def dqn_learing(env,
q_func,
optimizer_spec,
exploration,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000):
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
with torch.no_grad():
return model(Variable(obs)).data.max(1)[1].cpu()
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function, i.e. build the model.
######
# YOUR CODE HERE
Q = q_func(input_arg, num_actions)
Q_target = q_func(input_arg, num_actions)
######
# Construct Q network optimizer function
optimizer = optimizer_spec.constructor(Q.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
LOG_EVERY_N_STEPS = 10000
for t in count():
### 1. Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
### 2. Step the env and store the transition
# At this point, "last_obs" contains the latest observation that was
# recorded from the simulator. Here, your code needs to store this
# observation and its outcome (reward, next observation, etc.) into
# the replay buffer while stepping the simulator forward one step.
# At the end of this block of code, the simulator should have been
# advanced one step, and the replay buffer should contain one more
# transition.
# Specifically, last_obs must point to the new latest observation.
# Useful functions you'll need to call:
# obs, reward, done, info = env.step(action)
# this steps the environment forward one step
# obs = env.reset()
# this resets the environment if you reached an episode boundary.
# Don't forget to call env.reset() to get a new observation if done
# is true!!
# Note that you cannot use "last_obs" directly as input
# into your network, since it needs to be processed to include context
# from previous frames. You should check out the replay buffer
# implementation in dqn_utils.py to see what functionality the replay
# buffer exposes. The replay buffer has a function called
# encode_recent_observation that will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
# Don't forget to include epsilon greedy exploration!
# And remember that the first time you enter this loop, the model
# may not yet have been initialized (but of course, the first step
# might as well be random, since you haven't trained your net...)
#####
idx = replay_buffer.store_frame(last_obs)
encoded_obs = replay_buffer.encode_recent_observation()
if (t > learning_starts):
action = select_epilson_greedy_action(Q, encoded_obs, t)
else:
action = random.randrange(num_actions)
obs, reward, done, _ = env.step(action)
replay_buffer.store_effect(idx, action, reward, done)
if (done):
last_obs = env.reset()
else:
last_obs = obs
#####
# at this point, the environment should have been advanced one step (and
# reset if done was true), and last_obs should point to the new latest
# observation
### 3. Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Here, you should perform training. Training consists of four steps:
# 3.a: use the replay buffer to sample a batch of transitions (see the
# replay buffer code for function definition, each batch that you sample
# should consist of current observations, current actions, rewards,
# next observations, and done indicator).
# Note: Move the variables to the GPU if avialable
# 3.b: fill in your own code to compute the Bellman error. This requires
# evaluating the current and next Q-values and constructing the corresponding error.
# Note: don't forget to clip the error between [-1,1], multiply is by -1 (since pytorch minimizes) and
# maskout post terminal status Q-values (see ReplayBuffer code).
# 3.c: train the model. To do this, use the bellman error you calculated perviously.
# Pytorch will differentiate this error for you, to backward the error use the following API:
# current.backward(d_error.data.unsqueeze(1))
# Where "current" is the variable holding current Q Values and d_error is the clipped bellman error.
# Your code should produce one scalar-valued tensor.
# Note: don't forget to call optimizer.zero_grad() before the backward call and
# optimizer.step() after the backward call.
# 3.d: periodically update the target network by loading the current Q network weights into the
# target_Q network. see state_dict() and load_state_dict() methods.
# you should update every target_update_freq steps, and you may find the
# variable num_param_updates useful for this (it was initialized to 0)
#####
# YOUR CODE HERE
#
# Alpha (learning rate) from the q function update isn't present in our code -- its in OptimizerSpec in main.
# Move to GPU if possible
# done flag in loop ---- SKIPPED IF DONE IS TRUE
# clipping the error between -1 and 1 -- OK
# backward the error meaning?
# Suggestion for changing parameters - change exploration scehdule (main)
#
# Q.cuda()
obs_batch, act_batch, reward_batch, next_obs_batch, done_mask = replay_buffer.sample(
batch_size=batch_size)
states = Variable(torch.from_numpy(obs_batch).type(dtype) / 255.0)
actions = Variable(torch.from_numpy(act_batch).long())
rewards = Variable(torch.from_numpy(reward_batch).float())
next_states = Variable(
torch.from_numpy(next_obs_batch).type(dtype) / 255.0)
not_dones = Variable(torch.from_numpy(1 - done_mask).type(dtype))
if USE_CUDA:
states = states.cuda()
actions = actions.cuda()
rewards = rewards.cuda()
next_states = next_states.cuda()
Q.train()
Q_target.eval()
predicted_rewards = Q(states).gather(1,
actions.unsqueeze(1)) #Q(s,a)
next_max_Q = Q_target(next_states).detach().max(1)[
0] #.unsqueeze(1) #Q_target(s,a)
next_Q_values = not_dones * next_max_Q
target_Q_values = rewards + (gamma * next_Q_values) #r + Q_target
bellman_error = target_Q_values - predicted_rewards.squeeze(1)
clipped_bellman_error = bellman_error.clamp(-1, 1) * (-1.0)
optimizer.zero_grad()
predicted_rewards.backward(clipped_bellman_error.data.unsqueeze(1))
optimizer.step()
num_param_updates += +1
if (num_param_updates % target_update_freq == 0):
Q_target.load_state_dict(Q.state_dict())
# for obs,act,reward,next_obs,done in zip(obs_batch,act_batch,reward_batch,next_obs_batch,done_mask):
# if(done == 1.0):
# continue
# obs = Variable(torch.from_numpy(obs, ).type(dtype).unsqueeze(0) / 255.0, requires_grad=True)
# next_obs = Variable(torch.from_numpy(next_obs).type(dtype).unsqueeze(0) / 255.0, requires_grad=False)
# current_Q = Q(obs)
# predicted_reward = Variable(current_Q[0][act].unsqueeze(0), requires_grad=True)
# target_reward = Q_target(next_obs).data.max(1)[0]
# loss = loss_fn(reward + gamma * target_reward, predicted_reward).clamp(-1, 1) * (-1.0)
# optimizer.zero_grad()
# # should be current.backward(d_error.data.unsqueeze(1))
# # but it crashes on misfitting dims
# predicted_reward.backward(loss.data.unsqueeze(1))
# optimizer.step()
#####
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
sys.stdout.flush()
# Dump statistics to pickle
with open('statistics.pkl', 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % 'statistics.pkl')
def dqn_learing(env,
q_func,
optimizer_spec,
exploration,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000):
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
return model(Variable(obs, volatile=True)).data.max(1)[1].cpu()
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function
Q = q_func(input_arg, num_actions).type(dtype)
target_Q = q_func(input_arg, num_actions).type(dtype)
# Construct Q network optimizer function
optimizer = optimizer_spec.constructor(Q.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
LOG_EVERY_N_STEPS = 10000
for t in count():
### Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
### Step the env and store the transition
# Store lastest observation in replay memory and last_idx can be used to store action, reward, done
last_idx = replay_buffer.store_frame(last_obs)
# encode_recent_observation will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
recent_observations = replay_buffer.encode_recent_observation()
# Choose random action if not yet start learning
if t > learning_starts:
action = select_epilson_greedy_action(Q, recent_observations, t)[0,
0]
else:
action = random.randrange(num_actions)
# Advance one step
obs, reward, done, _ = env.step(action)
# clip rewards between -1 and 1
reward = max(-1.0, min(reward, 1.0))
# Store other info in replay memory
replay_buffer.store_effect(last_idx, action, reward, done)
