def test_minibatches_per_step(self):
_epochs = self.epochs
self.epochs = 2
rl_parameters = RLParameters(gamma=0.95,
target_update_rate=0.9,
maxq_learning=True)
rainbow_parameters = RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False)
training_parameters1 = TrainingParameters(
layers=self.layers,
activations=self.activations,
minibatch_size=1024,
minibatches_per_step=1,
learning_rate=0.25,
optimizer="ADAM",
)
training_parameters2 = TrainingParameters(
layers=self.layers,
activations=self.activations,
minibatch_size=128,
minibatches_per_step=8,
learning_rate=0.25,
optimizer="ADAM",
)
env1 = Env(self.state_dims, self.action_dims)
env2 = Env(self.state_dims, self.action_dims)
model_parameters1 = DiscreteActionModelParameters(
actions=env1.actions,
rl=rl_parameters,
rainbow=rainbow_parameters,
training=training_parameters1,
)
model_parameters2 = DiscreteActionModelParameters(
actions=env2.actions,
rl=rl_parameters,
rainbow=rainbow_parameters,
training=training_parameters2,
)
# minibatch_size / 8, minibatches_per_step * 8 should give the same result
logger.info("Training model 1")
trainer1 = self._train(model_parameters1, env1)
SummaryWriterContext._reset_globals()
logger.info("Training model 2")
trainer2 = self._train(model_parameters2, env2)
weight1 = trainer1.q_network.fc.layers[-1].weight.detach().numpy()
weight2 = trainer2.q_network.fc.layers[-1].weight.detach().numpy()
# Due to numerical stability this tolerance has to be fairly high
self.assertTrue(np.allclose(weight1, weight2, rtol=0.0, atol=1e-3))
self.epochs = _epochs
def get_sarsa_parameters_factorized(self):
return ContinuousActionModelParameters(
rl=RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
reward_burnin=100,
maxq_learning=False,
),
training=TrainingParameters(
layers=[],
activations=[],
factorization_parameters=FactorizationParameters(
state=FeedForwardParameters(
layers=[-1, 128, 64, 32],
activations=["relu", "relu", "linear"]),
action=FeedForwardParameters(
layers=[-1, 128, 64, 32],
activations=["relu", "relu", "linear"]),
),
minibatch_size=self.minibatch_size,
learning_rate=0.05,
optimizer="ADAM",
),
knn=KnnParameters(model_type="DQN"),
in_training_cpe_evaluation=InTrainingCPEParameters(
mdp_sampled_rate=0.1),
)
def test_input_validation(self):
with self.assertRaises(Exception):
# layers and activations sizes incompatible
MLTrainer(
"Test model",
TrainingParameters([2, 1], ['linear', 'relu'], 100, 0.001,
'ADAM'))
with self.assertRaises(Exception):
# All values in layers should be positive integers
MLTrainer(
"Test model",
TrainingParameters([-1, 1], ['linear'], 100, 0.001, 'ADAM'))
MLTrainer(
"Test model",
TrainingParameters([1.3, 1], ['linear'], 100, 0.001, 'ADAM'))
def test_trainer_maxq(self):
env = Env(self.state_dims, self.action_dims)
env.seed(42)
maxq_parameters = DiscreteActionModelParameters(
actions=env.actions,
rl=RLParameters(
gamma=0.99,
target_update_rate=1.0,
reward_burnin=100,
maxq_learning=True,
),
training=TrainingParameters(
layers=self.layers,
activations=self.activations,
minibatch_size=self.minibatch_size,
learning_rate=1.0,
optimizer="ADAM",
),
)
maxq_trainer = DiscreteActionTrainer(maxq_parameters,
env.normalization)
# predictor = maxq_trainer.predictor()
logger.info("Generating constant_reward MDPs..")
states, actions, rewards, next_states, next_actions, is_terminal, possible_next_actions = env.generate_samples_discrete(
self.num_samples)
logger.info("Preprocessing constant_reward MDPs..")
