Several questions regarding my implementation of PPO on Pytorch

Hi! First time posting here!
I’ve been learning RL this summer and this week I’ve tried to make a PPO implementation on Pytorch with the help of some repositories from github with similiar algorithms.
The code runs OpenAI’s Lunar Lander but I have several errors that I have not been able to fix, the biggest one being that the algorithm quickly converges to doing the same action regardless of the state. The other major problem I’ve found is that even though I’m using backwards() only once, I get an error asking me to set retain_graph to True.
Because of that, I see no improvement of the rewards obtained over 1000 steps, I don’t know if the algorithm needs more steps to be able to see an improvement.

I’m really sorry if this kind of problems have no place in this forums, I just didn’t know where to post this.
Also I’m sorry for the messy code, it’s my first time doing this kind of algorithms, and I’m fairly new with pytorch and machine learning in general.

Tanks a lot in advance!!


import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
from torch.distributions import Categorical
import gym

class actorCritic(nn.Module):
    def __init__(self):
        super(actorCritic, self).__init__()
        self.fc = nn.Sequential(
        nn.Linear(8, 16),
        nn.Linear(16, 32), 
        nn.Linear(32, 64), 
        self.pi = nn.Linear(64, 4)
        self.value = nn.Linear(64, 1)
    def forward(self, x):
        x = self.fc(x)
        pi_1 = self.pi(x)
        pi_out = F.softmax(pi_1, dim=-1)
        value_out = self.value(x)
        return pi_out, value_out

def GAE(rewards, values, masks):
    gamma = 0.99
    lamb = 0.95
    advan_t = 0
    sizes = rewards.size()

    advantages = torch.zeros(1, sizes[1])
    for t in reversed(range(sizes[1])):
        delta = rewards[0][t] + gamma*values[0][t+1]*masks[0][t] - values[0][t]
        advan_t = delta + gamma*lamb*advan_t*mask[0][t]
        advantages[0][t] = advan_t
    real_values = values[:,:sizes[1]] + advantages

    return advantages, real_values
def plot_rewards(rewards):
    plt.savefig('TruePPO 500 steps.png')    
def interact(times, states):

    rewards = torch.zeros(1, times)

    actions = torch.zeros(1, times)
    mask = torch.ones(1, times)
    for steps in range(times):
        action_probs, _ = network(states[steps])
        m = Categorical(action_probs)
        action = int(m.sample())

        obs, reward, done, _ = env.step(action)

        if done:
            obs = env.reset()
            mask[0][steps] = 0

        states[steps+1] = torch.from_numpy(obs).float()
        rewards[0][steps] = reward
        actions[0][steps] = action
    return states, rewards, actions, mask


total_steps = 1000
batch_size = 10
env = gym.make('LunarLander-v2')
network = actorCritic()
old_network = actorCritic()
optimizer = torch.optim.Adam(network.parameters(), lr = 0.001)
states = torch.zeros(batch_size+1, 8)
steps = 0
obs_ = env.reset()
obs = torch.from_numpy(obs_).float()
states[0] = obs
reward_means = []
nn_paramD = network.state_dict()

while steps < total_steps:
    print (steps)
    states, rewards, actions, mask = interact(batch_size, states)
    #calculate values, GAE, normalize advantages, randomize, calculate loss, backprop,
    _, values = network(states)
    values = values.view(-1, batch_size+1)

    advantages, v_targ = GAE(rewards, values, mask)

    advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-5)

    for n in range(rewards.size()[1]):

        probabilities, _ = network(states[n])
        print (probabilities)
        m = Categorical(probabilities)
        action_prob = m.probs[int(actions[0][n])]

        entropia = m.entropy()
        old_probabilities, _ = old_network(states[n])
        m_old = Categorical(old_probabilities)
        old_action_prob = m.probs[actions[0][n].int()]
        ratio = action_prob / old_action_prob
        surr1 = ratio*advantages[0][n]
        surr2 = torch.clamp(ratio, min = (1.-0.2), max = (1.+0.2))
        policy_loss = -torch.min(surr1, surr2)

        value_loss = 0.5*(values[0][n]-v_targ[0][n])**2

        entropy_loss = -entropia

        total_loss = policy_loss + value_loss + 0.01*entropy_loss

        total_loss.backward(retain_graph = True)


    nn_paramD = network.state_dict()
    steps += 1

There is a lot of detail there, or in any RL implementation for that matter. Taking a quick look, there are a number of quirks and ? in your impl. Without digging deep into slicing and off by one level detail, a few thoughts:

  1. It’s a great learning experience to start from scratch, but there is a lot to get right, and just a small mistake to get wrong and nothing will work. I’d spend a little more time looking at other (working) reference impl to guide you.

  2. You’re calling zero grad once for your whole inner loop. That is likely why you get the retain_graph warning. Doing one zero and multiple backward, and one step is a way of accumulating your gradients before an update. What you’re doing is all kinds of crazy :slight_smile: You should generally do the following within the loop:

# most RL algorithms also do a nn.utils.clip_grad_norm_(model.parameters(), self.max_grad_norm)
  1. For PPO specifically, your notion of ‘batches’ and samples in general is not typical. Usually you build up more trajectories from your simulator and then sample larger batches from them multiple times (PPO epochs).

  2. From outer loop to outer loop the states (observations) aren’t being rolled over. You leave next state in states[steps+1], but you don’t bring that last observation into states[0] for the start of the next. Unless I’m missing something you’re losing all continuity in your observations from one outer loop to another, given that the environment episodes don’t necessarily line up with your rollout sampling.

1 Like

Thank you very much for the reply Ross! You helped me a ton and I’ll definitely look into some other implementations first. Thanks!!