Skip to content

Instantly share code, notes, and snippets.

Embed
What would you like to do?
Training a Neural Network ATARI Pong agent with Policy Gradients from raw pixels
""" Trains an agent with (stochastic) Policy Gradients on Pong. Uses OpenAI Gym. """
import numpy as np
import cPickle as pickle
import gym
from chainer import cuda
import cupy as cp
import time, threading
#backend
be = cp
# hyperparameters
A = 3 # 2, 3 for no-ops
H = 200 # number of hidden layer neurons
update_freq = 10
batch_size = 1000 # every how many episodes to do a param update?
learning_rate = 1e-3
gamma = 0.99 # discount factor for reward
decay_rate = 0.99 # decay factor for RMSProp leaky sum of grad^2
resume = 0 # resume from previous checkpoint?
render = 0
device = 1
# model initialization
D = 80 * 80 # input dimensionality: 80x80 grid
with cp.cuda.Device(0):
if resume:
model = pickle.load(open('save.p', 'rb'))
print('resuming')
else:
model = {}
model['W1'] = np.random.randn(D,H) / np.sqrt(D) # "Xavier" initialization
model['W2'] = np.random.randn(H,A) / np.sqrt(H)
grad_buffer = { k : np.zeros_like(v) for k,v in model.iteritems() } # update buffers that add up gradients over a batch
rmsprop_cache = { k : np.zeros_like(v) for k,v in model.iteritems() } # rmsprop memory
def sigmoid(x):
return 1.0 / (1.0 + np.exp(-x)) # sigmoid "squashing" function to interval [0,1]
def softmax(x):
#if(len(x.shape)==1):
# x = x[np.newaxis,...]
probs = np.exp(x - np.max(x, axis=1, keepdims=True))
probs /= np.sum(probs, axis=1, keepdims=True)
return probs
def prepro(I):
""" prepro 210x160x3 uint8 frame into 6400 (80x80) 1D float vector """
I = I[35:195] # crop
I = I[::2,::2,0] # downsample by factor of 2
I[I == 144] = 0 # erase background (background type 1)
I[I == 109] = 0 # erase background (background type 2)
I[I != 0] = 1 # everything else (paddles, ball) just set to 1
return I.astype(np.float).ravel()
def discount_rewards(r):
""" take 1D float array of rewards and compute discounted reward """
discounted_r = np.zeros_like(r)
running_add = 0
for t in reversed(xrange(0, r.size)):
if r[t] != 0: running_add = 0 # reset the sum, since this was a game boundary (pong specific!)
running_add = running_add * gamma + r[t]
discounted_r[t] = running_add
return discounted_r
def policy_forward(x):
if(len(x.shape)==1):
x = x[np.newaxis,...]
h = x.dot(model['W1'])
h[h<0] = 0 # ReLU nonlinearity
logp = h.dot(model['W2'])
#p = sigmoid(logp)
p = softmax(logp)
return p, h # return probability of taking action 2, and hidden state
def policy_backward(eph, epdlogp):
""" backward pass. (eph is array of intermediate hidden states) """
dW2 = eph.T.dot(epdlogp)
dh = epdlogp.dot(model['W2'].T)
dh[eph <= 0] = 0 # backpro prelu
t = time.time()
if(be == cp):
dh_gpu = cuda.to_gpu(dh, device=0)
epx_gpu = cuda.to_gpu(epx.T, device=0)
dW1 = cuda.to_cpu( epx_gpu.dot(dh_gpu) )
else:
dW1 = epx.T.dot(dh)
print((time.time()-t0)*1000, ' ms, @final bprop')
return {'W1':dW1, 'W2':dW2}
env = gym.make("Pong-v0")
observation = env.reset()
prev_x = None # used in computing the difference frame
xs,hs,dlogps,drs = [],[],[],[]
running_reward = None
reward_sum = 0
episode_number = 0
while True:
t0 = time.time()
if render:
t = time.time()
env.render()
print((time.time()-t)*1000, ' ms, @rendering')
t = time.time()
# preprocess the observation, set input to network to be difference image
cur_x = prepro(observation)
x = cur_x - prev_x if prev_x is not None else np.zeros(D)
prev_x = cur_x
#print((time.time()-t)*1000, ' ms, @prepo')
# forward the policy network and sample an action from the returned probability
t = time.time()
aprob, h = policy_forward(x)
#action = 2 if np.random.uniform() < aprob else 3 # roll the dice!
#print((time.time()-t)*1000, ' ms, @forward')
# roll the dice, in the softmax loss
u = np.random.uniform()
aprob_cum = np.cumsum(aprob)
a = np.where(u <= aprob_cum)[0][0]
action = a+2
#print(u, a, aprob_cum)
# record various intermediates (needed later for backprop)
t = time.time()
xs.append(x) # observation
hs.append(h) # hidden state
#softmax loss gradient
dlogsoftmax = aprob.copy()
dlogsoftmax[0,a] -= 1 #-discounted reward
dlogps.append(dlogsoftmax)
# step the environment and get new measurements
t = time.time()
observation, reward, done, info = env.step(action)
reward_sum += reward
#print((time.time()-t)*1000, ' ms, @env.step')
drs.append(reward) # record reward (has to be done after we call step() to get reward for previous action)
#print((time.time()-t0)*1000, ' ms, @whole.step')
if done: # an episode finished
episode_number += 1
t = time.time()
# stack together all inputs, hidden states, action gradients, and rewards for this episode
epx = np.vstack(xs)
eph = np.vstack(hs)
epdlogp = np.vstack(dlogps)
epr = np.vstack(drs)
xs,hs,dlogps,drs = [],[],[],[] # reset array memory
print(epdlogp.shape)
# compute the discounted reward backwards through time
discounted_epr = discount_rewards(epr)
# standardize the rewards to be unit normal (helps control the gradient estimator variance)
discounted_epr -= np.mean(discounted_epr)
discounted_epr /= np.std(discounted_epr)
epdlogp *= discounted_epr # modulate the gradient with advantage (PG magic happens right here.)
grad = policy_backward(eph, epdlogp)
for k in model: grad_buffer[k] += grad[k] # accumulate grad over batch
# perform rmsprop parameter update every batch_size episodes
if episode_number % update_freq == 0: #update_freq used to be batch_size
for k,v in model.iteritems():
g = grad_buffer[k] # gradient
rmsprop_cache[k] = decay_rate * rmsprop_cache[k] + (1 - decay_rate) * g**2
model[k] -= learning_rate * g / (np.sqrt(rmsprop_cache[k]) + 1e-5)
grad_buffer[k] = np.zeros_like(v) # reset batch gradient buffer
# boring book-keeping
running_reward = reward_sum if running_reward is None else running_reward * 0.99 + reward_sum * 0.01
print 'resetting env. episode reward total was %f. running mean: %f' % (reward_sum, running_reward)
if episode_number % 100 == 0: pickle.dump(model, open('save.p', 'wb'))
reward_sum = 0
observation = env.reset() # reset env
prev_x = None
print((time.time()-t)*1000, ' ms, @backprop')
if reward != 0: # Pong has either +1 or -1 reward exactly when game ends.
print ('ep %d: game finished, reward: %f' % (episode_number, reward)) + ('' if reward == -1 else ' !!!!!!!!')
@etienne87