tdps = env.preprocess_samples_discrete(
states,
actions,
rewards,
next_states,
next_actions,
is_terminal,
possible_next_actions,
self.minibatch_size,
)
for epoch in range(self.epochs):
logger.info("Training.. " + str(epoch))
for tdp in tdps:
maxq_trainer.train_numpy(tdp, None)
logger.info(" ".join([
"Training epoch",
str(epoch),
"average q values",
str(np.mean(workspace.FetchBlob(maxq_trainer.q_score_output))),
"td_loss",
str(workspace.FetchBlob(maxq_trainer.loss_blob)),
]))
# Q value should converge to very close to 100
avg_q_value_after_training = np.mean(
workspace.FetchBlob(maxq_trainer.q_score_output))
self.assertLess(avg_q_value_after_training, 101)
self.assertGreater(avg_q_value_after_training, 99)
def get_sarsa_trainer_reward_boost(
self,
environment,
reward_shape,
dueling,
use_gpu=False,
use_all_avail_gpus=False,
):
rl_parameters = RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
reward_burnin=10,
maxq_learning=False,
reward_boost=reward_shape,
)
training_parameters = TrainingParameters(
layers=[-1, 128, -1] if dueling else [-1, -1],
activations=["relu", "linear"] if dueling else ["linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.05,
optimizer="ADAM",
)
return DQNTrainer(
DiscreteActionModelParameters(
actions=environment.ACTIONS,
rl=rl_parameters,
training=training_parameters,
rainbow=RainbowDQNParameters(
double_q_learning=True, dueling_architecture=dueling
),
),
environment.normalization,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
def get_sarsa_parameters(self, environment, reward_shape, dueling,
categorical, clip_grad_norm):
rl_parameters = RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
maxq_learning=False,
reward_boost=reward_shape,
)
training_parameters = TrainingParameters(
layers=[-1, 128, -1] if dueling else [-1, -1],
activations=["relu", "linear"] if dueling else ["linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.05,
optimizer="ADAM",
clip_grad_norm=clip_grad_norm,
)
return DiscreteActionModelParameters(
actions=environment.ACTIONS,
rl=rl_parameters,
training=training_parameters,
rainbow=RainbowDQNParameters(
double_q_learning=True,
dueling_architecture=dueling,
categorical=categorical,
),
)
def get_sarsa_trainer_reward_boost(self, environment, reward_shape):
rl_parameters = RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
reward_burnin=10,
maxq_learning=False,
reward_boost=reward_shape,
)
training_parameters = TrainingParameters(
layers=[-1, -1],
activations=["linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.125,
optimizer="ADAM",
)
return DiscreteActionTrainer(
DiscreteActionModelParameters(
actions=environment.ACTIONS,
rl=rl_parameters,
training=training_parameters,
rainbow=RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False),
in_training_cpe=InTrainingCPEParameters(mdp_sampled_rate=0.1),
),
environment.normalization,
)
def get_sarsa_parameters_factorized(self):
return ContinuousActionModelParameters(
rl=RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
reward_burnin=100,
maxq_learning=False,
),
training=TrainingParameters(
# These are used by reward network
layers=[-1, 256, 128, -1],
activations=["relu", "relu", "linear"],
factorization_parameters=FactorizationParameters(
state=FeedForwardParameters(layers=[-1, 128, 64],
activations=["relu",
"linear"]),
action=FeedForwardParameters(
layers=[-1, 128, 64], activations=["relu", "linear"]),
),
minibatch_size=self.minibatch_size,
learning_rate=0.03,
optimizer="ADAM",
),
rainbow=RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False),
in_training_cpe=InTrainingCPEParameters(mdp_sampled_rate=0.1),
)
def test_trainer_maxq(self):
env = Env(self.state_dims, self.action_dims)
env.seed(42)
maxq_parameters = DiscreteActionModelParameters(
actions=env.actions,
rl=RLParameters(
gamma=0.99,
target_update_rate=0.9,
reward_burnin=100,
maxq_learning=True,
),
rainbow=RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False),
training=TrainingParameters(
layers=self.layers,
activations=self.activations,
minibatch_size=self.minibatch_size,
learning_rate=0.25,
optimizer="ADAM",
),
)
maxq_trainer = DQNTrainer(maxq_parameters, env.normalization)
logger.info("Generating constant_reward MDPs..")
states, actions, rewards, next_states, next_actions, is_terminal, possible_actions, possible_next_actions = env.generate_samples_discrete(
self.num_samples)
logger.info("Preprocessing constant_reward MDPs..")
