I am try to train a DQN model with the following code. The GPU (cuda) usage is always lower than 25 percent. I know the tensorflow backend is consulting the GPU resources, but the usage is low. Is there any way I can improve the utilization of the GPU (When I train a CNN network, the GPU (cude) utilization is around 70 percent)?
How can I modify the code and make it run faster or use more GPU resources. Or it is very hard for the Dqn to work on a GPU?
import numpy as np
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Dropout, Conv2D, MaxPooling2D, Activation, Flatten
from tensorflow.keras.optimizers import Adam
import tensorflow as tf
from collections import deque
import time
import random
from tqdm import tqdm
import os
from PIL import Image
import cv2
DISCOUNT = 0.99
REPLAY_MEMORY_SIZE = 50_000 # How many last steps to keep for model training
MIN_REPLAY_MEMORY_SIZE = 1_000 # Minimum number of steps in a memory to start training
MINIBATCH_SIZE = 64 # How many steps (samples) to use for training
UPDATE_TARGET_EVERY = 5 # Terminal states (end of episodes)
MODEL_NAME = '2x256'
MIN_REWARD = -200 # For model save
MEMORY_FRACTION = 0.20
# Environment settings
EPISODES = 20_000
# Exploration settings
epsilon = 1 # not a constant, going to be decayed
EPSILON_DECAY = 0.99975
MIN_EPSILON = 0.001
# Stats settings
AGGREGATE_STATS_EVERY = 50 # episodes
SHOW_PREVIEW = False
class Blob:
def __init__(self, size):
self.size = size
self.x = np.random.randint(0, size)
self.y = np.random.randint(0, size)
def __str__(self):
return f"Blob ({self.x}, {self.y})"
def __sub__(self, other):
return (self.x-other.x, self.y-other.y)
def __eq__(self, other):
return self.x == other.x and self.y == other.y
def action(self, choice):
'''
Gives us 9 total movement options. (0,1,2,3,4,5,6,7,8)
'''
if choice == 0:
self.move(x=1, y=1)
elif choice == 1:
self.move(x=-1, y=-1)
elif choice == 2:
self.move(x=-1, y=1)
elif choice == 3:
self.move(x=1, y=-1)
elif choice == 4:
self.move(x=1, y=0)
elif choice == 5:
self.move(x=-1, y=0)
elif choice == 6:
self.move(x=0, y=1)
elif choice == 7:
self.move(x=0, y=-1)
elif choice == 8:
self.move(x=0, y=0)
def move(self, x=False, y=False):
# If no value for x, move randomly
if not x:
self.x += np.random.randint(-1, 2)
else:
self.x += x
# If no value for y, move randomly
if not y:
self.y += np.random.randint(-1, 2)
else:
self.y += y
# If we are out of bounds, fix!
if self.x < 0:
self.x = 0
elif self.x > self.size-1:
self.x = self.size-1
if self.y < 0:
self.y = 0
elif self.y > self.size-1:
self.y = self.size-1
class BlobEnv:
SIZE = 10
RETURN_IMAGES = True
MOVE_PENALTY = 1
ENEMY_PENALTY = 300
FOOD_REWARD = 25
OBSERVATION_SPACE_VALUES = (SIZE, SIZE, 3) # 4
ACTION_SPACE_SIZE = 9
PLAYER_N = 1 # player key in dict
FOOD_N = 2 # food key in dict
ENEMY_N = 3 # enemy key in dict
# the dict! (colors)
d = {1: (255, 175, 0),
2: (0, 255, 0),
3: (0, 0, 255)}
def reset(self):
self.player = Blob(self.SIZE)
self.food = Blob(self.SIZE)
while self.food == self.player:
self.food = Blob(self.SIZE)
self.enemy = Blob(self.SIZE)
while self.enemy == self.player or self.enemy == self.food:
self.enemy = Blob(self.SIZE)
self.episode_step = 0
if self.RETURN_IMAGES:
observation = np.array(self.get_image())
else:
observation = (self.player-self.food) + (self.player-self.enemy)
return observation
def step(self, action):
self.episode_step += 1
self.player.action(action)
#### MAYBE ###
#enemy.move()
#food.move()
##############
if self.RETURN_IMAGES:
new_observation = np.array(self.get_image())
else:
new_observation = (self.player-self.food) + (self.player-self.enemy)
if self.player == self.enemy:
reward = -self.ENEMY_PENALTY
elif self.player == self.food:
reward = self.FOOD_REWARD
else:
reward = -self.MOVE_PENALTY
done = False
if reward == self.FOOD_REWARD or reward == -self.ENEMY_PENALTY or self.episode_step >= 200:
done = True
return new_observation, reward, done
def render(self):
img = self.get_image()
img = img.resize((300, 300)) # resizing so we can see our agent in all its glory.
cv2.imshow("image", np.array(img)) # show it!
cv2.waitKey(1)
# FOR CNN #
def get_image(self):
env = np.zeros((self.SIZE, self.SIZE, 3), dtype=np.uint8) # starts an rbg of our size
env[self.food.x][self.food.y] = self.d[self.FOOD_N] # sets the food location tile to green color
env[self.enemy.x][self.enemy.y] = self.d[self.ENEMY_N] # sets the enemy location to red
env[self.player.x][self.player.y] = self.d[self.PLAYER_N] # sets the player tile to blue
img = Image.fromarray(env, 'RGB') # reading to rgb. Apparently. Even tho color definitions are bgr. ???
