procgen | Procgen Benchmark : Procedurally-Generated Game | Reinforcement Learning library
kandi X-RAY | procgen Summary
Support
Quality
Security
License
Reuse
Currently covering the most popular Java, JavaScript and Python libraries. See a Sample Here
procgen Key Features
procgen Examples and Code Snippets
Trending Discussions on Reinforcement Learning
Trending Discussions on Reinforcement Learning
QUESTION
I want to compile my DQN Agent but I get error: AttributeError: 'Adam' object has no attribute '_name',
DQN = buildAgent(model, actions)
DQN.compile(Adam(lr=1e-3), metrics=['mae'])
I tried adding fake _name but it doesn't work, I'm following a tutorial and it works on tutor's machine, it's probably some new update change but how to fix this
Here is my full code:
from keras.layers import Dense, Flatten
import gym
from keras.optimizer_v1 import Adam
from rl.agents.dqn import DQNAgent
from rl.policy import BoltzmannQPolicy
from rl.memory import SequentialMemory
env = gym.make('CartPole-v0')
states = env.observation_space.shape[0]
actions = env.action_space.n
episodes = 10
def buildModel(statez, actiones):
model = Sequential()
model.add(Flatten(input_shape=(1, statez)))
model.add(Dense(24, activation='relu'))
model.add(Dense(24, activation='relu'))
model.add(Dense(actiones, activation='linear'))
return model
model = buildModel(states, actions)
def buildAgent(modell, actionz):
policy = BoltzmannQPolicy()
memory = SequentialMemory(limit=50000, window_length=1)
dqn = DQNAgent(model=modell, memory=memory, policy=policy, nb_actions=actionz, nb_steps_warmup=10, target_model_update=1e-2)
return dqn
DQN = buildAgent(model, actions)
DQN.compile(Adam(lr=1e-3), metrics=['mae'])
DQN.fit(env, nb_steps=50000, visualize=False, verbose=1)
ANSWER
Answered 2022-Apr-16 at 15:05Your error came from importing Adam with from keras.optimizer_v1 import Adam, You can solve your problem with tf.keras.optimizers.Adam from TensorFlow >= v2 like below:
(The lr argument is deprecated, it's better to use learning_rate instead.)
# !pip install keras-rl2
import tensorflow as tf
from keras.layers import Dense, Flatten
import gym
from rl.agents.dqn import DQNAgent
from rl.policy import BoltzmannQPolicy
from rl.memory import SequentialMemory
env = gym.make('CartPole-v0')
states = env.observation_space.shape[0]
actions = env.action_space.n
episodes = 10
def buildModel(statez, actiones):
model = tf.keras.Sequential()
model.add(Flatten(input_shape=(1, statez)))
model.add(Dense(24, activation='relu'))
model.add(Dense(24, activation='relu'))
model.add(Dense(actiones, activation='linear'))
return model
def buildAgent(modell, actionz):
policy = BoltzmannQPolicy()
memory = SequentialMemory(limit=50000, window_length=1)
dqn = DQNAgent(model=modell, memory=memory, policy=policy,
nb_actions=actionz, nb_steps_warmup=10,
target_model_update=1e-2)
return dqn
model = buildModel(states, actions)
DQN = buildAgent(model, actions)
DQN.compile(tf.keras.optimizers.Adam(learning_rate=1e-3), metrics=['mae'])
DQN.fit(env, nb_steps=50000, visualize=False, verbose=1)
QUESTION
I'm having a hard time wrapping my head around what and when vectorized environments should be used. If you can provide an example of a use case, that would be great.
Documentation of vectorized environments in SB3: https://stable-baselines3.readthedocs.io/en/master/guide/vec_envs.html
ANSWER
Answered 2022-Mar-25 at 10:37Vectorized Environments are a method for stacking multiple independent environments into a single environment. Instead of executing and training an agent on 1 environment per step, it allows to train the agent on multiple environments per step.
Usually you also want these environment to have different seeds, in order to gain more diverse experience. This is very useful to speed up training.
I think they are called "vectorized" since each training step the agent observes multiple states (inserted in a vector), outputs multiple actions (one for each environment), which are inserted in a vector, and receives multiple rewards. Hence the "vectorized" term
QUESTION
I'm learning about policy gradients and I'm having hard time understanding how does the gradient passes through a random operation. From here: It is not possible to directly backpropagate through random samples. However, there are two main methods for creating surrogate functions that can be backpropagated through.