for epoch in range(self.epochs):
tdps = env.preprocess_samples_discrete(
states,
actions,
rewards,
next_states,
next_actions,
is_terminal,
possible_actions,
possible_next_actions,
self.minibatch_size,
)
logger.info("Training.. " + str(epoch))
for tdp in tdps:
maxq_trainer.train(tdp)
logger.info(" ".join([
"Training epoch",
str(epoch),
"average q values",
str(torch.mean(maxq_trainer.all_action_scores)),
]))
# Q value should converge to very close to 100
avg_q_value_after_training = torch.mean(maxq_trainer.all_action_scores)
self.assertLess(avg_q_value_after_training, 101)
self.assertGreater(avg_q_value_after_training, 99)
def main(params):
# Set minibatch size based on # of devices being used to train
params["training"]["minibatch_size"] *= minibatch_size_multiplier(
params["use_gpu"], params["use_all_avail_gpus"])
rl_parameters = RLParameters(**params["rl"])
training_parameters = TrainingParameters(**params["training"])
rainbow_parameters = RainbowDQNParameters(**params["rainbow"])
model_params = ContinuousActionModelParameters(
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters)
state_normalization = BaseWorkflow.read_norm_file(
params["state_norm_data_path"])
action_normalization = BaseWorkflow.read_norm_file(
params["action_norm_data_path"])
writer = SummaryWriter(log_dir=params["model_output_path"])
logger.info("TensorBoard logging location is: {}".format(writer.log_dir))
preprocess_handler = ParametricDqnPreprocessHandler(
Preprocessor(state_normalization, False),
Preprocessor(action_normalization, False),
PandasSparseToDenseProcessor(),
)
workflow = ParametricDqnWorkflow(
model_params,
preprocess_handler,
state_normalization,
action_normalization,
params["use_gpu"],
params["use_all_avail_gpus"],
)
train_dataset = JSONDatasetReader(
params["training_data_path"],
batch_size=training_parameters.minibatch_size)
eval_dataset = JSONDatasetReader(params["eval_data_path"], batch_size=16)
with summary_writer_context(writer):
workflow.train_network(train_dataset, eval_dataset,
int(params["epochs"]))
exporter = ParametricDQNExporter(
workflow.trainer.q_network,
PredictorFeatureExtractor(
state_normalization_parameters=state_normalization,
action_normalization_parameters=action_normalization,
),
ParametricActionOutputTransformer(),
)
return export_trainer_and_predictor(workflow.trainer,
params["model_output_path"],
exporter=exporter) # noqa
def test_pure_q_learning_all_cheat(self):
q_learning_parameters = DiscreteActionModelParameters(
actions=self._env.ACTIONS,
rl=self._rl_parameters_all_cheat_maxq,
training=TrainingParameters(
layers=[self._env.width * self._env.height, 1],
activations=['linear'],
minibatch_size=self.minibatch_size,
learning_rate=0.05,
optimizer='SGD',
lr_policy='fixed',
)
)
trainer = DiscreteActionTrainer(
q_learning_parameters,
self._env.normalization,
)
predictor = trainer.predictor()
policy = _build_policy(self._env, predictor, 1)
initial_state = self._env.reset()
iteration_result = _collect_samples(
self._env, policy, 20000, initial_state
)
num_iterations = 50
for _ in range(num_iterations):
tdps = self._env.preprocess_samples(
iteration_result.states,
iteration_result.actions,
iteration_result.rewards,
iteration_result.next_states,
iteration_result.next_actions,
iteration_result.is_terminals,
iteration_result.possible_next_actions,
None,
self.minibatch_size,
)
for tdp in tdps:
trainer.train_numpy(tdp, None)
initial_state = self._env.reset()
policy = _build_policy(self._env, predictor, 0.1)
iteration_result = _collect_samples(
self._env, policy, 20000, initial_state
)
policy = _build_policy(self._env, predictor, 0)
initial_state = self._env.reset()
iteration_result = _collect_samples(
self._env, policy, 1000, initial_state
)