# Resets the environment when reaching an episode boundary.
if done:
obs = env.reset()
last_obs = obs
### Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Use the replay buffer to sample a batch of transitions
# Note: done_mask[i] is 1 if the next state corresponds to the end of an episode,
# in which case there is no Q-value at the next state; at the end of an
# episode, only the current state reward contributes to the target
obs_batch, act_batch, rew_batch, next_obs_batch, done_mask = replay_buffer.sample(
batch_size)
# Convert numpy nd_array to torch variables for calculation
obs_batch = Variable(
torch.from_numpy(obs_batch).type(dtype) / 255.0)
act_batch = Variable(torch.from_numpy(act_batch).long())
rew_batch = Variable(torch.from_numpy(rew_batch))
next_obs_batch = Variable(
torch.from_numpy(next_obs_batch).type(dtype) / 255.0)
not_done_mask = Variable(torch.from_numpy(1 -
done_mask)).type(dtype)
if USE_CUDA:
act_batch = act_batch.cuda()
rew_batch = rew_batch.cuda()
# Compute current Q value, q_func takes only state and output value for every state-action pair
# We choose Q based on action taken.
current_Q_values = Q(obs_batch).gather(1, act_batch.unsqueeze(1))
# Compute next Q value based on which action gives max Q values
# Detach variable from the current graph since we don't want gradients for next Q to propagated
next_max_q = target_Q(next_obs_batch).detach().max(1)[0]
next_Q_values = not_done_mask * next_max_q
# Compute the target of the current Q values
target_Q_values = rew_batch + (gamma * next_Q_values)
# Compute Bellman error
bellman_error = target_Q_values - current_Q_values
# clip the bellman error between [-1 , 1]
clipped_bellman_error = bellman_error.clamp(-1, 1)
# Note: clipped_bellman_delta * -1 will be right gradient
d_error = clipped_bellman_error * -1.0
# Clear previous gradients before backward pass
optimizer.zero_grad()
# run backward pass
current_Q_values.backward(d_error.data.unsqueeze(1))
# Perfom the update
optimizer.step()
num_param_updates += 1
# Periodically update the target network by Q network to target Q network
if num_param_updates % target_update_freq == 0:
target_Q.load_state_dict(Q.state_dict())
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
sys.stdout.flush()
# Dump statistics to pickle
with open('statistics.pkl', 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % 'statistics.pkl')
d_error = clipped_bellman_error * -1.0
# Clear previous gradients before backward pass
optimizer.zero_grad()
# run backward pass
current_Q_values.backward(d_error.data.unsqueeze(1))
# Perfom the update
optimizer.step()
num_param_updates += 1
# Periodically update the target network by Q network to target Q network
if num_param_updates % target_update_freq == 0:
target_Q.load_state_dict(Q.state_dict())
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env, "Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
def stopping_criterion(env):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
def dqn_learing(env,
q_func,
optimizer_spec,
exploration,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000):
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
return model(Variable(obs,
volatile=True)).data.max(1)[1].view(1, 1)
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function
Q = q_func(input_arg, num_actions).type(dtype)
target_Q = q_func(input_arg, num_actions).type(dtype)
# Construct Q network optimizer function
optimizer = optimizer_spec.constructor(Q.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
LOG_EVERY_N_STEPS = 10000
rewards = []
episodes = []
reward_list = []
for t in count():
### Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
### Step the env and store the transition
# Store lastest observation in replay memory and last_idx can be used to store action, reward, done
last_idx = replay_buffer.store_frame(last_obs)
# encode_recent_observation will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
recent_observations = replay_buffer.encode_recent_observation()
# Choose random action if not yet start learning
if t > learning_starts:
#action = select_epilson_greedy_action(Q, recent_observations, t)[0, 0]
action = select_epilson_greedy_action(Q, recent_observations,
t)[0][0]
else:
action = random.randrange(num_actions)
# Advance one step
obs, reward, done, _ = env.step(action)
# clip rewards between -1 and 1
reward = max(-1.0, min(reward, 1.0))
if t == 20000000:
pylab.plot(episodes, reward_list, 'b')
pylab.savefig("./save_graph/breakout_dqn.png")
# Store other info in replay memory
replay_buffer.store_effect(last_idx, action, reward, done)
# Resets the environment when reaching an episode boundary.
if done:
obs = env.reset()
last_obs = obs
### Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Use the replay buffer to sample a batch of transitions
# Note: done_mask[i] is 1 if the next state corresponds to the end of an episode,
# in which case there is no Q-value at the next state; at the end of an
# episode, only the current state reward contributes to the target
obs_batch, act_batch, rew_batch, next_obs_batch, done_mask = replay_buffer.sample(
batch_size)
# Convert numpy nd_array to torch variables for calculation
obs_batch = Variable(
torch.from_numpy(obs_batch).type(dtype) / 255.0)
act_batch = Variable(torch.from_numpy(act_batch).long())
rew_batch = Variable(torch.from_numpy(rew_batch))
next_obs_batch = Variable(
torch.from_numpy(next_obs_batch).type(dtype) / 255.0)
not_done_mask = Variable(torch.from_numpy(1 -
done_mask)).type(dtype)
if USE_CUDA:
act_batch = act_batch.cuda()
rew_batch = rew_batch.cuda()
# Compute current Q value, q_func takes only state and output value for every state-action pair
# We choose Q based on action taken.
# current_Q_values = Q(obs_batch).gather(1, act_batch.unsqueeze(1))
current_Q_values = Q(obs_batch).gather(
1, act_batch.unsqueeze(1)
).squeeze(
) # squeeze the [batch_size x 1] Tensor to have a shape of batch_size
# Compute next Q value based on which action gives max Q values
# Detach variable from the current graph since we don't want gradients for next Q to propagated
next_max_q = target_Q(next_obs_batch).detach().max(1)[0]
next_Q_values = not_done_mask * next_max_q
# Compute the target of the current Q values
target_Q_values = rew_batch + (gamma * next_Q_values)
# # Compute Bellman error
# bellman_error = target_Q_values - current_Q_values
# # clip the bellman error between [-1 , 1]
# clipped_bellman_error = bellman_error.clamp(-1, 1)
# # Note: clipped_bellman_delta * -1 will be right gradient
# d_error = clipped_bellman_error * -1.0
loss = F.smooth_l1_loss(current_Q_values, target_Q_values)
# Clear previous gradients before backward pass
optimizer.zero_grad()
# run backward pass
# current_Q_values.backward(d_error.data.unsqueeze(1))
loss.backward()
# Clip the gradients to lie between -1 and +1
for params in Q.parameters():
params.grad.data.clamp_(-1, 1)
# Perfom the update
optimizer.step()
num_param_updates += 1
# Periodically update the target network by Q network to target Q network
if num_param_updates % target_update_freq == 0:
target_Q.load_state_dict(Q.state_dict())
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
sys.stdout.flush()
reward_list.append(mean_episode_reward)
episodes.append(episode_rewards)
# Dump statistics to pickle
with open('statistics.pkl', 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % 'statistics.pkl')
q_func=DQN,
optimizer_spec=optimizer_spec,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=REPLAY_BUFFER_SIZE,
batch_size=BATCH_SIZE,
gamma=GAMMA,
learning_starts=LEARNING_STARTS,
learning_freq=LEARNING_FREQ,
frame_history_len=FRAME_HISTORY_LEN,
target_update_freq=TARGER_UPDATE_FREQ,
statistics_file_name=run.statistics_file_name
)
if __name__ == '__main__':
