def fit_value(self, states):
"""Updates critic parameters.
Args:
states: Batch of states.
Returns:
Dictionary with information to track.
"""
actions, log_probs = self.actor(states,
sample=True,
with_log_probs=True)
q1, q2 = self.critic(states, actions)
q = tf.minimum(q1, q2) - self.alpha * log_probs
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(self.value.trainable_variables)
v = self.value(states)
value_loss = tf.losses.mean_squared_error(q, v)
grads = tape.gradient(value_loss, self.value.trainable_variables)
self.value_optimizer.apply_gradients(
zip(grads, self.value.trainable_variables))
if self.value_optimizer.iterations % self.target_update_period == 0:
critic.soft_update(self.value, self.value_target, tau=self.tau)
return {'v': tf.reduce_mean(v), 'value_loss': value_loss}
def fit_critic(self, states, actions, next_states, rewards, discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
next_actions = self.actor(next_states, sample=True)
next_q1, next_q2 = self.critic_target(next_states, next_actions)
target_q = rewards + self.discount * discounts * tf.minimum(
next_q1, next_q2)
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(self.critic.trainable_variables)
q1, q2 = self.critic(states, actions)
policy_actions = self.actor(states, sample=True)
q_pi1, q_pi2 = self.critic(states, policy_actions)
def discriminator_loss(real_output, fake_output):
diff = target_q - real_output
def my_exp(x):
return 1.0 + x + 0.5 * tf.square(x)
alpha = 10.0
real_loss = tf.reduce_mean(my_exp(diff / alpha) * alpha)
total_loss = real_loss + tf.reduce_mean(fake_output)
return total_loss
critic_loss1 = discriminator_loss(q1, q_pi1)
critic_loss2 = discriminator_loss(q2, q_pi2)
critic_loss = (critic_loss1 + critic_loss2)
critic_grads = tape.gradient(critic_loss,
self.critic.trainable_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grads, self.critic.trainable_variables))
critic.soft_update(self.critic, self.critic_target, tau=self.tau)
return {
'q1': tf.reduce_mean(q1),
'q2': tf.reduce_mean(q2),
'critic_loss': critic_loss,
'q_pi1': tf.reduce_mean(q_pi1),
'q_pi2': tf.reduce_mean(q_pi2)
}
def __init__(self,
observation_spec,
action_spec,
actor_lr=3e-4,
critic_lr=3e-4,
alpha_lr=3e-4,
discount=0.99,
tau=0.005,
target_update_period=1,
target_entropy=0.0,
use_soft_critic=False):
"""Creates networks.
Args:
observation_spec: environment observation spec.
action_spec: Action spec.
actor_lr: Actor learning rate.
critic_lr: Critic learning rate.
alpha_lr: Temperature learning rate.
discount: MDP discount.
tau: Soft target update parameter.
target_update_period: Target network update period.
target_entropy: Target entropy.
use_soft_critic: Whether to use soft critic representation.
"""
assert len(observation_spec.shape) == 1
state_dim = observation_spec.shape[0]
self.actor = policies.DiagGuassianPolicy(state_dim, action_spec)
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr)
self.log_alpha = tf.Variable(tf.math.log(0.1), trainable=True)
self.alpha_optimizer = tf.keras.optimizers.Adam(learning_rate=alpha_lr)
self.target_entropy = target_entropy
self.discount = discount
self.tau = tau
self.target_update_period = target_update_period
self.value = critic.CriticNet(state_dim)
self.value_target = critic.CriticNet(state_dim)
critic.soft_update(self.value, self.value_target, tau=1.0)
self.value_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
if use_soft_critic:
self.critic = critic.SoftCritic(state_dim, action_spec)
else:
action_dim = action_spec.shape[0]
self.critic = critic.Critic(state_dim, action_dim)
self.critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
def fit(self, states, actions, next_states, rewards, discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
next_v = self.value_target(next_states)
target_q = rewards + self.discount * discounts * next_v
all_vars = (list(self.actor.trainable_variables) +
list(self.value.trainable_variables))
with tf.GradientTape(watch_accessed_variables=False,
persistent=True) as tape:
tape.watch(all_vars)
actor_log_probs = self.actor.log_probs(states, actions)
q = self.value(states) + self.alpha * actor_log_probs
adv = tf.stop_gradient(target_q - q)
actor_loss = -tf.reduce_mean(actor_log_probs * adv)
critic_loss = tf.losses.mean_squared_error(target_q, q)
actor_grads = tape.gradient(actor_loss, self.actor.trainable_variables)
critic_grads = tape.gradient(critic_loss,
self.value.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_grads, self.actor.trainable_variables))
self.critic_optimizer.apply_gradients(
zip(critic_grads, self.value.trainable_variables))
del tape
if self.critic_optimizer.iterations % self.target_update_period == 0:
critic.soft_update(self.value, self.value_target, tau=self.tau)
return {
'q': tf.reduce_mean(q),
'critic_loss': critic_loss,
'actor_log_probs': tf.reduce_mean(actor_log_probs),
'adv': tf.reduce_mean(adv)
}
def __init__(self,
observation_spec,
action_spec,
actor_lr=1e-4,
critic_lr=3e-4,
alpha_lr=1e-4,
discount=0.99,
tau=0.005,
target_entropy=0.0):
"""Creates networks.
Args:
observation_spec: environment observation spec.
action_spec: Action spec.
actor_lr: Actor learning rate.
critic_lr: Critic learning rate.
alpha_lr: Temperature learning rate.
discount: MDP discount.
tau: Soft target update parameter.
target_entropy: Target entropy.
"""
assert len(observation_spec.shape) == 1
state_dim = observation_spec.shape[0]
beta_1 = 0.0
self.actor = policies.DiagGuassianPolicy(state_dim, action_spec)
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr,
beta_1=beta_1)
self.log_alpha = tf.Variable(tf.math.log(0.1), trainable=True)
self.alpha_optimizer = tf.keras.optimizers.Adam(learning_rate=alpha_lr,
beta_1=beta_1)
self.target_entropy = target_entropy
self.discount = discount
self.tau = tau
action_dim = action_spec.shape[0]
self.critic = critic.Critic(state_dim, action_dim)
self.critic_target = critic.Critic(state_dim, action_dim)
critic.soft_update(self.critic, self.critic_target, tau=1.0)
self.critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr, beta_1=beta_1)
def fit_critic(self, states, actions, next_states, rewards, discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
next_actions = self.actor(next_states, sample=True)
bc_log_probs = self.bc.policy.log_probs(next_states, next_actions)
next_target_q1, next_target_q2 = self.critic_target(
next_states, next_actions)
target_q = rewards + self.discount * discounts * (tf.minimum(
next_target_q1, next_target_q2) + self.alpha * bc_log_probs)
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(self.critic.trainable_variables)
q1, q2 = self.critic(states, actions)
critic_loss = (tf.losses.mean_squared_error(target_q, q1) +
tf.losses.mean_squared_error(target_q, q2))
critic_grads = tape.gradient(critic_loss,
self.critic.trainable_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grads, self.critic.trainable_variables))
critic.soft_update(self.critic, self.critic_target, tau=self.tau)
return {
'q1': tf.reduce_mean(q1),
'q2': tf.reduce_mean(q2),
'critic_loss': critic_loss
}
def __init__(self,
observation_spec,
action_spec,
actor_lr=3e-4,
critic_lr=3e-4,
alpha_lr=3e-4,
discount=0.99,
tau=0.005,
target_entropy=0.0,
f_reg=1.0,
reward_bonus=5.0,
num_augmentations=1,
env_name='',
batch_size=256):
"""Creates networks.