# 100% should be cheat. Will fix in the future.
self.assertGreater(
np.sum(np.array(iteration_result.actions) == 'C'), 800
)
def test_trainer_maxq(self):
environment = Gridworld()
maxq_sarsa_parameters = DiscreteActionModelParameters(
actions=environment.ACTIONS,
rl=RLParameters(
gamma=DISCOUNT,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=True,
),
training=TrainingParameters(
layers=[-1, 1],
activations=["linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.01,
optimizer="ADAM",
),
)
# construct the new trainer that using maxq
maxq_trainer = DiscreteActionTrainer(
maxq_sarsa_parameters, environment.normalization
)
samples = environment.generate_samples(100000, 1.0)
predictor = maxq_trainer.predictor()
tdps = environment.preprocess_samples(samples, self.minibatch_size)
evaluator = GridworldEvaluator(environment, True)
evaluator.evaluate(predictor)
print(
"Pre-Training eval: ",
evaluator.mc_loss[-1],
evaluator.reward_doubly_robust[-1],
)
self.assertGreater(evaluator.mc_loss[-1], 0.3)
for _ in range(5):
for tdp in tdps:
maxq_trainer.train_numpy(tdp, None)
evaluator.evaluate(predictor)
print(
"Post-Training eval: ",
evaluator.mc_loss[-1],
evaluator.reward_doubly_robust[-1],
)
self.assertLess(evaluator.mc_loss[-1], 0.1)
self.assertGreater(
evaluator.reward_doubly_robust[-1], evaluator.reward_doubly_robust[-2]
)
def get_sarsa_parameters(self):
return ContinuousActionModelParameters(
rl=RLParameters(gamma=DISCOUNT,
target_update_rate=1.0,
maxq_learning=False),
training=TrainingParameters(
layers=[-1, 256, 128, -1],
activations=["relu", "relu", "linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.05,
optimizer="ADAM",
),
rainbow=RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False),
)
def get_sarsa_parameters(self):
return ContinuousActionModelParameters(
rl=RLParameters(
gamma=DISCOUNT,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=False,
),
training=TrainingParameters(
layers=[-1, 200, 1],
activations=['linear', 'linear'],
minibatch_size=1024,
learning_rate=0.01,
optimizer='ADAM',
),
knn=KnnParameters(model_type='DQN', ))
def get_sarsa_parameters(self):
return ContinuousActionModelParameters(
rl=RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
reward_burnin=100,
maxq_learning=False,
),
training=TrainingParameters(
layers=[-1, 256, 128, -1],
activations=["relu", "relu", "linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.1,
optimizer="ADAM",
),
knn=KnnParameters(model_type="DQN"),
)
def get_sarsa_trainer(self, environment):
rl_parameters = RLParameters(gamma=DISCOUNT,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=False)
training_parameters = TrainingParameters(
layers=[-1, 1],
activations=['linear'],
minibatch_size=1024,
learning_rate=0.01,
optimizer='ADAM',
)
return DiscreteActionTrainer(
environment.normalization,
DiscreteActionModelParameters(actions=environment.ACTIONS,
rl=rl_parameters,
training=training_parameters))
def train_network(params):
logger.info("Running DQN workflow with params:")
logger.info(params)
action_names = np.array(params["actions"])
rl_parameters = RLParameters(**params["rl"])
training_parameters = TrainingParameters(**params["training"])
rainbow_parameters = RainbowDQNParameters(**params["rainbow"])
trainer_params = DiscreteActionModelParameters(
actions=params["actions"],
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters,
)
dataset = JSONDataset(params["training_data_path"],
batch_size=training_parameters.minibatch_size)
norm_data = JSONDataset(params["state_norm_data_path"])
state_normalization = read_norm_params(norm_data.read_all())
num_batches = int(len(dataset) / training_parameters.minibatch_size)
logger.info("Read in batch data set {} of size {} examples. Data split "
"into {} batches of size {}.".format(
params["training_data_path"],
len(dataset),
num_batches,
training_parameters.minibatch_size,
))
trainer = DQNTrainer(trainer_params, state_normalization,
params["use_gpu"])
for epoch in range(params["epochs"]):
for batch_idx in range(num_batches):
helpers.report_training_status(batch_idx, num_batches, epoch,
params["epochs"])
batch = dataset.read_batch(batch_idx)
tdp = preprocess_batch_for_training(action_names, batch,
state_normalization)
trainer.train(tdp)
logger.info("Training finished. Saving PyTorch model to {}".format(
params["pytorch_output_path"]))
helpers.save_model_to_file(trainer, params["pytorch_output_path"])
def test_adam_weights(self):
num_features = 4
num_outputs = 1
trainer = MLTrainer(
"Linear Regression",
TrainingParameters(layers=[num_features, num_outputs],
activations=['linear'],
minibatch_size=100,
learning_rate=0.1,
optimizer='ADAM'))
dist = get_weight_dist(trainer,
num_features=num_features,
num_outputs=num_outputs)
self.assertLess(dist, 0.1)
def test_trainer_maxq(self):
environment = Gridworld()
maxq_sarsa_parameters = DiscreteActionModelParameters(
actions=environment.ACTIONS,
rl=RLParameters(gamma=DISCOUNT,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=True),
training=TrainingParameters(
layers=[-1, 1],
activations=['linear'],
minibatch_size=self.minibatch_size,
learning_rate=0.01,
optimizer='ADAM',
))
# construct the new trainer that using maxq
maxq_trainer = DiscreteActionTrainer(
maxq_sarsa_parameters,
environment.normalization,
)
states, actions, rewards, next_states, next_actions, is_terminal,\
possible_next_actions, reward_timelines = \
environment.generate_samples(100000, 1.0)
predictor = maxq_trainer.predictor()
tdps = environment.preprocess_samples(
states,
actions,
rewards,
next_states,
next_actions,
is_terminal,
possible_next_actions,
reward_timelines,
self.minibatch_size,
)
evaluator = GridworldEvaluator(environment, True)
print("Pre-Training eval", evaluator.evaluate(predictor))
self.assertGreater(evaluator.evaluate(predictor), 0.3)
for _ in range(2):
for tdp in tdps:
maxq_trainer.stream_tdp(tdp, None)
evaluator.evaluate(predictor)
print("Post-Training eval", evaluator.evaluate(predictor))
self.assertLess(evaluator.evaluate(predictor), 0.1)
def test_sarsa_layer_validation(self):
env = Gridworld()
invalid_sarsa_params = DiscreteActionModelParameters(
actions=env.ACTIONS,
rl=RLParameters(gamma=DISCOUNT,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=False),
training=TrainingParameters(
layers=[-1, 3],
activations=['linear'],
minibatch_size=32,
learning_rate=0.1,
optimizer='SGD',
))
with self.assertRaises(Exception):
# layers[-1] should be 1
DiscreteActionTrainer(env.normalization, invalid_sarsa_params)
def main(args):
parser = argparse.ArgumentParser(
description="Train a RL net to play in an OpenAI Gym environment.")
