted_rewards</span><span class="p">()</span>
|
|
<span class="c1"># calculate advantages</span>
|
|
<span class="n">advantages</span> <span class="o">=</span> <span class="p">[]</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">policy_gradient_rescaler</span> <span class="o">==</span> <span class="n">PolicyGradientRescaler</span><span class="o">.</span><span class="n">A_VALUE</span><span class="p">:</span>
|
|
<span class="n">advantages</span> <span class="o">=</span> <span class="n">total_returns</span> <span class="o">-</span> <span class="n">current_state_values</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">policy_gradient_rescaler</span> <span class="o">==</span> <span class="n">PolicyGradientRescaler</span><span class="o">.</span><span class="n">GAE</span><span class="p">:</span>
|
|
<span class="c1"># get bootstraps</span>
|
|
<span class="n">episode_start_idx</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="n">advantages</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array</span><span class="p">([])</span>
|
|
<span class="c1"># current_state_values[batch.game_overs()] = 0</span>
|
|
<span class="k">for</span> <span class="n">idx</span><span class="p">,</span> <span class="n">game_over</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">game_overs</span><span class="p">()):</span>
|
|
<span class="k">if</span> <span class="n">game_over</span><span class="p">:</span>
|
|
<span class="c1"># get advantages for the rollout</span>
|
|
<span class="n">value_bootstrapping</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">((</span><span class="mi">1</span><span class="p">,))</span>
|
|
<span class="n">rollout_state_values</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">current_state_values</span><span class="p">[</span><span class="n">episode_start_idx</span><span class="p">:</span><span class="n">idx</span><span class="o">+</span><span class="mi">1</span><span class="p">],</span> <span class="n">value_bootstrapping</span><span class="p">)</span>
|
|
|
|
<span class="n">rollout_advantages</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> \
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">get_general_advantage_estimation_values</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">rewards</span><span class="p">()[</span><span class="n">episode_start_idx</span><span class="p">:</span><span class="n">idx</span><span class="o">+</span><span class="mi">1</span><span class="p">],</span>
|
|
<span class="n">rollout_state_values</span><span class="p">)</span>
|
|
<span class="n">episode_start_idx</span> <span class="o">=</span> <span class="n">idx</span> <span class="o">+</span> <span class="mi">1</span>
|
|
<span class="n">advantages</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">advantages</span><span class="p">,</span> <span class="n">rollout_advantages</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">screen</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span><span class="s2">"WARNING: The requested policy gradient rescaler is not available"</span><span class="p">)</span>
|
|
|
|
<span class="c1"># standardize</span>
|
|
<span class="n">advantages</span> <span class="o">=</span> <span class="p">(</span><span class="n">advantages</span> <span class="o">-</span> <span class="n">np</span><span class="o">.</span><span class="n">mean</span><span class="p">(</span><span class="n">advantages</span><span class="p">))</span> <span class="o">/</span> <span class="n">np</span><span class="o">.</span><span class="n">std</span><span class="p">(</span><span class="n">advantages</span><span class="p">)</span>
|
|
|
|
<span class="c1"># TODO: this will be problematic with a shared memory</span>
|
|
<span class="k">for</span> <span class="n">transition</span><span class="p">,</span> <span class="n">advantage</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">memory</span><span class="o">.</span><span class="n">transitions</span><span class="p">,</span> <span class="n">advantages</span><span class="p">):</span>
|
|
<span class="n">transition</span><span class="o">.</span><span class="n">info</span><span class="p">[</span><span class="s1">'advantage'</span><span class="p">]</span> <span class="o">=</span> <span class="n">advantage</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">action_advantages</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">advantages</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">train_value_network</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">dataset</span><span class="p">,</span> <span class="n">epochs</span><span class="p">):</span>
|
|
<span class="n">loss</span> <span class="o">=</span> <span class="p">[]</span>
|
|
<span class="n">batch</span> <span class="o">=</span> <span class="n">Batch</span><span class="p">(</span><span class="n">dataset</span><span class="p">)</span>
|
|
<span class="n">network_keys</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">input_embedders_parameters</span><span class="o">.</span><span class="n">keys</span><span class="p">()</span>
|
|
|
|
<span class="c1"># * Found not to have any impact *</span>
|
|
<span class="c1"># add a timestep to the observation</span>
|
|