They have an example of the score function:
probs = policy_network(state)
# Note that this is equivalent to what used to be called multinomial
m = Categorical(probs)
action = m.sample()
next_state, reward = env.step(action)
loss = -m.log_prob(action) * reward
loss.backward()
Which I tried to create an example of:
import torch
import torch.nn as nn
import torch.optim as optim
from torch.distributions import Normal
import matplotlib.pyplot as plt
from tqdm import tqdm
softplus = torch.nn.Softplus()
class Model_RL(nn.Module):
def __init__(self):
super(Model_RL, self).__init__()
self.fc1 = nn.Linear(1, 20)
self.fc2 = nn.Linear(20, 30)
self.fc3 = nn.Linear(30, 2)
def forward(self, x):
x1 = self.fc1(x)
x = torch.relu(x1)
x2 = self.fc2(x)
x = torch.relu(x2)
x3 = softplus(self.fc3(x))
return x3, x2, x1
# basic
net_RL = Model_RL()
features = torch.tensor([1.0])
x = torch.tensor([1.0])
y = torch.tensor(3.0)
baseline = 0
baseline_lr = 0.1
epochs = 3
opt_RL = optim.Adam(net_RL.parameters(), lr=1e-3)
losses = []
xs = []
for _ in tqdm(range(epochs)):
out_RL = net_RL(x)
mu, std = out_RL[0]
dist = Normal(mu, std)
print(dist)
a = dist.sample()
log_p = dist.log_prob(a)
out = features * a
reward = -torch.square((y - out))
baseline = (1-baseline_lr)*baseline + baseline_lr*reward
loss = -(reward-baseline)*log_p
opt_RL.zero_grad()
loss.backward()
opt_RL.step()
losses.append(loss.item())
This seems to work magically fine which again, I don't understand how the gradient passes through as they mentioned that it can't pass through the random operation (but then somehow it does).
Now since the gradient can't flow through the random operation I tried to replace mu, std = out_RL[0] with mu, std = out_RL[0].detach() and that caused the error: RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn. If the gradient doesn't pass through the random operation, I don't understand why would detaching a tensor before the operation matter.
ANSWER
Answered 2021-Nov-30 at 05:48It is indeed true that sampling is not a differentiable operation per se. However, there exist two (broad) ways to mitigate this - [1] The REINFORCE way and [2] The reparameterization way. Since your example is related to [1], I will stick my answer to REINFORCE.
What REINFORCE does is it entirely gets rid of sampling operation in the computation graph. However, the sampling operation remains outside the graph. So, your statement
.. how does the gradient passes through a random operation ..
isn't correct. It does not pass through any random operation. Let's see your example
mu, std = out_RL[0]
dist = Normal(mu, std)
a = dist.sample()
log_p = dist.log_prob(a)
Computation of a does not involve creating a computation graph. It is technically equivalent to plugging in some offline data from a dataset (as in supervised learning)
mu, std = out_RL[0]
dist = Normal(mu, std)
# a = dist.sample()
a = torch.tensor([1.23, 4.01, -1.2, ...], device='cuda')
log_p = dist.log_prob(a)
Since we don't have offline data beforehand, we create them on the fly and the .sample() method does merely that.
So, there is no random operation on the graph. The log_p depends on mu and std deterministically, just like any standard computation graph. If you cut the connection like this
mu, std = out_RL[0].detach()
.. of course it is going to complaint.
Also, do not get confused by this operation
dist = Normal(mu, std)
log_p = dist.log_prob(a)
as it does not contain any randomness by itself. This is merely a shortcut for writing the tedious log-likelihood formula for Normal distribution.
QUESTION
What is the connection between discount factor gamma and horizon in RL.
What I have learned so far is that the horizon is the agent`s time to live. Intuitively, agents with finite horizon will choose actions differently than if it has to live forever. In the latter case, the agent will try to maximize all the expected rewards it may get far in the future.
But the idea of the discount factor is also the same. Are the values of gamma near zero makes the horizon finite?
ANSWER
Answered 2022-Mar-13 at 17:50Horizon refers to how many steps into the future the agent cares about the reward it can receive, which is a little different from the agent's time to live. In general, you could potentially define any arbitrary horizon you want as the objective. You could define a 10 step horizon, in which the agent makes a decision that will enable it to maximize the reward it will receive in the next 10 time steps. Or we could choose a 100, or 1000, or n step horizon!
Usually, the n-step horizon is defined using n = 1 / (1-gamma). Therefore, 10 step horizon will be achieved using gamma = 0.9, while 100 step horizon can be achieved with gamma = 0.99
Therefore, any value of gamma less than 1 imply that the horizon is finite.