# Get Atari games.
benchmark = gym.benchmark_spec('Atari40M')
# Change the index to select a different game.
task = benchmark.tasks[3]
# Run training
seed = 0 # Use a seed of zero (you may want to randomize the seed!)
env = get_env(task, seed)
# do not record videos:
get_wrapper_by_name(env, "Monitor").video_callable = lambda episode_id: False
main(env, task.max_timesteps)
def dqn_learing(env,
q_func,
optimizer_spec,
exploration=LinearSchedule(1000000, 0.1),
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10):
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env, t) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
grad_norm_clipping: float or None
If not None gradients' norms are clipped to this value.
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
return model(Variable(obs, volatile=True)).data.max(1)[1].cpu()
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function
Q = q_func(input_arg, num_actions).type(dtype)
target_Q = q_func(input_arg, num_actions).type(dtype)
# Construct Q network optimizer function
# optimizer_func = construct_optimizer_func(Q, optimizer_spec)
optimizer = torch.optim.Adam(Q.parameters())
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
LOG_EVERY_N_STEPS = 10000
for t in count():
### Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
### Step the env and store the transition
# Store lastest observation in replay memory and last_idx can be used to store action, reward, done
last_idx = replay_buffer.store_frame(last_obs)
# encode_recent_observation will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
# recent_observations: shape(img_h, img_w, frame_history_len) are input to to the model
recent_observations = replay_buffer.encode_recent_observation(
).transpose(2, 0, 1)
# Choose random action if not yet start learning
if t > learning_starts:
action = select_epilson_greedy_action(Q, recent_observations, t)[0,
0]
else:
action = random.randrange(num_actions)
# Advance one step
obs, reward, done, _ = env.step(action)
replay_buffer.store_effect(last_idx, action, reward, done)
# Resets the environment when reaching an episode boundary.
if done:
obs = env.reset()
last_obs = obs
### Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Use the replay buffer to sample a batch of transitions
# Note: done_mask[i] is 1 if the next state corresponds to the end of an episode,
# in which case there is no Q-value at the next state; at the end of an
# episode, only the current state reward contributes to the target
obs_batch, act_batch, rew_batch, next_obs_batch, done_mask = replay_buffer.sample(
batch_size)
# Convert numpy nd_array to torch variables for calculation
obs_batch = Variable(
torch.from_numpy(obs_batch.transpose(0, 3, 1, 2)).type(dtype) /
255.0)
act_batch = Variable(torch.from_numpy(act_batch).long())
rew_batch = Variable(torch.from_numpy(rew_batch))
next_obs_batch = Variable(
torch.from_numpy(next_obs_batch.transpose(
0, 3, 1, 2)).type(dtype) / 255.0)
done_mask = torch.from_numpy(done_mask)
if USE_CUDA:
act_batch = act_batch.cuda()
rew_batch = rew_batch.cuda()
done_mask = done_mask.cuda()
# Compute current Q value, q_func takes only state and output value for every state-action pair
# We choose Q based on action taken.
current_Q_values = Q(obs_batch).gather(1, act_batch.unsqueeze(1))
# Compute next Q value, based on which acion gives max Q values
next_max_Q_values = Variable(torch.zeros(batch_size).type(dtype))
# # Detach variable from the current graph since we don't want gradients to propagated
next_max_Q_values[done_mask == 0] = target_Q(
next_obs_batch).detach().max(1)[0]
# Compute Bellman error, use huber loss to mitigate outlier impact
target_Q_values = rew_batch + (gamma * next_max_Q_values)
bellman_error = F.smooth_l1_loss(current_Q_values, target_Q_values)
# Construct and optimizer and clear previous gradients
optimizer = optimizer_func(t)
optimizer.zero_grad()
# run backward pass and clip the gradient
bellman_error.backward()
nn.utils.clip_grad_norm(Q.parameters(), grad_norm_clipping)
# Perfom the update
optimizer.step()
num_param_updates += 1
# Periodically update the target network by Q network to target Q network
if num_param_updates % target_update_freq == 0:
target_Q.load_state_dict(Q.state_dict())
### 4. Log progress
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
print("learning_rate %f" % optimizer_spec.lr_schedule.value(t))
sys.stdout.flush()
def dqn_learing(env,
q_func,
optimizer_spec,
exploration,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
statistics_file_name="statistics.pkl"):
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
statistics_file_name: str
Where to store the statistics file
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
print("STATISTICS_FILE_NAME: {}".format(statistics_file_name))
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(
torch_types.FloatTensor).unsqueeze(0) / 255.0
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
with torch.no_grad():
return model(Variable(obs)).data.max(1)[1].cpu()
else:
return random.randrange(num_actions)
# Initialize target q function and q function, i.e. build the model.
######
# YOUR CODE HERE
policy_net = q_func(input_arg, num_actions).to(device).type(
torch_types.FloatTensor) # Q
target_net = q_func(input_arg, num_actions).to(device).type(
torch_types.FloatTensor) # Q target
target_net.load_state_dict(
policy_net.state_dict()) # copies the state of policy Q into target
######
# Construct policy_net network optimizer function
optimizer = optimizer_spec.constructor(policy_net.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
LOG_EVERY_N_STEPS = 10000
for t in count():