parser.add_argument("-p",
"--parameters",
help="Path to JSON parameters file.")
parser.add_argument("-s",
"--score-bar",
help="Bar for averaged tests scores.",
type=float,
default=None)
parser.add_argument(
"-g",
"--gpu_id",
help="If set, will use GPU with specified ID. Otherwise will use CPU.",
default=USE_CPU)
args = parser.parse_args(args)
with open(args.parameters, 'r') as f:
params = json.load(f)
rl_settings = params['rl']
training_settings = params['training']
rl_settings['gamma'] = rl_settings['reward_discount_factor']
del rl_settings['reward_discount_factor']
training_settings['gamma'] = training_settings['learning_rate_decay']
del training_settings['learning_rate_decay']
env_type = params['env']
env = OpenAIGymEnvironment(env_type, rl_settings['epsilon'])
trainer_params = DiscreteActionModelParameters(
actions=env.actions,
rl=RLParameters(**rl_settings),
training=TrainingParameters(**training_settings))
device = core.DeviceOption(
caffe2_pb2.CPU if args.gpu_id == USE_CPU else caffe2_pb2.CUDA,
args.gpu_id)
with core.DeviceScope(device):
trainer = DiscreteActionTrainer(env.normalization,
trainer_params,
skip_normalization=True)
return run(env, trainer, "{} test run".format(env_type),
args.score_bar, **params["run_details"])
def test_sgd_weights(self):
num_features = 4
num_outputs = 1
trainer = MLTrainer(
"Linear Regression",
TrainingParameters(layers=[num_features, num_outputs],
activations=['linear'],
minibatch_size=100,
learning_rate=0.001,
optimizer='SGD',
gamma=0.9999,
lr_policy='step'))
dist = get_weight_dist(trainer,
num_features=num_features,
num_outputs=num_outputs)
self.assertLess(dist, 0.1)
def get_sarsa_parameters(self):
return ContinuousActionModelParameters(
rl=RLParameters(
gamma=DISCOUNT,
target_update_rate=1.0,
reward_burnin=100,
maxq_learning=False,
),
training=TrainingParameters(
layers=[-1, 256, 128, -1],
activations=["relu", "relu", "linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.05,
optimizer="ADAM",
),
knn=KnnParameters(model_type="DQN"),
rainbow=RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False),
in_training_cpe=InTrainingCPEParameters(mdp_sampled_rate=0.1),
)
def test_pure_q_learning_all_cheat(self):
q_learning_parameters = DiscreteActionModelParameters(
actions=self._env.ACTIONS,
rl=self._rl_parameters_all_cheat_maxq,
training=TrainingParameters(
layers=[self._env.width * self._env.height, 1],
activations=['linear'],
minibatch_size=32,
learning_rate=0.05,
optimizer='SGD',
lr_policy='fixed'))
trainer = DiscreteActionTrainer(self._env.normalization,
q_learning_parameters)
predictor = trainer.predictor()
policy = _build_policy(self._env, predictor, 1)
initial_state = self._env.reset()
iteration_result = _collect_samples(self._env, policy, 10000,
initial_state)
num_iterations = 50
for _ in range(num_iterations):
policy = _build_policy(self._env, predictor, 0)
tdp = self._env.preprocess_samples(
iteration_result.states,
iteration_result.actions,
iteration_result.rewards,
iteration_result.next_states,
iteration_result.next_actions,
iteration_result.is_terminals,
iteration_result.possible_next_actions,
None,
)
trainer.stream_tdp(tdp, None)
initial_state = iteration_result.current_state
initial_state = self._env.reset()
iteration_result = _collect_samples(self._env, policy, 10000,
initial_state)
self.assertTrue(np.all(np.array(iteration_result.actions) == 'C'))
def test_scaledoutput_regression_weights(self):
# this is testing if scaled/inner-product in final layer,
# using same case but transformed to (wx+b) * (1/y) = 1,
# in this test case, learned weight should be regularized on columns
num_features = 4
num_training_samples = 100
num_outputs = 2
training_inputs, training_outputs, weights, _ = (gen_training_data(
num_features, num_training_samples, num_outputs, 0.01))
trainer = MLTrainerIP("Linear Regression",
TrainingParameters(
layers=[num_features, num_outputs],
activations=['linear'],
minibatch_size=100,
learning_rate=0.1),
scaled_output=True)
training_labels = np.ones(