Args:
observation_spec: environment observation spec.
action_spec: Action spec.
actor_lr: Actor learning rate.
critic_lr: Critic learning rate.
alpha_lr: Temperature learning rate.
discount: MDP discount.
tau: Soft target update parameter.
target_entropy: Target entropy.
f_reg: Critic regularization weight.
reward_bonus: Bonus added to the rewards.
num_augmentations: Number of random crops
env_name: Env name
batch_size: batch size
"""
del num_augmentations, env_name
assert len(observation_spec.shape) == 1
state_dim = observation_spec.shape[0]
self.batch_size = batch_size
hidden_dims = (256, 256, 256)
self.actor = policies.DiagGuassianPolicy(state_dim,
action_spec,
hidden_dims=hidden_dims)
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr)
self.log_alpha = tf.Variable(tf.math.log(1.0), trainable=True)
self.alpha_optimizer = tf.keras.optimizers.Adam(learning_rate=alpha_lr)
self.target_entropy = target_entropy
self.discount = discount
self.tau = tau
self.bc = behavioral_cloning.BehavioralCloning(observation_spec,
action_spec,
mixture=True)
action_dim = action_spec.shape[0]
self.critic = critic.Critic(state_dim,
action_dim,
hidden_dims=hidden_dims)
self.critic_target = critic.Critic(state_dim,
action_dim,
hidden_dims=hidden_dims)
critic.soft_update(self.critic, self.critic_target, tau=1.0)
self.critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
self.f_reg = f_reg
self.reward_bonus = reward_bonus
self.model_dict = {
'critic': self.critic,
'actor': self.actor,
'critic_target': self.critic_target,
'actor_optimizer': self.actor_optimizer,
'critic_optimizer': self.critic_optimizer,
'alpha_optimizer': self.alpha_optimizer
}
def fit_critic(self, states, actions, next_states, rewards, discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
next_actions = self.actor(next_states, sample=True)
policy_actions = self.actor(states, sample=True)
next_target_q1, next_target_q2 = self.dist_critic(next_states,
next_actions,
target=True)
target_q = rewards + self.discount * discounts * tf.minimum(
next_target_q1, next_target_q2)
critic_variables = self.critic.trainable_variables
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(critic_variables)
q1, q2 = self.dist_critic(states, actions, stop_gradient=True)
with tf.GradientTape(watch_accessed_variables=False,
persistent=True) as tape2:
tape2.watch([policy_actions])
q1_reg, q2_reg = self.critic(states, policy_actions)
q1_grads = tape2.gradient(q1_reg, policy_actions)
q2_grads = tape2.gradient(q2_reg, policy_actions)
q1_grad_norm = tf.reduce_sum(tf.square(q1_grads), axis=-1)
q2_grad_norm = tf.reduce_sum(tf.square(q2_grads), axis=-1)
del tape2
q_reg = tf.reduce_mean(q1_grad_norm + q2_grad_norm)
critic_loss = (tf.losses.mean_squared_error(target_q, q1) +
tf.losses.mean_squared_error(target_q, q2) +
self.f_reg * q_reg)
critic_grads = tape.gradient(critic_loss, critic_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grads, critic_variables))
critic.soft_update(self.critic, self.critic_target, tau=self.tau)
return {
'q1': tf.reduce_mean(q1),
'q2': tf.reduce_mean(q2),
'critic_loss': critic_loss,
'q1_grad': tf.reduce_mean(q1_grad_norm),
'q2_grad': tf.reduce_mean(q2_grad_norm)
}
def __init__(self,
observation_spec,
action_spec,
actor_lr=3e-4,
critic_lr=3e-4,
alpha_lr=3e-4,
discount=0.99,
tau=0.005,
target_entropy=0.0,
f_reg=1.0,
reward_bonus=5.0,
num_augmentations=1,
rep_learn_keywords='outer',
env_name='',
batch_size=256,
n_quantiles=5,
temp=0.1,
num_training_levels=200,
latent_dim=256,
n_levels_nce=5,
popart_norm_beta=0.1):
"""Creates networks.
Args:
observation_spec: environment observation spec.
action_spec: Action spec.
actor_lr: Actor learning rate.
critic_lr: Critic learning rate.
alpha_lr: Temperature learning rate.
discount: MDP discount.
tau: Soft target update parameter.
target_entropy: Target entropy.
f_reg: Critic regularization weight.
reward_bonus: Bonus added to the rewards.
num_augmentations: Number of DrQ augmentations (crops)
rep_learn_keywords: Representation learning loss to add (see below)
env_name: Env name
batch_size: Batch size
n_quantiles: Number of GVF quantiles
temp: Temperature of NCE softmax
num_training_levels: Number of training MDPs (Procgen=200)
latent_dim: Latent dimensions of auxiliary MLPs
n_levels_nce: Number of MDPs to use contrastive loss on
popart_norm_beta: PopArt normalization constant
For `rep_learn_keywords`, pick from:
stop_grad_FQI: whether to stop_grad TD/FQI critic updates?
linear_Q: use a linear critic?
successor_features: uses ||SF|| as cumulant
gvf_termination: uses +1 if done else 0 as cumulant
gvf_action_count: uses state-cond. action counts as cumulant
nce: uses the multi-class dot-product InfoNCE objective
cce: uses MoCo Categorical CrossEntropy objective
energy: uses SimCLR + pairwise GVF distance (not fully tested)
If no cumulant is specified, the reward will be taken as default one.