### 1. Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
### 2. Step the env and store the transition
# At this point, "last_obs" contains the latest observation that was
# recorded from the simulator. Here, your code needs to store this
# observation and its outcome (reward, next observation, etc.) into
# the replay buffer while stepping the simulator forward one step.
# At the end of this block of code, the simulator should have been
# advanced one step, and the replay buffer should contain one more
# transition.
# Specifically, last_obs must point to the new latest observation.
# Useful functions you'll need to call:
# obs, reward, done, info = env.step(action)
# this steps the environment forward one step
# obs = env.reset()
# this resets the environment if you reached an episode boundary.
# Don't forget to call env.reset() to get a new observation if done
# is true!!
# Note that you cannot use "last_obs" directly as input
# into your network, since it needs to be processed to include context
# from previous frames. You should check out the replay buffer
# implementation in dqn_utils.py to see what functionality the replay
# buffer exposes. The replay buffer has a function called
# encode_recent_observation that will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
# Don't forget to include epsilon greedy exploration!
# And remember that the first time you enter this loop, the model
# may not yet have been initialized (but of course, the first step
# might as well be random, since you haven't trained your net...)
#####
# YOUR CODE HERE
stored_frame_idx = replay_buffer.store_frame(last_obs)
last_obs_encoded = replay_buffer.encode_recent_observation()
action = select_epilson_greedy_action(policy_net, last_obs_encoded, t)
obs, reward, done, info = env.step(action)
replay_buffer.store_effect(stored_frame_idx, action, reward, done)
if done:
obs = env.reset()
last_obs = obs
#####
# at this point, the environment should have been advanced one step (and
# reset if done was true), and last_obs should point to the new latest
# observation
### 3. Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Here, you should perform training. Training consists of four steps:
# 3.a: use the replay buffer to sample a batch of transitions (see the
# replay buffer code for function definition, each batch that you sample
# should consist of current observations, current actions, rewards,
# next observations, and done indicator).
# Note: Move the variables to the GPU if avialable
# 3.b: fill in your own code to compute the Bellman error. This requires
# evaluating the current and next Q-values and constructing the corresponding error.
# Note: don't forget to clip the error between [-1,1], multiply is by -1 (since pytorch minimizes) and
# maskout post terminal status Q-values (see ReplayBuffer code).
# 3.c: train the model. To do this, use the bellman error you calculated perviously.
# Pytorch will differentiate this error for you, to backward the error use the following API:
# current.backward(d_error.data.unsqueeze(1))
# Where "current" is the variable holding current Q Values and d_error is the clipped bellman error.
# Your code should produce one scalar-valued tensor.
# Note: don't forget to call optimizer.zero_grad() before the backward call and
# optimizer.step() after the backward call.
# 3.d: periodically update the target network by loading the current Q network weights into the
# target_Q network. see state_dict() and load_state_dict() methods.
# you should update every target_update_freq steps, and you may find the
# variable num_param_updates useful for this (it was initialized to 0)
#####
# YOUR CODE HERE
sample = replay_buffer.sample(batch_size)
obs_batch, actions_batch, rewards_batch, next_obs_batch, done_mask = sample
# convert batches to pytorch tensors:
obs_batch = torch.from_numpy(obs_batch).to(device).type(
torch_types.FloatTensor) / 255.0
next_obs_batch = torch.from_numpy(next_obs_batch).to(device).type(
torch_types.FloatTensor) / 255.0
actions_batch = torch.from_numpy(actions_batch).to(device).type(
torch_types.LongTensor)
rewards_batch = torch.from_numpy(rewards_batch).to(device).type(
torch_types.FloatTensor)
non_final_mask = 1 - torch.from_numpy(done_mask).to(device).type(
torch_types.FloatTensor)
# inspired by https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html:
# Compute Q(s_t, a) - the model computes Q(s_t), then we select the
# columns of actions taken
state_action_values = policy_net(obs_batch).gather(
1, actions_batch.unsqueeze(1)).squeeze(1)
# Compute V(s_{t+1}) for all next states.
next_state_values = target_net(next_obs_batch).max(
1)[0].detach() * non_final_mask
# Compute the expected Q values
expected_state_action_values = (next_state_values *
gamma) + rewards_batch
# Compute loss
d_error = state_action_values - expected_state_action_values # = -bellman_error
d_error.clamp_(-1, 1)
# Optimize the model
optimizer.zero_grad()
state_action_values.backward(d_error)
optimizer.step()
num_param_updates += 1
# Periodically update target network:
if num_param_updates % target_update_freq == 0:
target_net.load_state_dict(policy_net.state_dict())
#####
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t >= learning_starts:
print("Timestep %d" % (t, ))
print(" mean reward (100 episodes) %f" % mean_episode_reward)
print(" best mean reward %f" % best_mean_episode_reward)
print(" episodes %d" % len(episode_rewards))
print(" exploration %f" % exploration.value(t))
sys.stdout.flush()
# Dump statistics to pickle
with open(statistics_file_name, 'wb') as f:
pickle.dump(Statistic, f)
def train(self):
num_param_updates = 0
loss_acc_since_last_log = 0.0
param_updates_since_last_log = 0
num_episodes = 0
state = self.env.reset()[..., np.newaxis]
for t in tqdm(range(self.total_timesteps)):
last_idx = self.memory.store_frame(state)
recent_observations = self.memory.encode_recent_observation()
# Choose random action if learning hasn't started yet
if t > self.learning_start:
action = self.select_epsilon_greedy_action(
recent_observations, t).item()
else:
action = random.randrange(self.num_actions)
# Advance a step
next_state, reward, done, _ = self.env.step(action)
next_state = next_state[..., np.newaxis]
# Store result in memory
self.memory.store_effect(last_idx, action, reward, done)
# Reset if done (life lost, due to atari wrapper)
if done:
next_state = self.env.reset()
next_state = next_state[..., np.newaxis]
state = next_state
# Train network using experience replay when
# memory is sufficiently large.
if (t > self.learning_start and t % self.learning_freq == 0
and self.memory.can_sample(self.batch_size)):
# Sample from replay buffer
(
state_batch,
act_batch,
r_batch,
next_state_batch,
done_mask,
) = self.memory.sample(self.batch_size)
state_batch = torch.from_numpy(state_batch).type(
self.dtype) / 255.0
act_batch = torch.from_numpy(act_batch).long().to(self.device)
r_batch = torch.from_numpy(r_batch).to(self.device)
next_state_batch = (
torch.from_numpy(next_state_batch).type(self.dtype) /
255.0)
not_done_mask = torch.from_numpy(1 - done_mask).type(
self.dtype)
# Calculate current Q value
current_Q_vals = self.Q(state_batch).gather(
1, act_batch.unsqueeze(1))
# Calculate next Q value based on action that gives max Q vals
next_max_Q = self.target_Q(next_state_batch).detach().max(
dim=1)[0]
next_Q_vals = not_done_mask * next_max_Q
# Calculate target of current Q values
target_Q_vals = r_batch + (self.gamma * next_Q_vals)
# Calculate loss and backprop
loss = F.smooth_l1_loss(current_Q_vals.squeeze(),
target_Q_vals)
self.optimizer.zero_grad()
loss.backward()
for param in self.Q.parameters():
param.grad.data.clamp_(-1, 1)
# Update weights
self.optimizer.step()
num_param_updates += 1
# Store stats
loss_acc_since_last_log += loss.item()
param_updates_since_last_log += 1
# Update target network periodically
if num_param_updates % self.target_update_freq == 0:
self.target_Q.load_state_dict(self.Q.state_dict())
# Save model checkpoint
if num_param_updates % self.checkpoint_frequency == 0:
save_model_checkpoint(
self.Q,
self.optimizer,
t,
f"{self.out_dir}/checkpoints/{self.model_name}_{num_param_updates}",
)
# Log progress
if (num_param_updates % (self.log_freq // 2) == 0
and param_updates_since_last_log > 0):
self.writer.add_scalar(
"Mean Loss per Update (Updates)",
loss_acc_since_last_log / param_updates_since_last_log,
num_param_updates,
)
loss_acc_since_last_log = 0.0
param_updates_since_last_log = 0
if num_param_updates % self.log_freq == 0:
wrapper = get_wrapper_by_name(self.env, "Monitor")
episode_rewards = wrapper.get_episode_rewards()
mean_reward = round(np.mean(episode_rewards[-101:-1]), 2)
sum_reward = np.sum(episode_rewards[-101:-1])
episode_lengths = wrapper.get_episode_lengths()
mean_duration = round(np.mean(episode_lengths[-101:-1]), 2)
sum_duration = np.sum(episode_lengths[-101:-1])
self.writer.add_scalar(
f"Mean Reward (epoch = {self.log_freq} updates)",
mean_reward,
num_param_updates // self.log_freq,
)
self.writer.add_scalar(
f"Mean Duration (epoch = {self.log_freq} updates)",
mean_duration,
num_param_updates // self.log_freq,
)
self.writer.add_scalar(
f"Mean Reward per Timestep (epoch = {self.log_freq} updates)",
round(sum_reward / sum_duration, 2),
num_param_updates // self.log_freq,
)
if done:
num_episodes += 1
# Save model
save_model(self.Q, f"{self.out_dir}/{self.model_name}.model")
self.env.close()
print(f"Number of Episodes: {num_episodes}")
return self.Q
def should_stop(self):
return (get_wrapper_by_name(self.env, "Monitor").get_total_steps() >=
self.max_steps)
def dqn_learing(env,
q_func,
optimizer_spec,
exploration,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000):
print("running new version")
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
return model(Variable(obs, volatile=True)).data.max(1)[1].cpu()
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function, i.e. build the model.