QUESTION
I am trying to set a Deep-Q-Learning agent with a custom environment in OpenAI Gym. I have 4 continuous state variables with individual limits and 3 integer action variables with individual limits.
Here is the code:
#%% import
from gym import Env
from gym.spaces import Discrete, Box, Tuple
import numpy as np
#%%
class Custom_Env(Env):
def __init__(self):
# Define the state space
#State variables
self.state_1 = 0
self.state_2 = 0
self.state_3 = 0
self.state_4_currentTimeSlots = 0
#Define the gym components
self.action_space = Box(low=np.array([0, 0, 0]), high=np.array([10, 20, 27]), dtype=np.int)
self.observation_space = Box(low=np.array([20, -20, 0, 0]), high=np.array([22, 250, 100, 287]),dtype=np.float16)
def step(self, action ):
# Update state variables
self.state_1 = self.state_1 + action [0]
self.state_2 = self.state_2 + action [1]
self.state_3 = self.state_3 + action [2]
#Calculate reward
reward = self.state_1 + self.state_2 + self.state_3
#Set placeholder for info
info = {}
#Check if it's the end of the day
if self.state_4_currentTimeSlots >= 287:
done = True
if self.state_4_currentTimeSlots < 287:
done = False
#Move to the next timeslot
self.state_4_currentTimeSlots +=1
state = np.array([self.state_1,self.state_2, self.state_3, self.state_4_currentTimeSlots ])
#Return step information
return state, reward, done, info
def render (self):
pass
def reset (self):
self.state_1 = 0
self.state_2 = 0
self.state_3 = 0
self.state_4_currentTimeSlots = 0
state = np.array([self.state_1,self.state_2, self.state_3, self.state_4_currentTimeSlots ])
return state
#%% Set up the environment
env = Custom_Env()
#%% Create a deep learning model with keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten
from tensorflow.keras.optimizers import Adam
def build_model(states, actions):
model = Sequential()
model.add(Dense(24, activation='relu', input_shape=states))
model.add(Dense(24, activation='relu'))
model.add(Dense(actions[0] , activation='linear'))
return model
states = env.observation_space.shape
actions = env.action_space.shape
print("env.observation_space: ", env.observation_space)
print("env.observation_space.shape : ", env.observation_space.shape )
print("action_space: ", env.action_space)
print("action_space.shape : ", env.action_space.shape )
model = build_model(states, actions)
print(model.summary())
#%% Build Agent wit Keras-RL
from rl.agents import DQNAgent
from rl.policy import BoltzmannQPolicy
from rl.memory import SequentialMemory
def build_agent (model, actions):
policy = BoltzmannQPolicy()
memory = SequentialMemory(limit = 50000, window_length=1)
dqn = DQNAgent (model = model, memory = memory, policy=policy,
nb_actions=actions, nb_steps_warmup=10, target_model_update= 1e-2)
return dqn
dqn = build_agent(model, actions)
dqn.compile(Adam(lr=1e-3), metrics = ['mae'])
dqn.fit (env, nb_steps = 4000, visualize=False, verbose = 1)
When I run this code I get the following error message
ValueError: Model output "Tensor("dense_23/BiasAdd:0", shape=(None, 3), dtype=float32)" has invalid shape. DQN expects a model that has one dimension for each action, in this case (3,).
thrown by the line dqn = DQNAgent (model = model, memory = memory, policy=policy, nb_actions=actions, nb_steps_warmup=10, target_model_update= 1e-2)
Can anyone tell me, why this problem is occuring and how to solve this issue? I assume it has something to do with the built model and thus with the action and state spaces. But I could not figure out what exactly the problem is.
Reminder on the bounty: My bounty is expiring quite soon and unfortunately, I still have not received any answer. If you at least have a guess how to tackle that problem, I'll highly appreciate if you share your thoughts with me and I would be quite thankful for it.
ANSWER
Answered 2021-Dec-23 at 11:19As we talked about in the comments, it seems that the Keras-rl library is no longer supported (the last update in the repository was in 2019), so it's possible that everything is inside Keras now. I take a look at Keras documentation and there are no high-level functions to build a reinforcement learning model, but is possible to use lower-level functions to this.
- Here is an example of how to use Deep Q-Learning with Keras: link
Another solution may be to downgrade to Tensorflow 1.0 as it seems the compatibility problem occurs due to some changes in version 2.0. I didn't test, but maybe the Keras-rl + Tensorflow 1.0 may work.