return img
env = BlobEnv()
# For stats
ep_rewards = [-200]
# For more repetitive results
random.seed(1)
np.random.seed(1)
tf.compat.v1.set_random_seed(1)
# Memory fraction, used mostly when trai8ning multiple agents
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=MEMORY_FRACTION)
#backend.set_session(tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)))
# Create models folder
if not os.path.isdir('models'):
os.makedirs('models')
# Agent class
class DQNAgent:
def __init__(self):
# Main model
self.model = self.create_model()
# Target network
self.target_model = self.create_model()
self.target_model.set_weights(self.model.get_weights())
# An array with last n steps for training
self.replay_memory = deque(maxlen=REPLAY_MEMORY_SIZE)
# Used to count when to update target network with main network's weights
self.target_update_counter = 0
def create_model(self):
model = Sequential()
model.add(Conv2D(256, (3, 3), input_shape=env.OBSERVATION_SPACE_VALUES)) # OBSERVATION_SPACE_VALUES = (10, 10, 3) a 10x10 RGB image.
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Conv2D(256, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Flatten()) # this converts our 3D feature maps to 1D feature vectors
model.add(Dense(64))
model.add(Dense(env.ACTION_SPACE_SIZE, activation='linear')) # ACTION_SPACE_SIZE = how many choices (9)
model.compile(loss="mse", optimizer=Adam(lr=0.001), metrics=['accuracy'])
return model
# Adds step's data to a memory replay array
# (observation space, action, reward, new observation space, done)
def update_replay_memory(self, transition):
self.replay_memory.append(transition)
# Trains main network every step during episode
def train(self, terminal_state, step):
# Start training only if certain number of samples is already saved
if len(self.replay_memory) < MIN_REPLAY_MEMORY_SIZE:
return
# Get a minibatch of random samples from memory replay table
minibatch = random.sample(self.replay_memory, MINIBATCH_SIZE)
# Get current states from minibatch, then query NN model for Q values
current_states = np.array([transition[0] for transition in minibatch])/255
current_qs_list = self.model.predict(current_states, verbose=0)
# Get future states from minibatch, then query NN model for Q values
# When using target network, query it, otherwise main network should be queried
new_current_states = np.array([transition[3] for transition in minibatch])/255
future_qs_list = self.target_model.predict(new_current_states, verbose=0)
X = []
y = []
# Now we need to enumerate our batches
for index, (current_state, action, reward, new_current_state, done) in enumerate(minibatch):
# If not a terminal state, get new q from future states, otherwise set it to 0
# almost like with Q Learning, but we use just part of equation here
if not done:
max_future_q = np.max(future_qs_list[index])
new_q = reward + DISCOUNT * max_future_q
else:
new_q = reward
# Update Q value for given state
current_qs = current_qs_list[index]
current_qs[action] = new_q
# And append to our training data
X.append(current_state)
y.append(current_qs)
# Fit on all samples as one batch, log only on terminal state
self.model.fit(np.array(X)/255, np.array(y), batch_size=MINIBATCH_SIZE, verbose=0, shuffle=False)
# Update target network counter every episode
if terminal_state:
self.target_update_counter += 1
# If counter reaches set value, update target network with weights of main network
if self.target_update_counter > UPDATE_TARGET_EVERY:
self.target_model.set_weights(self.model.get_weights())
self.target_update_counter = 0
# Queries main network for Q values given current observation space (environment state)
def get_qs(self, state):
return self.model.predict(np.array(state).reshape(-1, *state.shape)/255)[0]
agent = DQNAgent()
# Iterate over episodes
for episode in tqdm(range(1, EPISODES + 1), ascii=True, unit='episodes'):
# Restarting episode - reset episode reward and step number
episode_reward = 0
step = 1
# Reset environment and get initial state
current_state = env.reset()
# Reset flag and start iterating until episode ends
done = False
while not done:
# This part stays mostly the same, the change is to query a model for Q values
if np.random.random() > epsilon:
# Get action from Q table
action = np.argmax(agent.get_qs(current_state))
else:
# Get random action
action = np.random.randint(0, env.ACTION_SPACE_SIZE)
new_state, reward, done = env.step(action)
# Transform new continous state to new discrete state and count reward
episode_reward += reward
if SHOW_PREVIEW and not episode % AGGREGATE_STATS_EVERY:
env.render()
# Every step we update replay memory and train main network
agent.update_replay_memory((current_state, action, reward, new_state, done))
agent.train(done, step)
current_state = new_state
step += 1
# Append episode reward to a list and log stats (every given number of episodes)
ep_rewards.append(episode_reward)
if not episode % AGGREGATE_STATS_EVERY or episode == 1:
average_reward = sum(ep_rewards[-AGGREGATE_STATS_EVERY:])/len(ep_rewards[-AGGREGATE_STATS_EVERY:])
min_reward = min(ep_rewards[-AGGREGATE_STATS_EVERY:])
max_reward = max(ep_rewards[-AGGREGATE_STATS_EVERY:])
# Save model, but only when min reward is greater or equal a set value
if min_reward >= MIN_REWARD:
agent.model.save(f'models/{MODEL_NAME}__{max_reward:_>7.2f}max_{average_reward:_>7.2f}avg_{min_reward:_>7.2f}min__{int(time.time())}.model')
# Decay epsilon
if epsilon > MIN_EPSILON:
epsilon *= EPSILON_DECAY
epsilon = max(MIN_EPSILON, epsilon)
I would suggest you to profile your application with dlprof or Tensorflow profiler to understand what causes the underutilization. It also gives you suggestions about how to solve that. As far as I can see from your code, your workload is very small for a modern GPU anyways. So, it seems normal that your application does not take advantage of every CUDA/Tensor core available in your GPU.