""" ---------------------------- OUR CODE ---------------------------- """
Q = q_func(input_arg, num_actions) # The parameters are random
Qtag = q_func(input_arg, num_actions)
if (USE_CUDA):
Q.cuda()
Qtag.cuda()
Qtag.load_state_dict(Q.state_dict())
# Construct Q network optimizer function
optimizer = optimizer_spec.constructor(Q.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
""" ------------------------------------------------------------------ """
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
reward = None
done = None
info = None
LOG_EVERY_N_STEPS = 10000
startTime = time.time()
for t in count():
""" Tsuf: ---- Stuff for debigging times for various places --- """
T1 = 0
t1Tmp = 0
T2 = 0
t2Tmp = 0
T3 = 0
t3Tmp = 0
T4 = 0
t4Tmp = 0
T5 = 0
t5Tmp = 0
T6 = 0
t6Tmp = 0
T7 = 0
t7Tmp = 0
T8 = 0
t8Tmp = 0
""" ----------------------------------------------------------- """
### 1. Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
#if (t>1000000):
# break
### 2. Step the env and store the transition
# At this point, "last_obs" contains the latest observation that was
# recorded from the simulator. Here, your code needs to store this
# observation and its outcome (reward, next observation, etc.) into
# the replay buffer while stepping the simulator forward one step.
# At the end of this block of code, the simulator should have been
# advanced one step, and the replay buffer should contain one more
# transition.
# Specifically, last_obs must point to the new latest observation.
# Useful functions you'll need to call:
# obs, reward, done, info = env.step(action)
# this steps the environment forward one step
# obs = env.reset()
# this resets the environment if you reached an episode boundary.
# Don't forget to call env.reset() to get a new observation if done
# is true!!
# Note that you cannot use "last_obs" directly as input
# into your network, since it needs to be processed to include context
# from previous frames. You should check out the replay buffer
# implementation in dqn_utils.py to see what functionality the replay
# buffer exposes. The replay buffer has a function called
# encode_recent_observation that will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
# Don't forget to include epsilon greedy exploration!
# And remember that the first time you enter this loop, the model
# may not yet have been initialized (but of course, the first step
# might as well be random, since you haven't trained your net...)
""" -------------------------- OUR CODE -------------------------- """
#store last_obs, and get latest obs's as the input for the n.n
t1Tmp = time.time()
cur_idx = replay_buffer.store_frame(last_obs)
next_input = replay_buffer.encode_recent_observation()
T1 += time.time() - t1Tmp
#take random action or use the net
t2Tmp = time.time()
action = select_epilson_greedy_action(
Q, next_input, t) #the returned action is on the CPU
T2 += time.time() - t2Tmp
#see what happens after we take that action
t3Tmp = time.time()
last_obs, reward, done, info = env.step(
action) #the returned parameters are on the CPU
T3 += time.time() - t3Tmp
# print(t)
# env.render()
#store the results on the replay buffer
replay_buffer.store_effect(cur_idx, action, reward, done) #on the CPU
#if the simulation is done, reset the environment
if (done):
last_obs = env.reset()
""" -------------------------------------------------------------- """
# at this point, the environment should have been advanced one step (and
# reset if done was true), and last_obs should point to the new latest
# observation
### 3. Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Here, you should perform training. Training consists of four steps:
# 3.a: use the replay buffer to sample a batch of transitions (see the
# replay buffer code for function definition, each batch that you sample
# should consist of current observations, current actions, rewards,
# next observations, and done indicator).
# Note: Move the variables to the GPU if avialable
# 3.b: fill in your own code to compute the Bellman error. This requires
# evaluating the current and next Q-values and constructing the corresponding error.
# Note: don't forget to clip the error between [-1,1], multiply is by -1 (since pytorch minimizes) and
# maskout post terminal status Q-values (see ReplayBuffer code).
# 3.c: train the model. To do this, use the bellman error you calculated perviously.
# Pytorch will differentiate this error for you, to backward the error use the following API:
# current.backward(d_error.data.unsqueeze(1))
# Where "current" is the variable holding current Q Values and d_error is the clipped bellman error.
# Your code should produce one scalar-valued tensor.
# Note: don't forget to call optimizer.zero_grad() before the backward call and
# optimizer.step() after the backward call.
# 3.d: periodically update the target network by loading the current Q network weights into the
# target_Q network. see state_dict() and load_state_dict() methods.
# you should update every target_update_freq steps, and you may find the
# variable num_param_updates useful for this (it was initialized to 0)
""" ------------------------ OUR CODE ------------------------ """
#sample a batch of history samples
t4Tmp = time.time()
obs_batch, act_batch, rew_batch, next_obs_batch, done_mask = replay_buffer.sample(
batch_size) #on CPU
obs_batch = torch.from_numpy(obs_batch).type(
dtype) / 255.0 # When available, move the samples batch to GPU
next_obs_batch = torch.from_numpy(next_obs_batch).type(
dtype) / 255.0 #GPU
T4 += time.time() - t4Tmp
#see which Q values the current network gives, for all obs's
t5Tmp = time.time()
inter_Qs = Q(
Variable(obs_batch)) #input is on GPU, output is on GPU
inter_Qs_chosen = Variable(
torch.zeros(batch_size).type(dtype)) #GPU
#take the action that was chosen before
for i in range(batch_size):
inter_Qs_chosen[i] = inter_Qs[i, act_batch[i]]
#take only the intermediate (non-terminal) obs's
inter_idx = np.where(done_mask == False)[0] #CPU
inter_next_obs_batch = next_obs_batch[inter_idx, :, :, :]
T5 += time.time() - t5Tmp
#see what the "target" (backuped) network says for the intermediate ones
t6Tmp = time.time()
inter_next_Qs = Qtag(
Variable(inter_next_obs_batch,
volatile=True)).data.max(1)[0] #All on GPU
T6 += time.time() - t6Tmp
#calculate the bellman errors
t7Tmp = time.time()
#for final obs's, the target is just the reward
targets = torch.from_numpy(rew_batch).type(
dtype) #Moved rew_batch to GPU (as 'targets')
for (i, idx) in enumerate(inter_idx):
targets[idx] += gamma * inter_next_Qs[i] #The bellman item
# errors = -(inter_Qs_chosen.data - targets)**2 #EQUATION COULD BE WRONG!! [on GPU]
# for i in range(len(errors)):
# if errors[i]<-1:
# errors[i] = -1
# elif errors[i]>1:
# errors[i] = 1
errors = inter_Qs_chosen.data - targets
errors.clamp(-1, 1)
T7 += time.time() - t7Tmp
#train the network! (:
t8Tmp = time.time()
optimizer.zero_grad()
inter_Qs_chosen.backward(
errors) #COULD BE WRONG WAY!! [Everything is on GPU (: ]
optimizer.step()
T8 += time.time() - t8Tmp
num_param_updates += 1
if (num_param_updates % target_update_freq == 0):
Qtag.load_state_dict(Q.state_dict())
""" ---------------------------------------------------------- """
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
Statistic["running_times"].append(int(time.time() - startTime))
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
if (PRINT_TIMES):
print("-----------------------")
print(T1)
print(T2)
print(T3)
print(T4)
print(T5)
print(T6)
print(T7)
print(T8)
print("-----------------------")
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
sys.stdout.flush()
# Dump statistics to pickle
with open('statistics.pkl', 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % 'statistics.pkl')
def dqn_learing(
env,
q_func,
optimizer_spec,
exploration,
feature_tested,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
):
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
# added bool_flag for double save statistic
bool_flag = False
STATS_FILE_NAME = 'statistics ' + feature_tested + '.pkl'
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
with torch.no_grad():
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
return model(Variable(obs)).data.max(1)[1].cpu()
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function, i.e. build the model.
######
# YOUR CODE HERE
if USE_CUDA:
Q = q_func(num_actions=num_actions).cuda()
Q_target = q_func(num_actions=num_actions).cuda()
else:
Q = q_func(num_actions=num_actions)
Q_target = q_func(num_actions=num_actions)
Q_target.load_state_dict(Q.state_dict())
# Check & load pretrained model
if os.path.isfile('Q_params' + feature_tested + '.pkl'):
print('Load Q parameters ...')
Q.load_state_dict(torch.load('Q_params' + feature_tested + '.pkl'))
if os.path.isfile('target_Q_params' + feature_tested + '.pkl'):
print('Load target Q parameters ...')
Q_target.load_state_dict(
torch.load('target_Q_params' + feature_tested + '.pkl'))
######
# Construct Q network optimizer function
optimizer = optimizer_spec.constructor(Q.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
Statistic = {
"starting_Q_values": [],
"mean_episode_rewards": [],
"best_mean_episode_rewards": []
}
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
# load prev Stats
start = 0
if os.path.isfile(STATS_FILE_NAME):
with open(STATS_FILE_NAME, 'rb') as f:
Statistic = pickle.load(f)
mean_episode_reward = Statistic["mean_episode_rewards"][-1]
best_mean_episode_reward = Statistic["best_mean_episode_rewards"][
-1]
start = len(Statistic["mean_episode_rewards"])
print('Load %s ...' % STATS_FILE_NAME)
done = False
###############
# RUN ENV #
###############
num_param_updates = 0
last_obs = env.reset()
LOG_EVERY_N_STEPS = 50000 #10000
for t in count(start):
### 1. Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
return Statistic
break
# couldnt handle stopping_criterion, this works:
if t > 4500000:
return Statistic
break
### 2. Step the env and store the transition
# At this point, "last_obs" contains the latest observation that was
# recorded from the simulator. Here, your code needs to store this
# observation and its outcome (reward, next observation, etc.) into
# the replay buffer while stepping the simulator forward one step.
# At the end of this block of code, the simulator should have been
# advanced one step, and the replay buffer should contain one more
# transition.
# Specifically, last_obs must point to the new latest observation.
# Useful functions you'll need to call:
# obs, reward, done, info = env.step(action)
# this steps the environment forward one step
# obs = env.reset()
# this resets the environment if you reached an episode boundary.
# Don't forget to call env.reset() to get a new observation if done
# is true!!
# Note that you cannot use "last_obs" directly as input
# into your network, since it needs to be processed to include context
# from previous frames. You should check out the replay buffer
# implementation in dqn_utils.py to see what functionality the replay
# buffer exposes. The replay buffer has a function called
# encode_recent_observation that will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
# Don't forget to include epsilon greedy exploration!
# And remember that the first time you enter this loop, the model
# may not yet have been initialized (but of course, the first step
# might as well be random, since you haven't trained your net...)
#####
# YOUR CODE HERE
idx = replay_buffer.store_frame(last_obs)
encoded_obs = replay_buffer.encode_recent_observation()
action = select_epilson_greedy_action(Q, encoded_obs, t)
# if started a new game - store Q-func
######
# done not initialized since last time
if t > learning_starts and done:
# a very expensive statistic - so don't log frequently
with torch.no_grad():
obs = torch.from_numpy(encoded_obs).type(dtype).unsqueeze(
0) / 255.0
item = torch.max(Q(Variable(obs))).item()
Statistic["starting_Q_values"].append(item)
######
# this steps the environment forward one step
last_obs, reward, done, info = env.step(action)
replay_buffer.store_effect(idx, action, reward, done)
if done:
last_obs = env.reset()
#####
# at this point, the environment should have been advanced one step (and
# reset if done was true), and last_obs should point to the new latest
# observation
### 3. Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Here, you should perform training. Training consists of four steps:
# 3.a: use the replay buffer to sample a batch of transitions (see the
# replay buffer code for function definition, each batch that you sample
# should consist of current observations, current actions, rewards,
# next observations, and done indicator).
# Note: Move the variables to the GPU if avialable
# 3.b: fill in your own code to compute the Bellman error. This requires
# evaluating the current and next Q-values and constructing the corresponding error.
# Note: don't forget to clip the error between [-1,1], multiply is by -1 (since pytorch minimizes) and
# maskout post terminal status Q-values (see ReplayBuffer code).
# 3.c: train the model. To do this, use the bellman error you calculated perviously.
# Pytorch will differentiate this error for you, to backward the error use the following API:
# current.backward(d_error.data.unsqueeze(1))
# Where "current" is the variable holding current Q Values and d_error is the clipped bellman error.
# Your code should produce one scalar-valued tensor.
# Note: don't forget to call optimizer.zero_grad() before the backward call and
# optimizer.step() after the backward call.
# 3.d: periodically update the target network by loading the current Q network weights into the
# target_Q network. see state_dict() and load_state_dict() methods.
# you should update every target_update_freq steps, and you may find the
# variable num_param_updates useful for this (it was initialized to 0)
#####
# YOUR CODE HERE
# 3.a sample a batch of transitions
sample = replay_buffer.sample(batch_size)
obs_batch, action_batch, reward_batch, next_obs_batch, done_mask = sample
# move variables to GPU if available
obs_batch = Variable(
torch.from_numpy(obs_batch).type(dtype)) / 255.0
action_batch = Variable(
torch.from_numpy(action_batch).type(dtype).long().view(-1, 1))
reward_batch = Variable(torch.from_numpy(reward_batch).type(dtype))
next_obs_batch = Variable(
torch.from_numpy(next_obs_batch).type(dtype)) / 255.0
done_mask = Variable(torch.from_numpy(done_mask).type(dtype))
# 3.b compute the Bellman error
# evaluating the current and next Q-values
state_action_values = Q(obs_batch).gather(1, action_batch)
next_state_values = Q_target(next_obs_batch).detach()
# maskout post terminal status Q-values
masked_next_state_values = next_state_values.max(1)[0] * (
1 - done_mask)
# constructing the corresponding error
expected_state_action_values = (masked_next_state_values *
gamma) + reward_batch
bellman_error = expected_state_action_values.unsqueeze(
1) - state_action_values
# clip the error between [-1,1]
clipped_bellman_error = bellman_error.clamp(-1, 1)
optimizer.zero_grad()
# multiply by -1 (since pytorch minimizes)
state_action_values.backward(-clipped_bellman_error)
# 3.c: train the model
optimizer.step()
# 3.d periodically update the target network
num_param_updates += 1
if num_param_updates % target_update_freq == 0:
Q_target.load_state_dict(Q.state_dict())
#####
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
sys.stdout.flush()
# Save the trained model
torch.save(Q.state_dict(), 'Q_params' + feature_tested + '.pkl')
torch.save(Q_target.state_dict(),
'target_Q_params' + feature_tested + '.pkl')
# Dump statistics to pickle
#double save
if bool_flag:
bool_flag = False
with open(STATS_FILE_NAME, 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % STATS_FILE_NAME)
else:
bool_flag = True
with open('copy_' + STATS_FILE_NAME, 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % 'copy_' + STATS_FILE_NAME)
plt.clf()
plt.xlabel('Num of Games')
plt.ylabel('Q-values on starting state')
plt.plot(range(len(Statistic["starting_Q_values"])),
Statistic["starting_Q_values"],
label='Q-values')
plt.legend()
plt.title(feature_tested)
plt.savefig('Q-value-Performance' + feature_tested + '.png')
plt.clf()
plt.xlabel('Timesteps')
plt.ylabel('Mean Reward (past 100 episodes)')
num_items = len(Statistic["mean_episode_rewards"])
plt.plot(range(num_items),
Statistic["mean_episode_rewards"],
label='mean reward')
plt.plot(range(num_items),
Statistic["best_mean_episode_rewards"],
label='best mean rewards')
plt.legend()
plt.title(feature_tested)
plt.savefig('DeepQ-Performance' + feature_tested + '.png')
def dqn_learing(env,
q_func,
optimizer_spec,
exploration,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000):
"""Run Deep Q-learning algorithm.
You can specify your own convnet using q_func.
All schedules are w.r.t. total number of steps taken in the environment.
Parameters
----------
env: gym.Env
gym environment to train on.
q_func: function
Model to use for computing the q function. It should accept the
following named arguments:
input_channel: int
number of channel of input.
num_actions: int
number of actions
optimizer_spec: OptimizerSpec
Specifying the constructor and kwargs, as well as learning rate schedule
for the optimizer
exploration: Schedule (defined in utils.schedule)
schedule for probability of chosing random action.
stopping_criterion: (env) -> bool
should return true when it's ok for the RL algorithm to stop.
takes in env and the number of steps executed so far.
replay_buffer_size: int
How many memories to store in the replay buffer.
batch_size: int
How many transitions to sample each time experience is replayed.
gamma: float
Discount Factor
learning_starts: int
After how many environment steps to start replaying experiences
learning_freq: int
How many steps of environment to take between every experience replay
frame_history_len: int
How many past frames to include as input to the model.
target_update_freq: int
How many experience replay rounds (not steps!) to perform between
each update to the target Q network
"""
assert type(env.observation_space) == gym.spaces.Box
assert type(env.action_space) == gym.spaces.Discrete
Statistic['parameters'] = {
'replay_buffer_size': replay_buffer_size,
'batch_size': batch_size,
'gamma': gamma,
'frame_history_len': frame_history_len,
'learning_starts': learning_starts,
'learning_freq': learning_freq,
'target_update_freq': target_update_freq,
'name': env.env.unwrapped.spec.id
}
###############
# BUILD MODEL #
###############
if len(env.observation_space.shape) == 1:
# This means we are running on low-dimensional observations (e.g. RAM)
input_arg = env.observation_space.shape[0]
else:
img_h, img_w, img_c = env.observation_space.shape
input_arg = frame_history_len * img_c
num_actions = env.action_space.n
# Construct an epilson greedy policy with given exploration schedule
def select_epilson_greedy_action(model, obs, t):
sample = random.random()
eps_threshold = exploration.value(t)
if sample > eps_threshold:
obs = torch.from_numpy(obs).type(dtype).unsqueeze(0) / 255.0
# Use volatile = True if variable is only used in inference mode, i.e. don’t save the history
with torch.no_grad():
return model(Variable(obs)).data.max(1)[1].cpu()
else:
return torch.IntTensor([[random.randrange(num_actions)]])
# Initialize target q function and q function, i.e. build the model.
######
Q = q_func(input_arg, num_actions).type(dtype)
target_Q = q_func(input_arg, num_actions).type(dtype)
if USE_CUDA:
Q = Q.cuda()
target_Q = target_Q.cuda()
######
# Construct Q network optimizer function
optimizer = optimizer_spec.constructor(Q.parameters(),
**optimizer_spec.kwargs)
# Construct the replay buffer
replay_buffer = ReplayBuffer(replay_buffer_size, frame_history_len)
###############
# RUN ENV #
###############
num_param_updates = 0
mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
last_obs = env.reset()
LOG_EVERY_N_STEPS = 10000
filename = 'statistics.pkl'
# Google Drive
try:
import google.colab
IN_COLAB = True
except:
IN_COLAB = False
if IN_COLAB:
run_in_colab_message()
try:
from google.colab import auth
import logging
from pydrive.auth import GoogleAuth
from pydrive.drive import GoogleDrive
from oauth2client.client import GoogleCredentials
logging.getLogger('googleapicliet.discovery_cache').setLevel(
logging.ERROR)
auth.authenticate_user()
gauth = GoogleAuth()
gauth.credentials = GoogleCredentials.get_application_default()
drive = GoogleDrive(gauth)
except:
pass
iter_time = time()
for t in count():
### 1. Check stopping criterion
if stopping_criterion is not None and stopping_criterion(env):
