def test_train(self):
model_config = core.ModelConfig("CartPole-v0")
tc = core.StepsTrainContext()
dqn_agent = tfagents.TfDqnAgent(model_config=model_config)
dqn_agent.train(train_context=tc,
callbacks=[duration.Fast(),
log.Iteration()])
def test_train(self):
model_config = core.ModelConfig(_mountaincar_continuous_name)
tc = core.StepsTrainContext()
dqn_agent = tfagents.TfSacAgent(model_config=model_config)
dqn_agent.train(
train_context=tc,
callbacks=[duration.Fast(),
log.Iteration(),
log.Agent()])
def test_dueling_dqn_train(self):
model_config = core.ModelConfig("CartPole-v0", fc_layers=(100, ))
tc: core.StepsTrainContext = core.StepsTrainContext()
tc.num_iterations = 20000
tc.num_steps_buffer_preload = 1000
tc.num_iterations_between_eval = 1000
tc.max_steps_per_episode = 200
dqnAgent = tforce.TforceDuelingDqnAgent(model_config=model_config)
dqnAgent.train(train_context=tc,
callbacks=[log.Iteration(eval_only=True),
log.Agent()])
def test_dqn_train(self):
from easyagents.backends import tforce
model_config = core.ModelConfig(_cartpole_name, fc_layers=(100, 100))
tc: core.StepsTrainContext = core.StepsTrainContext()
tc.num_iterations = 10000
tc.num_steps_buffer_preload = 500
tc.num_iterations_between_eval = 500
tc.max_steps_per_episode = 200
dqn_agent = tforce.TforceDqnAgent(model_config=model_config)
dqn_agent.train(train_context=tc, callbacks=[log.Iteration(eval_only=True), log.Agent()])
(min_r, avg_r, max_r) = tc.eval_rewards[tc.episodes_done_in_training]
assert avg_r > 50
def train(self,
callbacks: Union[List[core.AgentCallback], core.AgentCallback,
None] = None,
num_iterations: int = 20000,
max_steps_per_episode: int = 500,
num_steps_per_iteration: int = 1,
num_steps_buffer_preload=1000,
num_steps_sampled_from_buffer=64,
num_iterations_between_eval: int = 1000,
num_episodes_per_eval: int = 10,
learning_rate: float = 0.001,
train_context: core.StepsTrainContext = None,
default_plots: bool = None):
"""Trains a new model using the gym environment passed during instantiation.
Args:
callbacks: list of callbacks called during training and evaluation
num_iterations: number of times the training is repeated (with additional data)
max_steps_per_episode: maximum number of steps per episode
num_steps_per_iteration: number of steps played per training iteration
num_steps_buffer_preload: number of initial collect steps to preload the buffer
num_steps_sampled_from_buffer: the number of steps sampled from buffer for each iteration training
num_iterations_between_eval: number of training iterations before the current policy is evaluated.
if 0 no evaluation is performed.
num_episodes_per_eval: number of episodes played to estimate the average return and steps
learning_rate: the learning rate used in the next iteration's policy training (0,1]
train_context: training configuration to be used. if set overrides all other training context arguments.
default_plots: if set adds a set of default callbacks (plot.State, plot.Rewards, plot.Loss,...).
if None default callbacks are only added if the callbacks list is empty
Returns:
train_context: the training configuration containing the loss and sum of rewards encountered
during training
"""
if train_context is None:
train_context = core.StepsTrainContext()
train_context.num_iterations = num_iterations
train_context.max_steps_per_episode = max_steps_per_episode
train_context.num_steps_per_iteration = num_steps_per_iteration
train_context.num_steps_buffer_preload = num_steps_buffer_preload
train_context.num_steps_sampled_from_buffer = num_steps_sampled_from_buffer
train_context.num_iterations_between_eval = num_iterations_between_eval
train_context.num_episodes_per_eval = num_episodes_per_eval
train_context.learning_rate = learning_rate
super().train(train_context=train_context,
callbacks=callbacks,
default_plots=default_plots)
return train_context