<span class="c1"># current_states_with_timestep = self.concat_state_and_timestep(dataset)</span>
|
|
|
|
<span class="n">mix_fraction</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">algorithm</span><span class="o">.</span><span class="n">value_targets_mix_fraction</span>
|
|
<span class="n">total_returns</span> <span class="o">=</span> <span class="n">batch</span><span class="o">.</span><span class="n">n_step_discounted_rewards</span><span class="p">(</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">j</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">epochs</span><span class="p">):</span>
|
|
<span class="n">curr_batch_size</span> <span class="o">=</span> <span class="n">batch</span><span class="o">.</span><span class="n">size</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">optimizer_type</span> <span class="o">!=</span> <span class="s1">'LBFGS'</span><span class="p">:</span>
|
|
<span class="n">curr_batch_size</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">batch_size</span>
|
|
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">size</span> <span class="o">//</span> <span class="n">curr_batch_size</span><span class="p">):</span>
|
|
<span class="c1"># split to batches for first order optimization techniques</span>
|
|
<span class="n">current_states_batch</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">k</span><span class="p">:</span> <span class="n">v</span><span class="p">[</span><span class="n">i</span> <span class="o">*</span> <span class="n">curr_batch_size</span><span class="p">:(</span><span class="n">i</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span> <span class="o">*</span> <span class="n">curr_batch_size</span><span class="p">]</span>
|
|
<span class="k">for</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span> <span class="ow">in</span> <span class="n">batch</span><span class="o">.</span><span class="n">states</span><span class="p">(</span><span class="n">network_keys</span><span class="p">)</span><span class="o">.</span><span class="n">items</span><span class="p">()</span>
|
|
<span class="p">}</span>
|
|
<span class="n">total_return_batch</span> <span class="o">=</span> <span class="n">total_returns</span><span class="p">[</span><span class="n">i</span> <span class="o">*</span> <span class="n">curr_batch_size</span><span class="p">:(</span><span class="n">i</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span> <span class="o">*</span> <span class="n">curr_batch_size</span><span class="p">]</span>
|
|
<span class="n">old_policy_values</span> <span class="o">=</span> <span class="n">force_list</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">target_network</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span>
|
|
<span class="n">current_states_batch</span><span class="p">)</span><span class="o">.</span><span class="n">squeeze</span><span class="p">())</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">optimizer_type</span> <span class="o">!=</span> <span class="s1">'LBFGS'</span><span class="p">:</span>
|
|
<span class="n">targets</span> <span class="o">=</span> <span class="n">total_return_batch</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">current_values</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">current_states_batch</span><span class="p">)</span>
|
|
<span class="n">targets</span> <span class="o">=</span> <span class="n">current_values</span> <span class="o">*</span> <span class="p">(</span><span class="mi">1</span> <span class="o">-</span> <span class="n">mix_fraction</span><span class="p">)</span> <span class="o">+</span> <span class="n">total_return_batch</span> <span class="o">*</span> <span class="n">mix_fraction</span>
|
|
|
|
<span class="n">inputs</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">copy</span><span class="p">(</span><span class="n">current_states_batch</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">input_index</span><span class="p">,</span> <span class="nb">input</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">old_policy_values</span><span class="p">):</span>
|
|
<span class="n">name</span> <span class="o">=</span> <span class="s1">'output_0_</span><span class="si">{}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">input_index</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">name</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">inputs</span><span class="p">:</span>
|
|
<span class="n">inputs</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="nb">input</span>
|
|
|
|
<span class="n">value_loss</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">accumulate_gradients</span><span class="p">(</span><span class="n">inputs</span><span class="p">,</span> <span class="n">targets</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">apply_gradients_to_online_network</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">task_parameters</span><span class="p">,</span> <span class="n">DistributedTaskParameters</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">apply_gradients_to_global_network</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">reset_accumulated_gradients</span><span class="p">()</span>
|
|
|
|