This comment has been minimized.

Copy link
Owner Author

@etienne87 etienne87 commented Jun 9, 2016

special thanx to Marin Toromanoff's correction on update (- instead of +). the modif works with gpu for backward (the only place worth it) with cupy (you can install it via sudo pip install chainer) & several actions, so we can try great Karpathy's code on other games maybe? In my trial this works fairly well even with noop

@SalemAmeen

This comment has been minimized.

Copy link

@SalemAmeen SalemAmeen commented Jun 13, 2016

Where I can find the dataset for this experiment?

@dwhit

This comment has been minimized.

Copy link

@dwhit dwhit commented Jun 15, 2016

why do you set the learning rate to 10 if the episode number is divisible by 100? it's never reset, so effectively the learning rate is 10 after episode 100.

@etienne87

This comment has been minimized.

Copy link
Owner Author

@etienne87 etienne87 commented Jun 16, 2016

sorry for late answer! (i am noob with gist, i haven't received any notification)
@SalemAmeen, you need to install open-ai gym. follow original post of karpathy : http://karpathy.github.io/2016/05/31/rl/
@dwhit, thanks for correcting, it was not in my original experiment.

@steleu

This comment has been minimized.

Copy link

@steleu steleu commented Jul 28, 2016

Thanks etienne87 and Karpathy for the code.

Just wonder if it would even works better if a deeper neural network is used (e.g Convolutional neural network).
Also, may I ask why set action = a+2?

@neighborzhang

This comment has been minimized.

Copy link

@neighborzhang neighborzhang commented Dec 13, 2016

I also have the same concern. Would it be better to have more layers?

@mariolew

This comment has been minimized.

Copy link

@mariolew mariolew commented Jan 13, 2017

@etienne87 Hi, I think the line
model[k] -= learning_rate * g / (np.sqrt(rmsprop_cache[k]) + 1e-5)
should use - instead of +, too.
But when I changed + to - with Karpathy's original post, it doesn't work, I wonder how? Could you please explain it for me?
Thanks.

@yxchng

This comment has been minimized.

Copy link

@yxchng yxchng commented Feb 21, 2017

Why this code gives nan when test when CartPole environment? Can't seem to figure about the bug though. Do you have any idea?

@unos21

This comment has been minimized.

Copy link

@unos21 unos21 commented Apr 20, 2017

What are the extra packages (other than that of karpathy) that need to be installed to run this code?

@normferns

This comment has been minimized.

Copy link

@normferns normferns commented Apr 27, 2017

@etienne87 and @mariolew:

I posted the following comment on @karpathy's gist in regard to the use of + or -.

"RMSProp is presented in CS231 in the context of gradient descent, wherein the goal is to move the parameters downward (in the negative direction of the gradient) in order to minimize a loss function.

Here, in the Monte-Carlo Policy Gradient method, we are using gradient ascent; we are trying to move the parameters upward (in the positive direction of the gradient) in order to maximize an objective function.

This is why a plus sign is used here, whereas a minus sign is used in the class notes."

@normferns

This comment has been minimized.

Copy link

@normferns normferns commented Apr 27, 2017

@yxchng:

This error probably results when you attempt to divide by the standard deviation in line 184. If I am not mistaken, the CartPole environment returns an immediate reward of 1 at each decision point until the end of the episode. So a reward vector would look like a list of ones, e.g. [1,1,1,1,1,1,1]. With the code as written, the vector of returns (discounted_epr) would remain unchanged, i.e it would remain [1,1,1,1,1,1,1]. The standard deviation is therefore zero, and therefore your code would be attempting to divide by zero.

@HadXu

This comment has been minimized.

Copy link

@HadXu HadXu commented Feb 13, 2018

can you share pytorch version? I encounter a problems,thanks

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.