"""
del actor_lr, critic_lr, alpha_lr, target_entropy
self.action_spec = action_spec
self.num_augmentations = num_augmentations
self.rep_learn_keywords = rep_learn_keywords.split('__')
self.batch_size = batch_size
self.env_name = env_name
self.stop_grad_fqi = 'stop_grad_FQI' in self.rep_learn_keywords
critic_kwargs = {'hidden_dims': (1024, 1024)}
self.latent_dim = latent_dim
self.n_levels_nce = n_levels_nce
hidden_dims = hidden_dims_per_level = (self.latent_dim,
self.latent_dim)
self.num_training_levels = int(num_training_levels)
self.n_quantiles = n_quantiles
self.temp = temp
# Make 2 sets of weights:
# - Critic
# - Critic (target)
# Optionally, make a 3rd set for per-level critics
if observation_spec.shape == (64, 64, 3):
# IMPALA for Procgen
def conv_stack():
return make_impala_cnn_network(depths=[16, 32, 32],
use_batch_norm=False,
dropout_rate=0.)
state_dim = 256
else:
# Reduced architecture for DMC
def conv_stack():
return ConvStack(observation_spec.shape)
state_dim = 50
conv_stack_critic = conv_stack()
conv_target_stack_critic = conv_stack()
if observation_spec.shape == (64, 64, 3):
conv_stack_critic.output_size = state_dim
conv_target_stack_critic.output_size = state_dim
critic_kwargs['encoder'] = ImageEncoder(conv_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=True)
# Note: the target critic does not share any weights.
critic_kwargs['encoder_target'] = ImageEncoder(
conv_target_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=True)
conv_stack_critic_per_level = conv_stack()
conv_target_stack_critic_per_level = conv_stack()
if observation_spec.shape == (64, 64, 3):
conv_stack_critic_per_level.output_size = state_dim
conv_target_stack_critic_per_level.output_size = state_dim
self.encoder_per_level = ImageEncoder(conv_stack_critic_per_level,
feature_dim=state_dim,
bprop_conv_stack=True)
self.encoder_per_level_target = ImageEncoder(
conv_target_stack_critic_per_level,
feature_dim=state_dim,
bprop_conv_stack=True)
criticCL.soft_update(self.encoder_per_level,
self.encoder_per_level_target,
tau=1.0)
if self.num_augmentations == 0:
dummy_state = tf.constant(
np.zeros([1] + list(observation_spec.shape)))
else: # account for padding of +4 everywhere and then cropping out 68
dummy_state = tf.constant(np.zeros(shape=[1, 68, 68, 3]))
dummy_enc = critic_kwargs['encoder'](dummy_state)
@tf.function
def init_models():
"""This function initializes all auxiliary networks (state and action encoders) with dummy input (Procgen-specific, 68x68x3, 15 actions).
"""
critic_kwargs['encoder'](dummy_state)
critic_kwargs['encoder_target'](dummy_state)
self.encoder_per_level(dummy_state)
self.encoder_per_level_target(dummy_state)
init_models()
action_dim = action_spec.maximum.item() + 1
self.action_dim = action_dim
self.discount = discount
self.tau = tau
self.reg = f_reg
self.reward_bonus = reward_bonus
self.critic = criticCL.Critic(state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder'],
discrete_actions=True,
linear='linear_Q'
in self.rep_learn_keywords)
self.critic_target = criticCL.Critic(
state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder_target'],
discrete_actions=True,
linear='linear_Q' in self.rep_learn_keywords)
self.critic_optimizer = tf.keras.optimizers.Adam(learning_rate=3e-4)
self.task_critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=3e-4)
self.br_optimizer = tf.keras.optimizers.Adam(learning_rate=3e-4)
if 'cce' in self.rep_learn_keywords:
self.classifier = tf.keras.Sequential([
tf.keras.layers.Dense(self.latent_dim, use_bias=True),
tf.keras.layers.ReLU(),
tf.keras.layers.Dense(self.n_quantiles, use_bias=True)
],
name='classifier')
elif 'nce' in self.rep_learn_keywords:
self.embedding = tf.keras.Sequential([
tf.keras.layers.Dense(self.latent_dim, use_bias=True),
tf.keras.layers.ReLU(),
tf.keras.layers.Dense(self.latent_dim, use_bias=True)
],
name='embedding')
# This snipet initializes all auxiliary networks (state and action encoders)
# with dummy input (Procgen-specific, 68x68x3, 15 actions).
dummy_state = tf.zeros((1, 68, 68, 3), dtype=tf.float32)
phi_s = self.critic.encoder(dummy_state)
phi_a = tf.eye(action_dim, dtype=tf.float32)
if 'linear_Q' in self.rep_learn_keywords:
_ = self.critic.critic1.state_encoder(phi_s)
_ = self.critic.critic2.state_encoder(phi_s)
_ = self.critic.critic1.action_encoder(phi_a)
_ = self.critic.critic2.action_encoder(phi_a)
_ = self.critic_target.critic1.state_encoder(phi_s)
_ = self.critic_target.critic2.state_encoder(phi_s)
_ = self.critic_target.critic1.action_encoder(phi_a)
_ = self.critic_target.critic2.action_encoder(phi_a)
if 'cce' in self.rep_learn_keywords:
self.classifier(phi_s)
elif 'nce' in self.rep_learn_keywords:
self.embedding(phi_s)
self.target_critic_to_use = self.critic_target
self.critic_to_use = self.critic
criticCL.soft_update(self.critic, self.critic_target, tau=1.0)
self.cce = tf.keras.losses.SparseCategoricalCrossentropy(
reduction=tf.keras.losses.Reduction.NONE, from_logits=True)
self.bc = None
if 'successor_features' in self.rep_learn_keywords:
self.output_dim_level = self.latent_dim
elif 'gvf_termination' in self.rep_learn_keywords:
self.output_dim_level = 1
elif 'gvf_action_count' in self.rep_learn_keywords:
self.output_dim_level = action_dim
else:
self.output_dim_level = action_dim
self.task_critic_one = criticCL.Critic(
state_dim,
self.output_dim_level * self.num_training_levels,
hidden_dims=hidden_dims_per_level,
encoder=None, # critic_kwargs['encoder'],
discrete_actions=True,
cross_norm=False)
self.task_critic_target_one = criticCL.Critic(
state_dim,
self.output_dim_level * 200,
hidden_dims=hidden_dims_per_level,
encoder=None, # critic_kwargs['encoder'],
discrete_actions=True,
cross_norm=False)
self.task_critic_one(dummy_enc,
actions=None,
training=False,
return_features=False,
stop_grad_features=False)
self.task_critic_target_one(dummy_enc,
actions=None,
training=False,
return_features=False,
stop_grad_features=False)
criticCL.soft_update(self.task_critic_one,
self.task_critic_target_one,
tau=1.0)
# Normalization constant beta, set to best default value as per PopArt paper
self.reward_normalizer = popart.PopArt(
running_statistics.EMAMeanStd(popart_norm_beta))
self.reward_normalizer.init()
if 'CLIP' in self.rep_learn_keywords or 'clip' in self.rep_learn_keywords:
self.loss_temp = tf.Variable(tf.constant(0.0, dtype=tf.float32),
name='loss_temp',
trainable=True)
self.model_dict = {
'critic': self.critic,
'critic_target': self.critic_target,
'critic_optimizer': self.critic_optimizer,
'br_optimizer': self.br_optimizer
}
self.model_dict['encoder_perLevel'] = self.encoder_per_level
self.model_dict[
'encoder_perLevel_target'] = self.encoder_per_level_target
self.model_dict['task_critic'] = self.task_critic_one
self.model_dict['task_critic_target'] = self.task_critic_target_one
def fit_task_critics(self, mb_states, mb_actions, mb_next_states,
mb_next_actions, mb_rewards, mb_discounts, level_ids):
"""Updates per-level critic parameters.