<span class="n">loss</span><span class="o">.</span><span class="n">append</span><span class="p">([</span><span class="n">value_loss</span><span class="p">[</span><span class="mi">0</span><span class="p">]])</span>
|
|
<span class="n">loss</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">mean</span><span class="p">(</span><span class="n">loss</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">loss</span>
|
|
|
|
<span class="k">def</span> <span class="nf">concat_state_and_timestep</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">dataset</span><span class="p">):</span>
|
|
<span class="n">current_states_with_timestep</span> <span class="o">=</span> <span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">transition</span><span class="o">.</span><span class="n">state</span><span class="p">[</span><span class="s1">'observation'</span><span class="p">],</span> <span class="n">transition</span><span class="o">.</span><span class="n">info</span><span class="p">[</span><span class="s1">'timestep'</span><span class="p">])</span>
|
|
<span class="k">for</span> <span class="n">transition</span> <span class="ow">in</span> <span class="n">dataset</span><span class="p">]</span>
|
|
<span class="n">current_states_with_timestep</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">expand_dims</span><span class="p">(</span><span class="n">current_states_with_timestep</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">current_states_with_timestep</span>
|
|
|
|
<span class="k">def</span> <span class="nf">train_policy_network</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">dataset</span><span class="p">,</span> <span class="n">epochs</span><span class="p">):</span>
|
|
<span class="n">loss</span> <span class="o">=</span> <span class="p">[]</span>
|
|
<span class="k">for</span> <span class="n">j</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">epochs</span><span class="p">):</span>
|
|
<span class="n">loss</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="s1">'total_loss'</span><span class="p">:</span> <span class="p">[],</span>
|
|
<span class="s1">'policy_losses'</span><span class="p">:</span> <span class="p">[],</span>
|
|
<span class="s1">'unclipped_grads'</span><span class="p">:</span> <span class="p">[],</span>
|
|
<span class="s1">'fetch_result'</span><span class="p">:</span> <span class="p">[]</span>
|
|
<span class="p">}</span>
|
|
<span class="c1">#shuffle(dataset)</span>
|
|
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">dataset</span><span class="p">)</span> <span class="o">//</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">batch_size</span><span class="p">):</span>
|
|
<span class="n">batch</span> <span class="o">=</span> <span class="n">Batch</span><span class="p">(</span><span class="n">dataset</span><span class="p">[</span><span class="n">i</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">batch_size</span><span class="p">:</span>
|
|
<span class="p">(</span><span class="n">i</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">batch_size</span><span class="p">])</span>
|
|
|
|
<span class="n">network_keys</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">input_embedders_parameters</span><span class="o">.</span><span class="n">keys</span><span class="p">()</span>
|
|
|
|
<span class="n">advantages</span> <span class="o">=</span> <span class="n">batch</span><span class="o">.</span><span class="n">info</span><span class="p">(</span><span class="s1">'advantage'</span><span class="p">)</span>
|
|
<span class="n">actions</span> <span class="o">=</span> <span class="n">batch</span><span class="o">.</span><span class="n">actions</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">spaces</span><span class="o">.</span><span class="n">action</span><span class="p">,</span> <span class="n">DiscreteActionSpace</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">len</span><span class="p">(</span><span class="n">actions</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="n">actions</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">expand_dims</span><span class="p">(</span><span class="n">actions</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="c1"># get old policy probabilities and distribution</span>
|
|
<span class="n">old_policy</span> <span class="o">=</span> <span class="n">force_list</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">target_network</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">states</span><span class="p">(</span><span class="n">network_keys</span><span class="p">)))</span>
|
|
|
|
<span class="c1"># calculate gradients and apply on both the local policy network and on the global policy network</span>
|
|
<span class="n">fetches</span> <span class="o">=</span> <span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">output_heads</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">kl_divergence</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">output_heads</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">entropy</span><span class="p">]</span>
|
|
|
|
<span class="n">inputs</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">copy</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">states</span><span class="p">(</span><span class="n">network_keys</span><span class="p">))</span>
|
|