There is also a branch of Keras-rl to support Tensorflow 2.0, the repository is archived, but there is a chance that it will work for you
QUESTION
Environment:
- Python: 3.9
- OS: Windows 10
When I try to create the ten armed bandits environment using the following code the error is thrown not sure of the reason.
import gym
import gym_armed_bandits
env = gym.make('ten-armed-bandits-v0')
The error:
---------------------------------------------------------------------------
KeyError Traceback (most recent call last)
File D:\00_PythonEnvironments\01_RL\lib\site-packages\gym\envs\registration.py:158, in EnvRegistry.spec(self, path)
157 try:
--> 158 return self.env_specs[id]
159 except KeyError:
160 # Parse the env name and check to see if it matches the non-version
161 # part of a valid env (could also check the exact number here)
KeyError: 'ten-armed-bandits-v0'
During handling of the above exception, another exception occurred:
UnregisteredEnv Traceback (most recent call last)
Input In [6], in
----> 1 env = gym.make('ten-armed-bandits-v0')
File D:\00_PythonEnvironments\01_RL\lib\site-packages\gym\envs\registration.py:235, in make(id, **kwargs)
234 def make(id, **kwargs):
--> 235 return registry.make(id, **kwargs)
File D:\00_PythonEnvironments\01_RL\lib\site-packages\gym\envs\registration.py:128, in EnvRegistry.make(self, path, **kwargs)
126 else:
127 logger.info("Making new env: %s", path)
--> 128 spec = self.spec(path)
129 env = spec.make(**kwargs)
130 return env
File D:\00_PythonEnvironments\01_RL\lib\site-packages\gym\envs\registration.py:203, in EnvRegistry.spec(self, path)
197 raise error.UnregisteredEnv(
198 "Toytext environment {} has been moved out of Gym. Install it via `pip install gym-legacy-toytext` and add `import gym_toytext` before using it.".format(
199 id
200 )
201 )
202 else:
--> 203 raise error.UnregisteredEnv("No registered env with id: {}".format(id))
UnregisteredEnv: No registered env with id: ten-armed-bandits-v0
When I check the environments available, I am able to see it there.
from gym import envs
print(envs.registry.all())
dict_values([EnvSpec(CartPole-v0), EnvSpec(CartPole-v1), EnvSpec(MountainCar-v0), EnvSpec(MountainCarContinuous-v0), EnvSpec(Pendulum-v1), EnvSpec(Acrobot-v1), EnvSpec(LunarLander-v2), EnvSpec(LunarLanderContinuous-v2), EnvSpec(BipedalWalker-v3), EnvSpec(BipedalWalkerHardcore-v3), EnvSpec(CarRacing-v0), EnvSpec(Blackjack-v1), EnvSpec(FrozenLake-v1), EnvSpec(FrozenLake8x8-v1), EnvSpec(CliffWalking-v0), EnvSpec(Taxi-v3), EnvSpec(Reacher-v2), EnvSpec(Pusher-v2), EnvSpec(Thrower-v2), EnvSpec(Striker-v2), EnvSpec(InvertedPendulum-v2), EnvSpec(InvertedDoublePendulum-v2), EnvSpec(HalfCheetah-v2), EnvSpec(HalfCheetah-v3), EnvSpec(Hopper-v2), EnvSpec(Hopper-v3), EnvSpec(Swimmer-v2), EnvSpec(Swimmer-v3), EnvSpec(Walker2d-v2), EnvSpec(Walker2d-v3), EnvSpec(Ant-v2), EnvSpec(Ant-v3), EnvSpec(Humanoid-v2), EnvSpec(Humanoid-v3), EnvSpec(HumanoidStandup-v2), EnvSpec(FetchSlide-v1), EnvSpec(FetchPickAndPlace-v1), EnvSpec(FetchReach-v1), EnvSpec(FetchPush-v1), EnvSpec(HandReach-v0), EnvSpec(HandManipulateBlockRotateZ-v0), EnvSpec(HandManipulateBlockRotateZTouchSensors-v0), EnvSpec(HandManipulateBlockRotateZTouchSensors-v1), EnvSpec(HandManipulateBlockRotateParallel-v0), EnvSpec(HandManipulateBlockRotateParallelTouchSensors-v0), EnvSpec(HandManipulateBlockRotateParallelTouchSensors-v1), EnvSpec(HandManipulateBlockRotateXYZ-v0), EnvSpec(HandManipulateBlockRotateXYZTouchSensors-v0), EnvSpec(HandManipulateBlockRotateXYZTouchSensors-v1), EnvSpec(HandManipulateBlockFull-v0), EnvSpec(HandManipulateBlock-v0), EnvSpec(HandManipulateBlockTouchSensors-v0), EnvSpec(HandManipulateBlockTouchSensors-v1), EnvSpec(HandManipulateEggRotate-v0), EnvSpec(HandManipulateEggRotateTouchSensors-v0), EnvSpec(HandManipulateEggRotateTouchSensors-v1), EnvSpec(HandManipulateEggFull-v0), EnvSpec(HandManipulateEgg-v0), EnvSpec(HandManipulateEggTouchSensors-v0), EnvSpec(HandManipulateEggTouchSensors-v1), EnvSpec(HandManipulatePenRotate-v0), EnvSpec(HandManipulatePenRotateTouchSensors-v0), EnvSpec(HandManipulatePenRotateTouchSensors-v1), EnvSpec(HandManipulatePenFull-v0), EnvSpec(HandManipulatePen-v0), EnvSpec(HandManipulatePenTouchSensors-v0), EnvSpec(HandManipulatePenTouchSensors-v1), EnvSpec(FetchSlideDense-v1), EnvSpec(FetchPickAndPlaceDense-v1), EnvSpec(FetchReachDense-v1), EnvSpec(FetchPushDense-v1), EnvSpec(HandReachDense-v0), EnvSpec(HandManipulateBlockRotateZDense-v0), EnvSpec(HandManipulateBlockRotateZTouchSensorsDense-v0), EnvSpec(HandManipulateBlockRotateZTouchSensorsDense-v1), EnvSpec(HandManipulateBlockRotateParallelDense-v0), EnvSpec(HandManipulateBlockRotateParallelTouchSensorsDense-v0), EnvSpec(HandManipulateBlockRotateParallelTouchSensorsDense-v1), EnvSpec(HandManipulateBlockRotateXYZDense-v0), EnvSpec(HandManipulateBlockRotateXYZTouchSensorsDense-v0), EnvSpec(HandManipulateBlockRotateXYZTouchSensorsDense-v1), EnvSpec(HandManipulateBlockFullDense-v0), EnvSpec(HandManipulateBlockDense-v0), EnvSpec(HandManipulateBlockTouchSensorsDense-v0), EnvSpec(HandManipulateBlockTouchSensorsDense-v1), EnvSpec(HandManipulateEggRotateDense-v0), EnvSpec(HandManipulateEggRotateTouchSensorsDense-v0), EnvSpec(HandManipulateEggRotateTouchSensorsDense-v1), EnvSpec(HandManipulateEggFullDense-v0), EnvSpec(HandManipulateEggDense-v0), EnvSpec(HandManipulateEggTouchSensorsDense-v0), EnvSpec(HandManipulateEggTouchSensorsDense-v1), EnvSpec(HandManipulatePenRotateDense-v0), EnvSpec(HandManipulatePenRotateTouchSensorsDense-v0), EnvSpec(HandManipulatePenRotateTouchSensorsDense-v1), EnvSpec(HandManipulatePenFullDense-v0), EnvSpec(HandManipulatePenDense-v0), EnvSpec(HandManipulatePenTouchSensorsDense-v0), EnvSpec(HandManipulatePenTouchSensorsDense-v1), EnvSpec(CubeCrash-v0), EnvSpec(CubeCrashSparse-v0), EnvSpec(CubeCrashScreenBecomesBlack-v0), EnvSpec(MemorizeDigits-v0), EnvSpec(three-armed-bandits-v0), EnvSpec(five-armed-bandits-v0), EnvSpec(ten-armed-bandits-v0), EnvSpec(MultiarmedBandits-v0)])
ANSWER
Answered 2022-Feb-08 at 08:01It could be a problem with your Python version: k-armed-bandits library was made 4 years ago, when Python 3.9 didn't exist. Besides this, the configuration files in the repo indicates that the Python version is 2.7 (not 3.9).
If you create an environment with Python 2.7 and follow the setup instructions it works correctly on Windows:
git clone gym_armed_bandits
cd gym_armed_bandits
pip install -e .
QUESTION
I have two different problems occurs at the same time.
I am having dimensionality problems with MaxPooling2d and having same dimensionality problem with DQNAgent.
The thing is, I can fix them seperately but cannot at the same time.
First Problem
I am trying to build a CNN network with several layers. After I build my model, when I try to run it, it gives me an error.