(training_inputs.shape[0], 1)).astype(np.float32)
training_scale = np.where(training_outputs == 0, 0,
1.0 / (num_outputs * training_outputs))
for _ in range(10000):
trainer.train_wexternal(training_inputs, training_labels,
training_scale)
trained_weights = np.concatenate(
[workspace.FetchBlob(w) for w in trainer.weights],
axis=0).transpose()
# regularized after training to keep same scale as given weights
scale_trained_weights = np.copy(trained_weights)
for i in range(trained_weights.shape[1]):
scaling = 1.0 / trained_weights[0, i] * weights[0, i]
scale_trained_weights[:, i] = trained_weights[:, i] * scaling
dist = np.linalg.norm(scale_trained_weights - weights)
self.assertLess(dist, 0.1)
def test_trainer_maxq(self):
env = Env(self.state_dims, self.action_dims)
maxq_parameters = DiscreteActionModelParameters(
actions=env.actions,
rl=RLParameters(gamma=0.95,
target_update_rate=0.9,
maxq_learning=True),
rainbow=RainbowDQNParameters(double_q_learning=True,
dueling_architecture=False),
training=TrainingParameters(
layers=self.layers,
activations=self.activations,
minibatch_size=1024,
learning_rate=0.25,
optimizer="ADAM",
),
)
# Q value should converge to very close to 20
trainer = self._train(maxq_parameters, env)
avg_q_value_after_training = torch.mean(trainer.all_action_scores)
self.assertLess(avg_q_value_after_training, 22)
self.assertGreater(avg_q_value_after_training, 18)
def get_sarsa_trainer_reward_boost(self, environment, reward_shape):
rl_parameters = RLParameters(
gamma=DISCOUNT,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=False,
reward_boost=reward_shape,
)
training_parameters = TrainingParameters(
layers=[-1, -1],
activations=["linear"],
minibatch_size=self.minibatch_size,
learning_rate=0.01,
optimizer="ADAM",
)
return DiscreteActionTrainer(
DiscreteActionModelParameters(
actions=environment.ACTIONS,
rl=rl_parameters,
training=training_parameters,
),
environment.normalization,
)
def run_gym(params, score_bar, gpu_id):
rl_settings = params['rl']
training_settings = params['training']
rl_settings['gamma'] = rl_settings['reward_discount_factor']
del rl_settings['reward_discount_factor']
training_settings['gamma'] = training_settings['learning_rate_decay']
del training_settings['learning_rate_decay']
env_type = params['env']
env = OpenAIGymEnvironment(env_type, rl_settings['epsilon'])
trainer_params = DiscreteActionModelParameters(
actions=env.actions,
rl=RLParameters(**rl_settings),
training=TrainingParameters(**training_settings))
device = core.DeviceOption(
caffe2_pb2.CPU if gpu_id == USE_CPU else caffe2_pb2.CUDA,
gpu_id,
)
with core.DeviceScope(device):
if env.img:
trainer = DiscreteActionConvTrainer(
DiscreteActionConvModelParameters(
fc_parameters=trainer_params,
cnn_parameters=CNNModelParameters(**params['cnn']),
num_input_channels=env.num_input_channels,
img_height=env.height,
img_width=env.width),
env.normalization,
)
else:
trainer = DiscreteActionTrainer(
trainer_params,
env.normalization,
)
return run(env, trainer, "{} test run".format(env_type), score_bar,
**params["run_details"])
def setUp(self):
super(self.__class__, self).setUp()
np.random.seed(0)
random.seed(0)
self.state_dim, self.action_dim = 2, 3
self._env = MockEnv(self.state_dim, self.action_dim)
self._rl_parameters = RLParameters(
gamma=0.9,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=False,
)
self._rl_parameters_maxq = RLParameters(
gamma=0.9,
target_update_rate=0.5,
reward_burnin=10,
maxq_learning=True,
)
self._rl_parameters = ContinuousActionModelParameters(
rl=self._rl_parameters,
training=TrainingParameters(
layers=[
-1, self._env.num_states * self._env.num_actions * 2, 1
],
activations=['linear', 'linear'],
minibatch_size=1024,
learning_rate=0.01,
optimizer='ADAM',
),
knn=KnnParameters(model_type='DQN', ))
self._trainer = ContinuousActionDQNTrainer(
self._env.normalization, self._env.normalization_action,
self._rl_parameters)
def create_trainer(model_type, params, rl_parameters, use_gpu, env):
if model_type == ModelType.PYTORCH_DISCRETE_DQN.value:
training_parameters = params["training"]
if isinstance(training_parameters, dict):
training_parameters = TrainingParameters(**training_parameters)
rainbow_parameters = params["rainbow"]
if isinstance(rainbow_parameters, dict):
rainbow_parameters = RainbowDQNParameters(**rainbow_parameters)
if env.img:
assert (training_parameters.cnn_parameters
is not None), "Missing CNN parameters for image input"
if isinstance(training_parameters.cnn_parameters, dict):
training_parameters.cnn_parameters = CNNParameters(
**training_parameters.cnn_parameters)
training_parameters.cnn_parameters.conv_dims[
0] = env.num_input_channels
training_parameters.cnn_parameters.input_height = env.height
training_parameters.cnn_parameters.input_width = env.width
training_parameters.cnn_parameters.num_input_channels = (
env.num_input_channels)
else:
assert (training_parameters.cnn_parameters is
None), "Extra CNN parameters for non-image input"
trainer_params = DiscreteActionModelParameters(
actions=env.actions,
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters,
)
trainer = DQNTrainer(trainer_params, env.normalization, use_gpu)
elif model_type == ModelType.PYTORCH_PARAMETRIC_DQN.value:
training_parameters = params["training"]
if isinstance(training_parameters, dict):
training_parameters = TrainingParameters(**training_parameters)
rainbow_parameters = params["rainbow"]
if isinstance(rainbow_parameters, dict):
rainbow_parameters = RainbowDQNParameters(**rainbow_parameters)
if env.img:
assert (training_parameters.cnn_parameters
is not None), "Missing CNN parameters for image input"
training_parameters.cnn_parameters.conv_dims[
0] = env.num_input_channels
else:
assert (training_parameters.cnn_parameters is
None), "Extra CNN parameters for non-image input"
trainer_params = ContinuousActionModelParameters(
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters)
trainer = ParametricDQNTrainer(trainer_params, env.normalization,
env.normalization_action, use_gpu)
elif model_type == ModelType.CONTINUOUS_ACTION.value:
training_parameters = params["shared_training"]
if isinstance(training_parameters, dict):
training_parameters = DDPGTrainingParameters(**training_parameters)
actor_parameters = params["actor_training"]
if isinstance(actor_parameters, dict):
actor_parameters = DDPGNetworkParameters(**actor_parameters)
critic_parameters = params["critic_training"]
if isinstance(critic_parameters, dict):
critic_parameters = DDPGNetworkParameters(**critic_parameters)
trainer_params = DDPGModelParameters(
rl=rl_parameters,
shared_training=training_parameters,
actor_training=actor_parameters,
critic_training=critic_parameters,
)
action_range_low = env.action_space.low.astype(np.float32)
action_range_high = env.action_space.high.astype(np.float32)
trainer = DDPGTrainer(
trainer_params,
env.normalization,
env.normalization_action,
torch.from_numpy(action_range_low).unsqueeze(dim=0),
torch.from_numpy(action_range_high).unsqueeze(dim=0),
use_gpu,
)
elif model_type == ModelType.SOFT_ACTOR_CRITIC.value:
trainer_params = SACModelParameters(
rl=rl_parameters,
training=SACTrainingParameters(
minibatch_size=params["sac_training"]["minibatch_size"],
use_2_q_functions=params["sac_training"]["use_2_q_functions"],
q_network_optimizer=OptimizerParameters(
**params["sac_training"]["q_network_optimizer"]),
value_network_optimizer=OptimizerParameters(
**params["sac_training"]["value_network_optimizer"]),
actor_network_optimizer=OptimizerParameters(
**params["sac_training"]["actor_network_optimizer"]),
entropy_temperature=params["sac_training"]
["entropy_temperature"],
),
q_network=FeedForwardParameters(**params["sac_q_training"]),
value_network=FeedForwardParameters(
**params["sac_value_training"]),
actor_network=FeedForwardParameters(
**params["sac_actor_training"]),
)
trainer = get_sac_trainer(env, trainer_params, use_gpu)
else:
raise NotImplementedError(
"Model of type {} not supported".format(model_type))
return trainer
def create_trainer(model_type, params, rl_parameters, use_gpu, env):