<span class="n">inputs</span><span class="p">[</span><span class="s1">'output_0_0'</span><span class="p">]</span> <span class="o">=</span> <span class="n">actions</span>
|
|
|
|
<span class="c1"># old_policy_distribution needs to be represented as a list, because in the event of discrete controls,</span>
|
|
<span class="c1"># it has just a mean. otherwise, it has both a mean and standard deviation</span>
|
|
<span class="k">for</span> <span class="n">input_index</span><span class="p">,</span> <span class="nb">input</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">old_policy</span><span class="p">):</span>
|
|
<span class="n">inputs</span><span class="p">[</span><span class="s1">'output_0_</span><span class="si">{}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">input_index</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)]</span> <span class="o">=</span> <span class="nb">input</span>
|
|
|
|
<span class="n">total_loss</span><span class="p">,</span> <span class="n">policy_losses</span><span class="p">,</span> <span class="n">unclipped_grads</span><span class="p">,</span> <span class="n">fetch_result</span> <span class="o">=</span>\
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">accumulate_gradients</span><span class="p">(</span>
|
|
<span class="n">inputs</span><span class="p">,</span> <span class="p">[</span><span class="n">advantages</span><span class="p">],</span> <span class="n">additional_fetches</span><span class="o">=</span><span class="n">fetches</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">apply_gradients_to_online_network</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">task_parameters</span><span class="p">,</span> <span class="n">DistributedTaskParameters</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">apply_gradients_to_global_network</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">reset_accumulated_gradients</span><span class="p">()</span>
|
|
|
|
<span class="n">loss</span><span class="p">[</span><span class="s1">'total_loss'</span><span class="p">]</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">total_loss</span><span class="p">)</span>
|
|
<span class="n">loss</span><span class="p">[</span><span class="s1">'policy_losses'</span><span class="p">]</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">policy_losses</span><span class="p">)</span>
|
|
<span class="n">loss</span><span class="p">[</span><span class="s1">'unclipped_grads'</span><span class="p">]</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">unclipped_grads</span><span class="p">)</span>
|
|
<span class="n">loss</span><span class="p">[</span><span class="s1">'fetch_result'</span><span class="p">]</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">fetch_result</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">unclipped_grads</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">unclipped_grads</span><span class="p">)</span>
|
|
|
|
<span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">loss</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
|
|
<span class="n">loss</span><span class="p">[</span><span class="n">key</span><span class="p">]</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">mean</span><span class="p">(</span><span class="n">loss</span><span class="p">[</span><span class="n">key</span><span class="p">],</span> <span class="mi">0</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">learning_rate_decay_rate</span> <span class="o">!=</span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="n">curr_learning_rate</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">get_variable_value</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">learning_rate</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">curr_learning_rate</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">curr_learning_rate</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">curr_learning_rate</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">network_wrappers</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">learning_rate</span>
|
|
|
|
<span class="c1"># log training parameters</span>
|
|
<span class="n">screen</span><span class="o">.</span><span class="n">log_dict</span><span class="p">(</span>
|
|
<span class="n">OrderedDict</span><span class="p">([</span>
|
|
<span class="p">(</span><span class="s2">"Surrogate loss"</span><span class="p">,</span> <span class="n">loss</span><span class="p">[</span><span class="s1">'policy_losses'</span><span class="p">][</span><span class="mi">0</span><span class="p">]),</span>
|
|
<span class="p">(</span><span class="s2">"KL divergence"</span><span class="p">,</span> <span class="n">loss</span><span class="p">[</span><span class="s1">'fetch_result'</span><span class="p">][</span><span class="mi">0</span><span class="p">]),</span>
|
|
<span class="p">(</span><span class="s2">"Entropy"</span><span class="p">,</span> <span class="n">loss</span><span class="p">[</span><span class="s1">'fetch_result'</span><span class="p">][</span><span class="mi">1</span><span class="p">]),</span>
|
|
<span class="p">(</span><span class="s2">"training epoch"</span><span class="p">,</span> <span class="n">j</span><span class="p">),</span>
|
|