!pip install PyOpenGL==3.1.* PyOpenGL-accelerate==3.1.*
!pip install tensorflow gym keras-rl2 gym[atari] keras pyvirtualdisplay
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Convolution2D, MaxPooling2D, Activation
from keras_visualizer import visualizer
from tensorflow.keras.optimizers import Adam
env = gym.make('Boxing-v0')
height, width, channels = env.observation_space.shape
actions = env.action_space.n
input_shape = (3, 210, 160, 3) ## input_shape = (batch_size, height, width, channels)
def build_model(height, width, channels, actions):
model = Sequential()
model.add(Convolution2D(32, (8,8), strides=(4,4), activation="relu", input_shape=input_shape, data_format="channels_last"))
model.add(MaxPooling2D(pool_size=(2, 2), data_format="channels_last"))
model.add(Convolution2D(64, (4,4), strides=(1,1), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2), data_format="channels_last"))
model.add(Convolution2D(64, (3,3), activation="relu"))
model.add(Flatten())
model.add(Dense(512, activation="relu"))
model.add(Dense(256, activation="relu"))
model.add(Dense(actions, activation="linear"))
return model
model = build_model(height, width, channels, actions)
It gives below error:
ValueError: Input 0 of layer "max_pooling2d_12" is incompatible with the layer: expected ndim=4, found ndim=5. Full shape received: (None, 3, 51, 39, 32)
Second Problem
My input_shape is (3, 210, 160, 3). I am using the first 3 on purpose due to I have to specify the batch_size before. If I do not specify it before and pass it as (210, 160, 3) to the build_model function, below build_agent function gives me an another error:
def build_agent(model, actions):
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(), attr="eps", value_max=1., value_min=.1, value_test=.2, nb_steps=10000)
memory = SequentialMemory(limit=1000, window_length=3)
dqn = DQNAgent(model=model, memory=memory, policy=policy,
enable_dueling_network=True, dueling_type="avg",
nb_actions=actions, nb_steps_warmup=1000)
return dqn
dqn = build_agent(model, actions)
dqn.compile(Adam(learning_rate=1e-4))
dqn.fit(env, nb_steps=10000, visualize=False, verbose=1)
ValueError: Error when checking input: expected conv2d_11_input to have 4 dimensions, but got array with shape (1, 3, 210, 160, 3)
Deleting batch size number in the model construction phase, removes the MaxPooling2D incompatibility error but throws DQNAgent dimensionality error. Adding the batch size to the model construction phase removes DQNAgent dimensionality error but throws the MaxPooling2D incompatibility error.
I am really stucked.
ANSWER
Answered 2022-Feb-01 at 07:31Issue is with input_shape. input_shape=input_shape[1:]
Working sample code
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Convolution2D, MaxPooling2D, Activation
from tensorflow.keras.optimizers import Adam
input_shape = (3, 210, 160, 3)
model = Sequential()
model.add(Convolution2D(32, (8,8), strides=(4,4), activation="relu", input_shape=input_shape[1:], data_format="channels_last"))
model.add(MaxPooling2D(pool_size=(2,2), data_format="channels_last"))
model.add(Convolution2D(64, (4,4), strides=(1,1), activation="relu"))
model.add(MaxPooling2D(pool_size=(2, 2), data_format="channels_last"))
model.add(Convolution2D(64, (3,3), activation="relu"))
model.add(Flatten())
model.add(Dense(512, activation="relu"))
model.add(Dense(256, activation="relu"))
model.add(Dense(2, activation="linear"))
model.summary()
Output
Model: "sequential_7"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_9 (Conv2D) (None, 51, 39, 32) 6176
max_pooling2d_5 (MaxPooling (None, 25, 19, 32) 0
2D)
conv2d_10 (Conv2D) (None, 22, 16, 64) 32832
max_pooling2d_6 (MaxPooling (None, 11, 8, 64) 0
2D)
conv2d_11 (Conv2D) (None, 9, 6, 64) 36928
flatten_1 (Flatten) (None, 3456) 0
dense_4 (Dense) (None, 512) 1769984
dense_5 (Dense) (None, 256) 131328
dense_6 (Dense) (None, 2) 514
=================================================================
Total params: 1,977,762
Trainable params: 1,977,762
Non-trainable params: 0
QUESTION
I have this custom callback to log the reward in my custom vectorized environment, but the reward appears in console as always [0] and is not logged in tensorboard at all
class TensorboardCallback(BaseCallback):
"""
Custom callback for plotting additional values in tensorboard.