if model_type == ModelType.PYTORCH_DISCRETE_DQN.value:
training_parameters = params["training"]
if isinstance(training_parameters, dict):
training_parameters = TrainingParameters(**training_parameters)
rainbow_parameters = params["rainbow"]
if isinstance(rainbow_parameters, dict):
rainbow_parameters = RainbowDQNParameters(**rainbow_parameters)
if env.img:
assert (
training_parameters.cnn_parameters is not None
), "Missing CNN parameters for image input"
if isinstance(training_parameters.cnn_parameters, dict):
training_parameters.cnn_parameters = CNNParameters(
**training_parameters.cnn_parameters
)
training_parameters.cnn_parameters.conv_dims[0] = env.num_input_channels
training_parameters.cnn_parameters.input_height = env.height
training_parameters.cnn_parameters.input_width = env.width
training_parameters.cnn_parameters.num_input_channels = (
env.num_input_channels
)
else:
assert (
training_parameters.cnn_parameters is None
), "Extra CNN parameters for non-image input"
trainer_params = DiscreteActionModelParameters(
actions=env.actions,
rl=rl_parameters,
training=training_parameters,
rainbow=rainbow_parameters,
)
trainer = DQNTrainer(trainer_params, env.normalization, use_gpu)
elif model_type == ModelType.PYTORCH_PARAMETRIC_DQN.value:
training_parameters = params["training"]
if isinstance(training_parameters, dict):
training_parameters = TrainingParameters(**training_parameters)
rainbow_parameters = params["rainbow"]
if isinstance(rainbow_parameters, dict):
rainbow_parameters = RainbowDQNParameters(**rainbow_parameters)
if env.img:
assert (
training_parameters.cnn_parameters is not None
), "Missing CNN parameters for image input"
training_parameters.cnn_parameters.conv_dims[0] = env.num_input_channels
else:
assert (
training_parameters.cnn_parameters is None
), "Extra CNN parameters for non-image input"
trainer_params = ContinuousActionModelParameters(
rl=rl_parameters, training=training_parameters, rainbow=rainbow_parameters
)
trainer = ParametricDQNTrainer(
trainer_params, env.normalization, env.normalization_action, use_gpu
)
elif model_type == ModelType.CONTINUOUS_ACTION.value:
training_parameters = params["shared_training"]
if isinstance(training_parameters, dict):
training_parameters = DDPGTrainingParameters(**training_parameters)
actor_parameters = params["actor_training"]
if isinstance(actor_parameters, dict):
actor_parameters = DDPGNetworkParameters(**actor_parameters)
critic_parameters = params["critic_training"]
if isinstance(critic_parameters, dict):
critic_parameters = DDPGNetworkParameters(**critic_parameters)
trainer_params = DDPGModelParameters(
rl=rl_parameters,
shared_training=training_parameters,
actor_training=actor_parameters,
critic_training=critic_parameters,
)
action_range_low = env.action_space.low.astype(np.float32)
action_range_high = env.action_space.high.astype(np.float32)
trainer = DDPGTrainer(
trainer_params,
env.normalization,
env.normalization_action,
torch.from_numpy(action_range_low).unsqueeze(dim=0),
torch.from_numpy(action_range_high).unsqueeze(dim=0),
use_gpu,
)
elif model_type == ModelType.SOFT_ACTOR_CRITIC.value:
trainer_params = SACModelParameters(
rl=rl_parameters,
training=SACTrainingParameters(
minibatch_size=params["sac_training"]["minibatch_size"],
use_2_q_functions=params["sac_training"]["use_2_q_functions"],
q_network_optimizer=OptimizerParameters(
**params["sac_training"]["q_network_optimizer"]
),
value_network_optimizer=OptimizerParameters(
**params["sac_training"]["value_network_optimizer"]
),
actor_network_optimizer=OptimizerParameters(
**params["sac_training"]["actor_network_optimizer"]
),
entropy_temperature=params["sac_training"]["entropy_temperature"],
),
q_network=FeedForwardParameters(**params["sac_q_training"]),
value_network=FeedForwardParameters(**params["sac_value_training"]),
actor_network=FeedForwardParameters(**params["sac_actor_training"]),
)
trainer = get_sac_trainer(env, trainer_params, use_gpu)
else:
raise NotImplementedError("Model of type {} not supported".format(model_type))
return trainer