<span class="p">(</span><span class="s2">"learning_rate"</span><span class="p">,</span> <span class="n">curr_learning_rate</span><span class="p">)</span>
|
|
<span class="p">]),</span>
|
|
<span class="n">prefix</span><span class="o">=</span><span class="s2">"Policy training"</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">total_kl_divergence_during_training_process</span> <span class="o">=</span> <span class="n">loss</span><span class="p">[</span><span class="s1">'fetch_result'</span><span class="p">][</span><span class="mi">0</span><span class="p">]</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">entropy</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">loss</span><span class="p">[</span><span class="s1">'fetch_result'</span><span class="p">][</span><span class="mi">1</span><span class="p">])</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kl_divergence</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">loss</span><span class="p">[</span><span class="s1">'fetch_result'</span><span class="p">][</span><span class="mi">0</span><span class="p">])</span>
|
|
<span class="k">return</span> <span class="n">loss</span><span class="p">[</span><span class="s1">'total_loss'</span><span class="p">]</span>
|
|
|
|
<span class="k">def</span> <span class="nf">update_kl_coefficient</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># John Schulman takes the mean kl divergence only over the last epoch which is strange but we will follow</span>
|
|
<span class="c1"># his implementation for now because we know it works well</span>
|
|
<span class="n">screen</span><span class="o">.</span><span class="n">log_title</span><span class="p">(</span><span class="s2">"KL = </span><span class="si">{}</span><span class="s2">"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">total_kl_divergence_during_training_process</span><span class="p">))</span>
|
|
|
|
<span class="c1"># update kl coefficient</span>
|
|
<span class="n">kl_target</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">algorithm</span><span class="o">.</span><span class="n">target_kl_divergence</span>
|
|
<span class="n">kl_coefficient</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">get_variable_value</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">output_heads</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">kl_coefficient</span><span class="p">)</span>
|
|
<span class="n">new_kl_coefficient</span> <span class="o">=</span> <span class="n">kl_coefficient</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">total_kl_divergence_during_training_process</span> <span class="o">></span> <span class="mf">1.3</span> <span class="o">*</span> <span class="n">kl_target</span><span class="p">:</span>
|
|
<span class="c1"># kl too high => increase regularization</span>
|
|
<span class="n">new_kl_coefficient</span> <span class="o">*=</span> <span class="mf">1.5</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">total_kl_divergence_during_training_process</span> <span class="o"><</span> <span class="mf">0.7</span> <span class="o">*</span> <span class="n">kl_target</span><span class="p">:</span>
|
|
<span class="c1"># kl too low => decrease regularization</span>
|
|
<span class="n">new_kl_coefficient</span> <span class="o">/=</span> <span class="mf">1.5</span>
|
|
|
|
<span class="c1"># update the kl coefficient variable</span>
|
|
<span class="k">if</span> <span class="n">kl_coefficient</span> <span class="o">!=</span> <span class="n">new_kl_coefficient</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">set_variable_value</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">output_heads</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">assign_kl_coefficient</span><span class="p">,</span>
|
|
<span class="n">new_kl_coefficient</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">output_heads</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">kl_coefficient_ph</span><span class="p">)</span>
|
|
|
|
<span class="n">screen</span><span class="o">.</span><span class="n">log_title</span><span class="p">(</span><span class="s2">"KL penalty coefficient change = </span><span class="si">{}</span><span class="s2"> -> </span><span class="si">{}</span><span class="s2">"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">kl_coefficient</span><span class="p">,</span> <span class="n">new_kl_coefficient</span><span class="p">))</span>
|
|
|
|
<span class="k">def</span> <span class="nf">post_training_commands</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">algorithm</span><span class="o">.</span><span class="n">use_kl_regularization</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">update_kl_coefficient</span><span class="p">()</span>
|
|
|
|
<span class="c1"># clean memory</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">call_memory</span><span class="p">(</span><span class="s1">'clean'</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_should_train_helper</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">wait_for_full_episode</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="n">_should_train_helper</span><span class="p">(</span><span class="kc">True</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">train</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">loss</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">_should_train</span><span class="p">(</span><span class="n">wait_for_full_episode</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