"""
def __init__(self, verbose=0):
super(TensorboardCallback, self).__init__(verbose)
def _on_step(self) -> bool:
self.logger.record('reward', self.training_env.get_attr('total_reward'))
return True
And this is part of the main function
model = PPO(
"MlpPolicy", env,
learning_rate=3e-4,
policy_kwargs=policy_kwargs,
verbose=1,
# as the environment is not serializable, we need to set a new instance of the environment
loaded_model = model = PPO.load("model", env=env)
loaded_model.set_env(env)
# and continue training
loaded_model.learn(1e+6, callback=TensorboardCallback())
tensorboard_log="./tensorboard/")
ANSWER
Answered 2021-Dec-25 at 01:10You need to add [0] as indexing,
so where you wrote self.logger.record('reward', self.training_env.get_attr('total_reward')) you just need to index with self.logger.record('reward', self.training_env.get_attr ('total_reward')[0])
class TensorboardCallback(BaseCallback):
"""
Custom callback for plotting additional values in tensorboard.
"""
def __init__(self, verbose=0):
super(TensorboardCallback, self).__init__(verbose)
def _on_step(self) -> bool:
self.logger.record('reward', self.training_env.get_attr('total_reward')[0])
return True
QUESTION
I followed a PyTorch tutorial to learn reinforcement learning(TRAIN A MARIO-PLAYING RL AGENT) but I am confused about the following code:
current_Q = self.net(state, model="online")[np.arange(0, self.batch_size), action] # Q_online(s,a)
What's the purpose of [np.arange(0, self.batch_size), action] after the neural network?(I know that TD_estimate takes in state and action, just confused about this on the programming side) What is this usage(put a list after self.net)?
More related code referenced from the tutorial:
class MarioNet(nn.Module):
def __init__(self, input_dim, output_dim):
super().__init__()
c, h, w = input_dim
if h != 84:
raise ValueError(f"Expecting input height: 84, got: {h}")
if w != 84:
raise ValueError(f"Expecting input width: 84, got: {w}")
self.online = nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=32, kernel_size=8, stride=4),
nn.ReLU(),
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=4, stride=2),
nn.ReLU(),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1),
nn.ReLU(),
nn.Flatten(),
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, output_dim),
)
self.target = copy.deepcopy(self.online)
# Q_target parameters are frozen.
for p in self.target.parameters():
p.requires_grad = False
def forward(self, input, model):
if model == "online":
return self.online(input)
elif model == "target":
return self.target(input)
self.net:
self.net = MarioNet(self.state_dim, self.action_dim).float()
Thanks for any help!
ANSWER
Answered 2021-Dec-23 at 11:07Essentially, what happens here is that the output of the net is being sliced to get the desired part of the Q table.
The (somewhat confusing) index of [np.arange(0, self.batch_size), action] indexes each axis. So, for axis with index 1, we pick the item indicated by action. For index 0, we pick all items between 0 and self.batch_size.
If self.batch_size is the same as the length of dimension 0 of this array, then this slice can be simplified to [:, action] which is probably more familiar to most users.
QUESTION
I'm trying to implement a DQN. As a warm up I want to solve CartPole-v0 with a MLP consisting of two hidden layers along with input and output layers. The input is a 4 element array [cart position, cart velocity, pole angle, pole angular velocity] and output is an action value for each action (left or right). I am not exactly implementing a DQN from the "Playing Atari with DRL" paper (no frame stacking for inputs etc). I also made a few non standard choices like putting done and the target network prediction of action value in the experience replay, but those choices shouldn't affect learning.
In any case I'm having a lot of trouble getting the thing to work. No matter how long I train the agent it keeps predicting a higher value for one action over another, for example Q(s, Right)> Q(s, Left) for all states s. Below is my learning code, my network definition, and some results I get from training
class DQN:
def __init__(self, env, steps_per_episode=200):
self.env = env
self.agent_network = MlpPolicy(self.env)
self.target_network = MlpPolicy(self.env)
self.target_network.load_state_dict(self.agent_network.state_dict())
self.target_network.eval()
self.optimizer = torch.optim.RMSprop(
self.agent_network.parameters(), lr=0.005, momentum=0.95
)
self.replay_memory = ReplayMemory()
self.gamma = 0.99
self.steps_per_episode = steps_per_episode
self.random_policy_stop = 1000
self.start_learning_time = 1000
self.batch_size = 32
def learn(self, episodes):
time = 0
for episode in tqdm(range(episodes)):
state = self.env.reset()
for step in range(self.steps_per_episode):
if time < self.random_policy_stop:
action = self.env.action_space.sample()
else:
action = select_action(self.env, time, state, self.agent_network)
new_state, reward, done, _ = self.env.step(action)
target_value_pred = predict_target_value(
new_state, reward, done, self.target_network, self.gamma
)
experience = Experience(
state, action, reward, new_state, done, target_value_pred
)
self.replay_memory.append(experience)
if time > self.start_learning_time: # learning step
experience_batch = self.replay_memory.sample(self.batch_size)
target_preds = extract_value_predictions(experience_batch)
agent_preds = agent_batch_preds(
experience_batch, self.agent_network
)
loss = torch.square(agent_preds - target_preds).sum()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if time % 1_000 == 0: # how frequently to update target net
self.target_network.load_state_dict(self.agent_network.state_dict())
self.target_network.eval()
state = new_state
time += 1
if done:
break
def agent_batch_preds(experience_batch: list, agent_network: MlpPolicy):
"""
Calculate the agent action value estimates using the old states and the
actual actions that the agent took at that step.