Args:
mb_states: Batch of states.
mb_actions: Batch of actions.
mb_next_states: Batch of next states.
mb_next_actions: Batch of next actions from training policy.
mb_rewards: Batch of rewards.
mb_discounts: Batch of masks indicating the end of the episodes.
level_ids: Batch of level ids
Returns:
Dictionary with information to track.
"""
if 'popart' in self.rep_learn_keywords:
# The PopArt normalization normalizes the GVF's cumulant signal so that
# it's not affected by the difference in scales across MDPs.
mb_rewards = self.reward_normalizer.normalize_target(mb_rewards)
trainable_variables = self.encoder_per_level.trainable_variables + self.task_critic_one.trainable_variables
next_action_indices = tf.stack([
tf.range(tf.shape(mb_next_actions)[0],
dtype=tf.int32), level_ids * self.output_dim_level +
tf.cast(mb_next_actions, dtype=tf.int32)
],
axis=-1)
action_indices = tf.stack([
tf.range(tf.shape(mb_actions)[0],
dtype=tf.int32), level_ids * self.output_dim_level +
tf.cast(mb_actions, dtype=tf.int32)
],
axis=-1)
level_ids = tf.stack([
tf.range(tf.shape(mb_next_actions)[0], dtype=tf.int32),
tf.cast(level_ids, dtype=tf.int32)
],
axis=-1)
next_states = [self.encoder_per_level_target(mb_next_states[0])]
next_q1, next_q2 = self.task_critic_target_one(next_states[0],
actions=None)
# Learn d-dimensional successor features
if 'successor_features' in self.rep_learn_keywords:
target_q = tf.concat(
[next_states[0]] * 200, 1) + self.discount * tf.expand_dims(
mb_discounts, 1) * tf.minimum(next_q1, next_q2)
# Learn discounted episode termination
elif 'gvf_termination' in self.rep_learn_keywords:
target_q = tf.expand_dims(
mb_discounts, 1) + self.discount * tf.expand_dims(
mb_discounts, 1) * tf.minimum(next_q1, next_q2)
# Learn discounted future action counts
elif 'gvf_action_count' in self.rep_learn_keywords:
target_q = tf.concat(
[tf.one_hot(mb_actions, depth=self.action_dim)] * 200,
1) + self.discount * tf.expand_dims(
mb_discounts, 1) * tf.minimum(next_q1, next_q2)
else:
target_q = tf.expand_dims(
mb_rewards, 1) + self.discount * tf.expand_dims(
mb_discounts, 1) * tf.minimum(next_q1, next_q2)
if ('successor_features' in self.rep_learn_keywords
or 'gvf_termination' in self.rep_learn_keywords
or 'gvf_action_count' in self.rep_learn_keywords):
target_q = tf.reshape(target_q, (-1, 200, self.output_dim_level))
target_q = tf.gather_nd(target_q, indices=level_ids)
else:
target_q = tf.gather_nd(target_q, indices=next_action_indices)
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(trainable_variables)
states = [self.encoder_per_level(mb_states[0])]
q1_all, q2_all = self.task_critic_one(states[0], actions=None)
q = tf.minimum(q1_all, q2_all)
if ('successor_features' in self.rep_learn_keywords
or 'gvf_termination' in self.rep_learn_keywords
or 'gvf_action_count' in self.rep_learn_keywords):
q1_all = tf.reshape(q1_all, (-1, 200, self.output_dim_level))
q2_all = tf.reshape(q2_all, (-1, 200, self.output_dim_level))
critic_loss = (
tf.losses.mean_squared_error(
target_q, tf.gather_nd(q1_all, indices=level_ids)) +
tf.losses.mean_squared_error(
target_q, tf.gather_nd(q2_all, indices=level_ids)))
else:
critic_loss = (tf.losses.mean_squared_error(
target_q, tf.gather_nd(q1_all, indices=action_indices)) +
tf.losses.mean_squared_error(
target_q,
tf.gather_nd(q2_all,
indices=action_indices)))
critic_grads = tape.gradient(critic_loss, trainable_variables)
self.task_critic_optimizer.apply_gradients(
zip(critic_grads, trainable_variables))
criticCL.soft_update(self.encoder_per_level,
self.encoder_per_level_target,
tau=self.tau)
criticCL.soft_update(self.task_critic_one,
self.task_critic_target_one,
tau=self.tau)
gn = tf.reduce_mean(
[tf.linalg.norm(v) for v in critic_grads if v is not None])
return {
'avg_level_critic_loss': tf.reduce_mean(critic_loss),
'avg_q': tf.reduce_mean(q),
'level_critic_grad_norm': gn
}
def fit_critic(self, states, actions, next_states, next_actions, rewards,
discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
next_actions: Batch of next actions from training policy.