break
### 2. Step the env and store the transition
# At this point, "last_obs" contains the latest observation that was
# recorded from the simulator. Here, your code needs to store this
# observation and its outcome (reward, next observation, etc.) into
# the replay buffer while stepping the simulator forward one step.
# At the end of this block of code, the simulator should have been
# advanced one step, and the replay buffer should contain one more
# transition.
# Specifically, last_obs must point to the new latest observation.
# Useful functions you'll need to call:
# obs, reward, done, info = env.step(action)
# this steps the environment forward one step
# obs = env.reset()
# this resets the environment if you reached an episode boundary.
# Don't forget to call env.reset() to get a new observation if done
# is true!!
# Note that you cannot use "last_obs" directly as input
# into your network, since it needs to be processed to include context
# from previous frames. You should check out the replay buffer
# implementation in dqn_utils.py to see what functionality the replay
# buffer exposes. The replay buffer has a function called
# encode_recent_observation that will take the latest observation
# that you pushed into the buffer and compute the corresponding
# input that should be given to a Q network by appending some
# previous frames.
# Don't forget to include epsilon greedy exploration!
# And remember that the first time you enter this loop, the model
# may not yet have been initialized (but of course, the first step
# might as well be random, since you haven't trained your net...)
#####
idx = replay_buffer.store_frame(last_obs)
enc_obs = replay_buffer.encode_recent_observation()
if t > learning_starts:
action = select_epilson_greedy_action(Q, enc_obs, t)
else:
action = torch.IntTensor([[random.randrange(num_actions)]])
obs, reward, done, info = env.step(action)
if done:
obs = env.reset()
replay_buffer.store_effect(idx, action, reward, done)
last_obs = obs
#####
# at this point, the environment should have been advanced one step (and
# reset if done was true), and last_obs should point to the new latest
# observation
### 3. Perform experience replay and train the network.
# Note that this is only done if the replay buffer contains enough samples
# for us to learn something useful -- until then, the model will not be
# initialized and random actions should be taken
if (t > learning_starts and t % learning_freq == 0
and replay_buffer.can_sample(batch_size)):
# Here, you should perform training. Training consists of four steps:
# 3.a: use the replay buffer to sample a batch of transitions (see the
# replay buffer code for function definition, each batch that you sample
# should consist of current observations, current actions, rewards,
# next observations, and done indicator).
# Note: Move the variables to the GPU if avialable
# 3.b: fill in your own code to compute the Bellman error. This requires
# evaluating the current and next Q-values and constructing the corresponding error.
# Note: don't forget to clip the error between [-1,1], multiply is by -1 (since pytorch minimizes) and
# maskout post terminal status Q-values (see ReplayBuffer code).
# 3.c: train the model. To do this, use the bellman error you calculated perviously.
# Pytorch will differentiate this error for you, to backward the error use the following API:
# current.backward(d_error.data.unsqueeze(1))
# Where "current" is the variable holding current Q Values and d_error is the clipped bellman error.
# Your code should produce one scalar-valued tensor.
# Note: don't forget to call optimizer.zero_grad() before the backward call and
# optimizer.step() after the backward call.
# 3.d: periodically update the target network by loading the current Q network weights into the
# target_Q network. see state_dict() and load_state_dict() methods.
# you should update every target_update_freq steps, and you may find the
# variable num_param_updates useful for this (it was initialized to 0)
#####
#3.a
obs_batch, act_batch, rew_batch, next_obs_batch, done_mask = replay_buffer.sample(
batch_size)
obs_batch = Variable(torch.from_numpy(obs_batch).type(dtype) /
255.,
requires_grad=True)
act_batch = Variable(torch.from_numpy(act_batch).type(torch.int64))
rew_batch = Variable(torch.from_numpy(rew_batch).type(dtype),
requires_grad=True)
next_obs_batch = Variable(
torch.from_numpy(next_obs_batch).type(dtype) / 255.,
requires_grad=True)
done_mask = Variable(torch.from_numpy(done_mask).type(torch.int64))
if USE_CUDA:
obs_batch = obs_batch.cuda()
act_batch = act_batch.cuda()
rew_batch = rew_batch.cuda()
next_obs_batch = next_obs_batch.cuda()
done_mask = done_mask.cuda()
# Q network
val = Q(obs_batch).gather(dim=1, index=act_batch.unsqueeze(1))
# Q target network
with torch.no_grad():
tar_val_t = target_Q(next_obs_batch).max(1)[0]
tar_val = torch.addcmul(rew_batch, gamma,
1 - done_mask.type(dtype), tar_val_t)
# 3.b error calculate
d_error = (tar_val - val.squeeze()).clamp_(-1, 1) * -1.
# d_error = torch.pow((tar_val - val.squeeze()).clamp_(-1, 1), 2) * -1.
# 3.c train Q network
optimizer.zero_grad()
val.backward(d_error.data.unsqueeze(1))
optimizer.step()
# 3.d update target network
num_param_updates += 1
if num_param_updates % target_update_freq == 0:
target_Q.load_state_dict(Q.state_dict())
#####
### 4. Log progress and keep track of statistics
episode_rewards = get_wrapper_by_name(env,
"Monitor").get_episode_rewards()
if len(episode_rewards) > 0:
mean_episode_reward = np.mean(episode_rewards[-100:])
if len(episode_rewards) > 100:
best_mean_episode_reward = max(best_mean_episode_reward,
mean_episode_reward)
Statistic["mean_episode_rewards"].append(mean_episode_reward)
Statistic["best_mean_episode_rewards"].append(best_mean_episode_reward)
if t % LOG_EVERY_N_STEPS == 0 and t > learning_starts:
print("Timestep %d" % (t, ))
print(f"Iteration time:{time()-iter_time:.2f}")
iter_time = time()
print("mean reward (100 episodes) %f" % mean_episode_reward)
print("best mean reward %f" % best_mean_episode_reward)
print("episodes %d" % len(episode_rewards))
print("exploration %f" % exploration.value(t))
sys.stdout.flush()
# Dump statistics to pickle
filename = f"{t}" + 'statistics.pkl' if IN_COLAB else 'statistics.pkl'
with open(filename, 'wb') as f:
pickle.dump(Statistic, f)
print("Saved to %s" % filename)
if IN_COLAB and t % (LOG_EVERY_N_STEPS * 10) == 0:
try:
stat_pkl = drive.CreateFile()
stat_pkl.SetContentFile(filename)
stat_pkl.Upload()
print("Uploaded to drive")
except Exception:
print("Exception during upload to drive")