"""
old_states = extract_old_states(experience_batch)
actions = extract_actions(experience_batch)
agent_preds = agent_network(old_states)
experienced_action_values = agent_preds.index_select(1, actions).diag()
return experienced_action_values
def extract_actions(experience_batch: list) -> list:
"""
Extract the list of actions from experience replay batch and torchify
"""
actions = [exp.action for exp in experience_batch]
actions = torch.tensor(actions)
return actions
class MlpPolicy(nn.Module):
"""
This class implements the MLP which will be used as the Q network. I only
intend to solve classic control problems with this.
"""
def __init__(self, env):
super(MlpPolicy, self).__init__()
self.env = env
self.input_dim = self.env.observation_space.shape[0]
self.output_dim = self.env.action_space.n
self.fc1 = nn.Linear(self.input_dim, 32)
self.fc2 = nn.Linear(32, 128)
self.fc3 = nn.Linear(128, 32)
self.fc4 = nn.Linear(32, self.output_dim)
def forward(self, x):
if type(x) != torch.Tensor:
x = torch.tensor(x).float()
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = F.relu(self.fc3(x))
x = self.fc4(x)
return x
Learning results:
Here I'm seeing one action always valued over the others (Q(right, s) > Q(left, s)). It's also clear that the network is predicting the same action values for every state.
Does anyone have an idea about what's going on? I've done a lot of debugging and careful reading of the original papers (also thought about "normalizing" the observation space even though the velocities can be infinite) and could be missing something obvious at this point. I can include more code for the helper functions if that would be useful.
ANSWER
Answered 2021-Dec-19 at 16:09There was nothing wrong with the network definition. It turns out the learning rate was too high and reducing it 0.00025 (as in the original Nature paper introducing the DQN) led to an agent which can solve CartPole-v0.
That said, the learning algorithm was incorrect. In particular I was using the wrong target action-value predictions. Note the algorithm laid out above does not use the most recent version of the target network to make predictions. This leads to poor results as training progresses because the agent is learning based on stale target data. The way to fix this is to just put (s, a, r, s', done) into the replay memory and then make target predictions using the most up to date version of the target network when sampling a mini batch. See the code below for an updated learning loop.
def learn(self, episodes):
time = 0
for episode in tqdm(range(episodes)):
state = self.env.reset()
for step in range(self.steps_per_episode):
if time < self.random_policy_stop:
action = self.env.action_space.sample()
else:
action = select_action(self.env, time, state, self.agent_network)
new_state, reward, done, _ = self.env.step(action)
experience = Experience(state, action, reward, new_state, done)
self.replay_memory.append(experience)
if time > self.start_learning_time: # learning step.
experience_batch = self.replay_memory.sample(self.batch_size)
target_preds = target_batch_preds(
experience_batch, self.target_network, self.gamma
)
agent_preds = agent_batch_preds(
experience_batch, self.agent_network
)
loss = torch.square(agent_preds - target_preds).sum()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if time % 1_000 == 0: # how frequently to update target net
self.target_network.load_state_dict(self.agent_network.state_dict())
self.target_network.eval()
state = new_state
time += 1
if done:
break
Community Discussions, Code Snippets contain sources that include Stack Exchange Network
Vulnerabilities
No vulnerabilities reported
Install procgen
If you want to change the environments or create new ones, you should build from source. You can get miniconda from https://docs.conda.io/en/latest/miniconda.html if you don't have it, or install the dependencies from environment.yml manually. On Windows you will also need "Visual Studio 16 2019" installed. The environment code is in C++ and is compiled into a shared library exposing the gym3.libenv C interface that is then loaded by python. The C++ code uses Qt for drawing.
Support
Find, review, and download reusable Libraries, Code Snippets, Cloud APIs from over 650 million Knowledge Items
Find more librariesExplore Kits - Develop, implement, customize Projects, Custom Functions and Applications with kandi kits
Save this library and start creating your kit
Share this Page