|
|
<span class="k">for</span> <span class="n">network</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="o">.</span><span class="n">values</span><span class="p">():</span>
|
|
<span class="n">network</span><span class="o">.</span><span class="n">set_is_training</span><span class="p">(</span><span class="kc">True</span><span class="p">)</span>
|
|
|
|
<span class="k">for</span> <span class="n">training_step</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">algorithm</span><span class="o">.</span><span class="n">num_consecutive_training_steps</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">sync</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'critic'</span><span class="p">]</span><span class="o">.</span><span class="n">sync</span><span class="p">()</span>
|
|
|
|
<span class="n">dataset</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">memory</span><span class="o">.</span><span class="n">transitions</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">fill_advantages</span><span class="p">(</span><span class="n">dataset</span><span class="p">)</span>
|
|
|
|
<span class="c1"># take only the requested number of steps</span>
|
|
<span class="n">dataset</span> <span class="o">=</span> <span class="n">dataset</span><span class="p">[:</span><span class="bp">self</span><span class="o">.</span><span class="n">ap</span><span class="o">.</span><span class="n">algorithm</span><span class="o">.</span><span class="n">num_consecutive_playing_steps</span><span class="o">.</span><span class="n">num_steps</span><span class="p">]</span>
|
|
|
|
<span class="n">value_loss</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">train_value_network</span><span class="p">(</span><span class="n">dataset</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">policy_loss</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">train_policy_network</span><span class="p">(</span><span class="n">dataset</span><span class="p">,</span> <span class="mi">10</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">value_loss</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">value_loss</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">policy_loss</span><span class="o">.</span><span class="n">add_sample</span><span class="p">(</span><span class="n">policy_loss</span><span class="p">)</span>
|
|
|
|
<span class="k">for</span> <span class="n">network</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="o">.</span><span class="n">values</span><span class="p">():</span>
|
|
<span class="n">network</span><span class="o">.</span><span class="n">set_is_training</span><span class="p">(</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">post_training_commands</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">training_iteration</span> <span class="o">+=</span> <span class="mi">1</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">update_log</span><span class="p">()</span> <span class="c1"># should be done in order to update the data that has been accumulated * while not playing *</span>
|
|
<span class="k">return</span> <span class="n">np</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">value_loss</span><span class="p">,</span> <span class="n">policy_loss</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">get_prediction</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">states</span><span class="p">):</span>
|
|
<span class="n">tf_input_state</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">prepare_batch_for_inference</span><span class="p">(</span><span class="n">states</span><span class="p">,</span> <span class="s2">"actor"</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">networks</span><span class="p">[</span><span class="s1">'actor'</span><span class="p">]</span><span class="o">.</span><span class="n">online_network</span><span class="o">.</span><span class="n">predict</span><span class="p">(</span><span class="n">tf_input_state</span><span class="p">)</span>
|
|
</pre></div>
|
|
|
|
</div>
|
|
|
|
</div>
|
|
<footer>
|
|
|
|
|
|
<hr/>
|
|
|
|
<div role="contentinfo">
|
|
<p>
|
|
© Copyright 2018, Intel AI Lab
|
|
|
|
</p>
|
|
</div>
|
|
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
|
|
|
|
</footer>
|
|
|
|
</div>
|
|
</div>
|
|
|
|
</section>
|
|
|
|
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<script type="text/javascript" id="documentation_options" data-url_root="../../../" src="../../../_static/documentation_options.js"></script>
|
|
<script type="text/javascript" src="../../../_static/jquery.js"></script>
|
|
<script type="text/javascript" src="../../../_static/underscore.js"></script>
|
|
<script type="text/javascript" src="../../../_static/doctools.js"></script>
|
|
<script async="async" type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
|
|
|
|
|
|
|
|
|
|
<script type="text/javascript" src="../../../_static/js/theme.js"></script>
|
|
|
|
<script type="text/javascript">
|
|
jQuery(function () {
|
|
SphinxRtdTheme.Navigation.enable(true);
|
|
});
|
|
</script>
|
|
|
|
</body>
|
|
</html> |