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
action_indices = tf.stack(
[tf.range(tf.shape(actions)[0], dtype=tf.int64), actions], axis=-1)
next_action_indices = tf.stack([
tf.range(tf.shape(next_actions)[0], dtype=tf.int64), next_actions
],
axis=-1)
if self.num_augmentations > 1:
target_q = 0.
for i in range(self.num_augmentations):
next_q1_i, next_q2_i = self.critic_target(
next_states[i],
actions=None,
stop_grad_features=self.stop_grad_fqi)
target_q_i = tf.expand_dims(
rewards, 1) + self.discount * tf.expand_dims(
discounts, 1) * tf.minimum(next_q1_i, next_q2_i)
target_q += target_q_i
target_q /= self.num_augmentations
elif self.num_augmentations == 1:
next_q1, next_q2 = self.critic_target(
next_states[0],
actions=None,
stop_grad_features=self.stop_grad_fqi)
target_q = tf.expand_dims(
rewards, 1) + self.discount * tf.expand_dims(
discounts, 1) * tf.minimum(next_q1, next_q2)
else:
next_q1, next_q2 = self.target_critic_to_use(
next_states,
actions=None,
stop_grad_features=self.stop_grad_fqi)
target_q = tf.expand_dims(
rewards, 1) + self.discount * tf.expand_dims(
discounts, 1) * tf.minimum(next_q1, next_q2)
target_q = tf.gather_nd(target_q, indices=next_action_indices)
trainable_variables = self.critic.trainable_variables
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(trainable_variables)
if self.num_augmentations > 1:
critic_loss = 0.
q1 = 0.
q2 = 0.
for i in range(self.num_augmentations):
q1_i, q2_i = self.critic_to_use(
states[i],
actions=None,
stop_grad_features=self.stop_grad_fqi)
critic_loss_i = (tf.losses.mean_squared_error(
target_q, tf.gather_nd(q1_i, indices=action_indices)) +
tf.losses.mean_squared_error(
target_q,
tf.gather_nd(q2_i,
indices=action_indices)))
q1 += q1_i
q2 += q2_i
critic_loss += critic_loss_i
q1 /= self.num_augmentations
q2 /= self.num_augmentations
critic_loss /= self.num_augmentations
elif self.num_augmentations == 1:
q1, q2 = self.critic_to_use(
states[0],
actions=None,
stop_grad_features=self.stop_grad_fqi)
q = tf.minimum(q1, q2)
critic_loss = (
tf.losses.mean_squared_error(
target_q, tf.gather_nd(q1, indices=action_indices)) +
tf.losses.mean_squared_error(
target_q, tf.gather_nd(q2, indices=action_indices)))
else:
q1, q2 = self.critic_to_use(
states,
actions=None,
stop_grad_features=self.stop_grad_fqi)
q = tf.minimum(q1, q2)
critic_loss = (
tf.losses.mean_squared_error(
target_q, tf.gather_nd(q1, indices=action_indices)) +
tf.losses.mean_squared_error(
target_q, tf.gather_nd(q2, indices=action_indices)))
# LSE from CQL
cql_logsumexp = tf.reduce_logsumexp(q, 1)
cql_loss = tf.reduce_mean(cql_logsumexp -
tf.gather_nd(q, indices=action_indices))
# Jointly optimize both losses
critic_loss = critic_loss + cql_loss
critic_grads = tape.gradient(critic_loss, trainable_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grads, trainable_variables))
criticCL.soft_update(self.critic, self.critic_target, tau=self.tau)
gn = tf.reduce_mean(
[tf.linalg.norm(v) for v in critic_grads if v is not None])
return {
'q1': tf.reduce_mean(q1),
'q2': tf.reduce_mean(q2),
'critic_loss': critic_loss,
'cql_loss': cql_loss,
'critic_grad_norm': gn
}
self.log_alpha = tf.Variable(tf.math.log(1.0), trainable=True)
self.log_cql_alpha = self.log_alpha
self.alpha_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr)
self.critic = critic.Critic(state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder'],
discrete_actions=True)
self.critic_target = critic.Critic(
state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder_target'],
discrete_actions=True)
critic.soft_update(self.critic, self.critic_target, tau=1.0)
self.critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
self.tau = tau
self.reg = reg
self.target_entropy = target_entropy
self.discount = discount
self.num_cql_actions = num_cql_actions
self.bc_pretraining_steps = bc_pretraining_steps
self.min_q_weight = min_q_weight
self.bc = None
self.model_dict = {
def __init__(self,
observation_spec,
action_spec,
embedding_dim=256,
hidden_dims=(256, 256),
sequence_length=2,
learning_rate=None,
discount=0.95,
target_update_period=1000,
num_augmentations=0,
rep_learn_keywords='outer',
batch_size=256):
"""Creates networks.
Args:
observation_spec: State spec.
action_spec: Action spec.
embedding_dim: Embedding size.
hidden_dims: List of hidden dimensions.
sequence_length: Expected length of sequences provided as input.
learning_rate: Learning rate.
discount: discount factor.
target_update_period: How frequently update target?
num_augmentations: Number of DrQ random crops.
rep_learn_keywords: Representation learning loss to add.
batch_size: batch size.
"""
super().__init__()
action_dim = action_spec.maximum.item() + 1
self.observation_spec = observation_spec
self.action_dim = action_dim
self.action_spec = action_spec
self.embedding_dim = embedding_dim
self.sequence_length = sequence_length
self.discount = discount
self.tau = 0.005
self.discount = 0.99
self.target_update_period = target_update_period
self.num_augmentations = num_augmentations
self.rep_learn_keywords = rep_learn_keywords.split('__')
self.batch_size = batch_size
critic_kwargs = {}
if observation_spec.shape == (64, 64, 3):
# IMPALA for Procgen
def conv_stack():
return make_impala_cnn_network(depths=[16, 32, 32],
use_batch_norm=False,
dropout_rate=0.)
state_dim = 256
else:
# Reduced architecture for DMC
def conv_stack():
return ConvStack(observation_spec.shape)
state_dim = 50
conv_stack_critic = conv_stack()
conv_target_stack_critic = conv_stack()
if observation_spec.shape == (64, 64, 3):
conv_stack_critic.output_size = state_dim
conv_target_stack_critic.output_size = state_dim
critic_kwargs['encoder'] = ImageEncoder(conv_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=True)
critic_kwargs['encoder_target'] = ImageEncoder(
conv_target_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=True)
self.embedder = tf_utils.EmbedNet(state_dim,
embedding_dim=self.embedding_dim,
hidden_dims=hidden_dims)
self.f_value = tf_utils.create_mlp(self.embedding_dim,
1,
hidden_dims=hidden_dims,
activation=tf.nn.swish)
self.f_value_target = tf_utils.create_mlp(self.embedding_dim,
1,
hidden_dims=hidden_dims,
activation=tf.nn.swish)
self.f_trans = tf_utils.create_mlp(self.embedding_dim +
self.embedding_dim,
self.embedding_dim,
hidden_dims=hidden_dims,
activation=tf.nn.swish)
self.f_out = tf_utils.create_mlp(self.embedding_dim +
self.embedding_dim,
2,
hidden_dims=hidden_dims,
activation=tf.nn.swish)
self.action_encoder = tf.keras.Sequential(
[
tf.keras.layers.Dense(
self.embedding_dim, use_bias=True
), # , kernel_regularizer=tf.keras.regularizers.l2(WEIGHT_DECAY)
tf.keras.layers.ReLU(),
tf.keras.layers.Dense(self.embedding_dim)
],
name='action_encoder')
if self.num_augmentations == 0:
dummy_state = tf.constant(
np.zeros(shape=[1] + list(observation_spec.shape)))
self.obs_spec = list(observation_spec.shape)
else: # account for padding of +4 everywhere and then cropping out 68
dummy_state = tf.constant(np.zeros(shape=[1, 68, 68, 3]))
self.obs_spec = [68, 68, 3]
@tf.function
def init_models():
critic_kwargs['encoder'](dummy_state)
critic_kwargs['encoder_target'](dummy_state)
self.action_encoder(
tf.cast(tf.one_hot([1], depth=action_dim), tf.float32))
init_models()
self.critic = critic.Critic(state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder'],
discrete_actions=True,
linear='linear_Q'
in self.rep_learn_keywords)
self.critic_target = critic.Critic(
state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder_target'],
discrete_actions=True,
linear='linear_Q' in self.rep_learn_keywords)
@tf.function
def init_models2():
dummy_state = tf.zeros((1, 68, 68, 3), dtype=tf.float32)
phi_s = self.critic.encoder(dummy_state)
phi_a = tf.eye(15, dtype=tf.float32)
if 'linear_Q' in self.rep_learn_keywords:
_ = self.critic.critic1.state_encoder(phi_s)
_ = self.critic.critic2.state_encoder(phi_s)
_ = self.critic.critic1.action_encoder(phi_a)
_ = self.critic.critic2.action_encoder(phi_a)
_ = self.critic_target.critic1.state_encoder(phi_s)
_ = self.critic_target.critic2.state_encoder(phi_s)
_ = self.critic_target.critic1.action_encoder(phi_a)
_ = self.critic_target.critic2.action_encoder(phi_a)
init_models2()
critic.soft_update(self.critic, self.critic_target, tau=1.0)
critic.soft_update(self.f_value, self.f_value_target, tau=1.0)
learning_rate = learning_rate or 1e-4
self.optimizer = tf.keras.optimizers.Adam(learning_rate=3e-4)
self.critic_optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
self.all_variables = (self.embedder.trainable_variables +
self.f_value.trainable_variables +
self.f_value_target.trainable_variables +
self.f_trans.trainable_variables +
self.f_out.trainable_variables +
self.critic.trainable_variables +
self.critic_target.trainable_variables)
self.model_dict = {
'action_encoder': self.action_encoder,
'f_out': self.f_out,
'f_trans': self.f_trans,
'f_value_target': self.f_value_target,
'f_value': self.f_value,
'embedder': self.embedder,
'critic': self.critic,
'critic_target': self.critic_target,
'critic_optimizer': self.critic_optimizer,
'optimizer': self.optimizer
}
def fit_embedding(self, states, actions, next_states, next_actions,
rewards, discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
next_actions: batch of next actions
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
states = tf.transpose(
tf.stack([states, next_states])[:, 0], (1, 0, 2, 3, 4))
batch_size = tf.shape(states)[0]
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(self.all_variables)
actions = tf.transpose(
tf.one_hot(tf.stack([actions, next_actions]),
depth=self.action_dim), (1, 0, 2))
actions = tf.reshape(
actions, [batch_size * self.sequence_length, self.action_dim])
actions = self.action_encoder(actions)
actions = tf.reshape(
actions,
[batch_size, self.sequence_length, self.embedding_dim])
all_states = tf.reshape(
states, [batch_size * self.sequence_length] + self.obs_spec)
all_features = self.critic.encoder(all_states)
all_embeddings = self.embedder(all_features, stop_gradient=False)
embeddings = tf.reshape(
all_embeddings,
[batch_size, self.sequence_length, self.embedding_dim])[:,
0, :]
all_pred_values = []
all_pred_rewards = []
all_pred_discounts = []
for idx in range(self.sequence_length):
pred_value = self.f_value(embeddings)[Ellipsis, 0]
pred_reward, pred_discount = tf.unstack(self.f_out(
tf.concat([embeddings, actions[:, idx, :]], -1)),
axis=-1)
pred_embeddings = embeddings + self.f_trans(
tf.concat([embeddings, actions[:, idx, :]], -1))
all_pred_values.append(pred_value)
all_pred_rewards.append(pred_reward)
all_pred_discounts.append(pred_discount)
embeddings = pred_embeddings
last_value = tf.stop_gradient(
self.f_value_target(embeddings)[Ellipsis,
0]) / (1 - self.discount)
all_true_values = []
# for idx in range(self.sequence_length - 1, -1, -1):
value = self.discount * discounts * last_value + rewards #[:, idx]
all_true_values.append(value)
last_value = value
all_true_values = all_true_values[::-1]
reward_error = tf.stack(all_pred_rewards, -1)[:, 0] - rewards
value_error = tf.stack(
all_pred_values,
-1) - (1 - self.discount) * tf.stack(all_true_values, -1)
reward_loss = tf.reduce_sum(tf.math.square(reward_error), -1)
value_loss = tf.reduce_sum(tf.math.square(value_error), -1)
loss = tf.reduce_mean(reward_loss + value_loss)
grads = tape.gradient(loss, self.all_variables)
self.optimizer.apply_gradients(zip(grads, self.all_variables))
if self.optimizer.iterations % self.target_update_period == 0:
critic.soft_update(self.f_value, self.f_value_target, tau=self.tau)
return {
'embed_loss': loss,
'reward_loss': tf.reduce_mean(reward_loss),
'value_loss': tf.reduce_mean(value_loss),
}
def fit_critic(self, states, actions, next_states, next_actions, rewards,
discounts):
"""Updates critic parameters.
Args:
states: Batch of states.
actions: Batch of actions.
next_states: Batch of next states.
next_actions: Batch of next actions from training policy.
rewards: Batch of rewards.
discounts: Batch of masks indicating the end of the episodes.
Returns:
Dictionary with information to track.
"""
next_q1, next_q2 = self.critic_target(next_states, next_actions)
target_q = rewards + self.discount * discounts * tf.minimum(
next_q1, next_q2)
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(self.critic.trainable_variables)
q1, q2 = self.critic(states, actions)
critic_loss = (tf.losses.mean_squared_error(target_q, q1) +
tf.losses.mean_squared_error(target_q, q2))
n_states = tf.repeat(states[tf.newaxis, :, :],
self.num_cql_actions, 0)
n_states = tf.reshape(n_states, [-1, n_states.get_shape()[-1]])
n_rand_actions = tf.random.uniform(
[tf.shape(n_states)[0],
actions.get_shape()[-1]], self.actor.action_spec.minimum,
self.actor.action_spec.maximum)
n_actions, n_log_probs = self.actor(n_states,
sample=True,
with_log_probs=True)
q1_rand, q2_rand = self.critic(n_states, n_rand_actions)
q1_curr_actions, q2_curr_actions = self.critic(n_states, n_actions)
log_u = -tf.reduce_mean(
tf.repeat((tf.math.log(2.0 * self.actor.action_scale) *
n_rand_actions.shape[-1])[tf.newaxis, :],
tf.shape(n_states)[0], 0), 1)
log_probs_all = tf.concat([n_log_probs, log_u], 0)
q1_all = tf.concat([q1_curr_actions, q1_rand], 0)
q2_all = tf.concat([q2_curr_actions, q2_rand], 0)
def get_qf_loss(q, log_probs):
q -= log_probs
q = tf.reshape(q, [-1, tf.shape(states)[0]])
return tf.math.reduce_logsumexp(q, axis=0)
min_qf1_loss = get_qf_loss(q1_all, log_probs_all)
min_qf2_loss = get_qf_loss(q2_all, log_probs_all)
cql_loss = tf.reduce_mean((min_qf1_loss - q1) +
(min_qf2_loss - q2))
critic_loss += self.min_q_weight * cql_loss
critic_grads = tape.gradient(critic_loss,
self.critic.trainable_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grads, self.critic.trainable_variables))
critic.soft_update(self.critic, self.critic_target, tau=self.tau)
return {
'q1': tf.reduce_mean(q1),
'q2': tf.reduce_mean(q2),
'critic_loss': critic_loss,
'cql_loss': cql_loss
}
def __init__(self,
observation_spec,
action_spec,
actor_lr=3e-4,
critic_lr=3e-4,
alpha_lr=3e-4,
discount=0.99,
tau=0.005,
target_entropy=0.0,
f_reg=1.0,
reward_bonus=5.0,
num_augmentations=1,
env_name='',
batch_size=256):
"""Creates networks.
Args:
observation_spec: environment observation spec.
action_spec: Action spec.
actor_lr: Actor learning rate.
critic_lr: Critic learning rate.
alpha_lr: Temperature learning rate.
discount: MDP discount.
tau: Soft target update parameter.
target_entropy: Target entropy.
f_reg: Critic regularization weight.
reward_bonus: Bonus added to the rewards.
num_augmentations: Number of DrQ augmentations (crops)
env_name: Env name
batch_size: Batch size
"""
self.num_augmentations = num_augmentations
self.discrete_actions = False if len(action_spec.shape) else True
self.batch_size = batch_size
actor_kwargs = {'hidden_dims': (1024, 1024)}
critic_kwargs = {'hidden_dims': (1024, 1024)}
# DRQ encoder params.
# https://github.com/denisyarats/drq/blob/master/config.yaml#L73
# Make 4 sets of weights:
# - BC
# - Actor
# - Critic
# - Critic (target)
if observation_spec.shape == (64, 64, 3):
# IMPALA for Procgen
def conv_stack():
return make_impala_cnn_network(depths=[16, 32, 32],
use_batch_norm=False,
dropout_rate=0.)
state_dim = 256
else:
# Reduced architecture for DMC
def conv_stack():
return ConvStack(observation_spec.shape)
state_dim = 50
conv_stack_bc = conv_stack()
conv_stack_actor = conv_stack()
conv_stack_critic = conv_stack()
conv_target_stack_critic = conv_stack()
if observation_spec.shape == (64, 64, 3):
conv_stack_bc.output_size = state_dim
conv_stack_actor.output_size = state_dim
conv_stack_critic.output_size = state_dim
conv_target_stack_critic.output_size = state_dim
# Combine and stop_grad some of the above conv stacks
actor_kwargs['encoder_bc'] = ImageEncoder(conv_stack_bc,
feature_dim=state_dim,
bprop_conv_stack=True)
actor_kwargs['encoder'] = ImageEncoder(conv_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=False)
critic_kwargs['encoder'] = ImageEncoder(conv_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=True)
# Note: the target critic does not share any weights.
critic_kwargs['encoder_target'] = ImageEncoder(
conv_target_stack_critic,
feature_dim=state_dim,
bprop_conv_stack=True)
if self.num_augmentations == 0:
dummy_state = tf.constant(
np.zeros(shape=[1] + list(observation_spec.shape)))
else: # account for padding of +4 everywhere and then cropping out 68
dummy_state = tf.constant(np.zeros(shape=[1, 68, 68, 3]))
@tf.function
def init_models():
actor_kwargs['encoder_bc'](dummy_state)
actor_kwargs['encoder'](dummy_state)
critic_kwargs['encoder'](dummy_state)
critic_kwargs['encoder_target'](dummy_state)
init_models()
if self.discrete_actions:
hidden_dims = ()
self.actor = policies.CategoricalPolicy(
state_dim,
action_spec,
hidden_dims=hidden_dims,
encoder=actor_kwargs['encoder'])
action_dim = action_spec.maximum.item() + 1
else:
hidden_dims = (256, 256, 256)
self.actor = policies.DiagGuassianPolicy(
state_dim,
action_spec,
hidden_dims=hidden_dims,
encoder=actor_kwargs['encoder'])
action_dim = action_spec.shape[0]
self.action_dim = action_dim
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr)
self.log_alpha = tf.Variable(tf.math.log(1.0), trainable=True)
self.alpha_optimizer = tf.keras.optimizers.Adam(learning_rate=alpha_lr)
self.target_entropy = target_entropy
self.discount = discount
self.tau = tau
self.bc = behavioral_cloning.BehavioralCloning(
observation_spec,
action_spec,
mixture=True,
encoder=actor_kwargs['encoder_bc'],
num_augmentations=self.num_augmentations,
env_name=env_name,
batch_size=batch_size)
self.critic = critic.Critic(state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder'])
self.critic_target = critic.Critic(
state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder_target'])
critic.soft_update(self.critic, self.critic_target, tau=1.0)
self.critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
self.f_reg = f_reg
self.reward_bonus = reward_bonus
self.model_dict = {
'critic': self.critic,
'critic_target': self.critic_target,
'actor': self.actor,
'bc': self.bc,
'critic_optimizer': self.critic_optimizer,
'alpha_optimizer': self.alpha_optimizer,
'actor_optimizer': self.actor_optimizer
}
def __init__(
self,
observation_spec,
action_spec,
embedding_dim = 256,
num_distributions=None,
hidden_dims = (256, 256),
sequence_length = 2,
learning_rate=None,
latent_dim = 256,
reward_weight = 1.0,
forward_weight
= 1.0, # Predict last state given prev actions/states.
inverse_weight = 1.0, # Predict last action given states.
state_prediction_mode = 'energy',
num_augmentations = 0,
rep_learn_keywords = 'outer',
batch_size = 256):
"""Creates networks.
Args:
observation_spec: State spec.
action_spec: Action spec.
embedding_dim: Embedding size.
num_distributions: Number of categorical distributions for discrete
embedding.
hidden_dims: List of hidden dimensions.
sequence_length: Expected length of sequences provided as input
learning_rate: Learning rate.
latent_dim: Dimension of the latent variable.
reward_weight: Weight on the reward loss.
forward_weight: Weight on the forward loss.
inverse_weight: Weight on the inverse loss.
state_prediction_mode: One of ['latent', 'energy'].
num_augmentations: Num of random crops
rep_learn_keywords: Representation learning loss to add.
batch_size: Batch size
"""
super().__init__()
action_dim = action_spec.maximum.item() + 1
self.observation_spec = observation_spec
self.action_dim = action_dim
self.action_spec = action_spec
self.embedding_dim = embedding_dim
self.num_distributions = num_distributions
self.sequence_length = sequence_length
self.latent_dim = latent_dim
self.reward_weight = reward_weight
self.forward_weight = forward_weight
self.inverse_weight = inverse_weight
self.state_prediction_mode = state_prediction_mode
self.num_augmentations = num_augmentations
self.rep_learn_keywords = rep_learn_keywords.split('__')
self.batch_size = batch_size
self.tau = 0.005
self.discount = 0.99
critic_kwargs = {}
if observation_spec.shape == (64, 64, 3):
# IMPALA for Procgen
def conv_stack():
return make_impala_cnn_network(
depths=[16, 32, 32], use_batch_norm=False, dropout_rate=0.)
state_dim = 256
else:
# Reduced architecture for DMC
def conv_stack():
return ConvStack(observation_spec.shape)
state_dim = 50
conv_stack_critic = conv_stack()
conv_target_stack_critic = conv_stack()
if observation_spec.shape == (64, 64, 3):
conv_stack_critic.output_size = state_dim
conv_target_stack_critic.output_size = state_dim
critic_kwargs['encoder'] = ImageEncoder(
conv_stack_critic, feature_dim=state_dim, bprop_conv_stack=True)
critic_kwargs['encoder_target'] = ImageEncoder(
conv_target_stack_critic, feature_dim=state_dim, bprop_conv_stack=True)
self.embedder = tf_utils.EmbedNet(
state_dim,
embedding_dim=self.embedding_dim,
num_distributions=self.num_distributions,
hidden_dims=hidden_dims)
if self.sequence_length > 2:
self.latent_embedder = tf_utils.RNNEmbedNet(
[self.sequence_length - 2, self.embedding_dim + self.embedding_dim],
embedding_dim=self.latent_dim)
self.reward_decoder = tf_utils.EmbedNet(
self.latent_dim + self.embedding_dim + self.embedding_dim,
embedding_dim=1,
hidden_dims=hidden_dims)
forward_decoder_out = (
self.embedding_dim if (self.state_prediction_mode
in ['latent', 'energy']) else self.input_dim)
forward_decoder_dists = (
self.num_distributions if
(self.state_prediction_mode in ['latent', 'energy']) else None)
self.forward_decoder = tf_utils.StochasticEmbedNet(
self.latent_dim + self.embedding_dim + self.embedding_dim,
embedding_dim=forward_decoder_out,
num_distributions=forward_decoder_dists,
hidden_dims=hidden_dims)
self.weight = tf.Variable(tf.eye(self.embedding_dim))
self.action_encoder = tf.keras.Sequential(
[
tf.keras.layers.Dense(
self.embedding_dim, use_bias=True
), # , kernel_regularizer=tf.keras.regularizers.l2(WEIGHT_DECAY)
tf.keras.layers.ReLU(),
tf.keras.layers.Dense(self.embedding_dim)
],
name='action_encoder')
if self.num_augmentations == 0:
dummy_state = tf.constant(
np.zeros(shape=[1] + list(observation_spec.shape)))
self.obs_spec = list(observation_spec.shape)
else: # account for padding of +4 everywhere and then cropping out 68
dummy_state = tf.constant(np.zeros(shape=[1, 68, 68, 3]))
self.obs_spec = [68, 68, 3]
@tf.function
def init_models():
critic_kwargs['encoder'](dummy_state)
critic_kwargs['encoder_target'](dummy_state)
self.action_encoder(
tf.cast(tf.one_hot([1], depth=action_dim), tf.float32))
init_models()
hidden_dims = (256, 256)
# self.actor = policies.CategoricalPolicy(state_dim, action_spec,
# hidden_dims=hidden_dims, encoder=actor_kwargs['encoder'])
self.critic = critic.Critic(
state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder'],
discrete_actions=True,
linear='linear_Q' in self.rep_learn_keywords)
self.critic_target = critic.Critic(
state_dim,
action_dim,
hidden_dims=hidden_dims,
encoder=critic_kwargs['encoder_target'],
discrete_actions=True,
linear='linear_Q' in self.rep_learn_keywords)
@tf.function
def init_models2():
dummy_state = tf.zeros((1, 68, 68, 3), dtype=tf.float32)
phi_s = self.critic.encoder(dummy_state)
phi_a = tf.eye(15, dtype=tf.float32)
if 'linear_Q' in self.rep_learn_keywords:
_ = self.critic.critic1.state_encoder(phi_s)
_ = self.critic.critic2.state_encoder(phi_s)
_ = self.critic.critic1.action_encoder(phi_a)
_ = self.critic.critic2.action_encoder(phi_a)
_ = self.critic_target.critic1.state_encoder(phi_s)
_ = self.critic_target.critic2.state_encoder(phi_s)
_ = self.critic_target.critic1.action_encoder(phi_a)
_ = self.critic_target.critic2.action_encoder(phi_a)
init_models2()
critic.soft_update(self.critic, self.critic_target, tau=1.0)
learning_rate = learning_rate or 1e-4
self.optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
self.critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=learning_rate)
self.all_variables = (
self.embedder.trainable_variables +
self.reward_decoder.trainable_variables +
self.forward_decoder.trainable_variables +
self.action_encoder.trainable_variables +
self.critic.trainable_variables +
self.critic_target.trainable_variables)
self.model_dict = {
'action_encoder': self.action_encoder,
'weight': self.weight,
'forward_decoder': self.forward_decoder,
'reward_decoder': self.reward_decoder,
'embedder': self.embedder,
'critic': self.critic,
'critic_target': self.critic_target,
'critic_optimizer': self.critic_optimizer,
'optimizer': self.optimizer
}