def _test_reward_boost(self, use_gpu=False, use_all_avail_gpus=False):
environment = Gridworld()
reward_boost = {"L": 100, "R": 200, "U": 300, "D": 400}
trainer = self.get_sarsa_trainer_reward_boost(
environment,
reward_boost,
False,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
predictor = trainer.predictor()
samples = environment.generate_samples(100000, 1.0, DISCOUNT)
rewards_update = []
for action, reward in zip(samples.actions, samples.rewards):
rewards_update.append(reward - reward_boost[action])
samples.rewards = rewards_update
evaluator = GridworldEvaluator(environment, False, DISCOUNT, False,
samples)
tdps = environment.preprocess_samples(samples,
self.minibatch_size,
use_gpu=use_gpu)
with tempfile.TemporaryDirectory() as tmpdirname:
tmp_path = os.path.join(tmpdirname, "model")
predictor.save(tmp_path, "minidb")
new_predictor = DQNPredictor.load(tmp_path, "minidb", False)
evaluator.evaluate(new_predictor)
print(
"Pre-Training eval: ",
evaluator.mc_loss[-1],
evaluator.value_doubly_robust[-1],
)
for tdp in tdps:
trainer.train(tdp, None)
predictor = trainer.predictor()
with tempfile.TemporaryDirectory() as tmpdirname:
tmp_path = os.path.join(tmpdirname, "model")
predictor.save(tmp_path, "minidb")
new_predictor = DQNPredictor.load(tmp_path, "minidb", False)
evaluator.evaluate(new_predictor)
print(
"Post-Training eval: ",
evaluator.mc_loss[-1],
evaluator.value_doubly_robust[-1],
)
self.assertLess(evaluator.mc_loss[-1], 0.1)
def test_predictor_export(self):
"""Verify that q-values before model export equal q-values after
model export. Meant to catch issues with export logic."""
environment = Gridworld()
trainer = trainer = self.get_sarsa_trainer(environment, False)
samples = Samples(
mdp_ids=["0"],
sequence_numbers=[0],
states=[{0: 1.0, 1: 1.0, 2: 1.0, 3: 1.0, 4: 1.0, 5: 1.0, 15: 1.0, 24: 1.0}],
actions=["D"],
action_probabilities=[0.5],
rewards=[0],
possible_actions=[["R", "D"]],
next_states=[{5: 1.0}],
next_actions=["U"],
terminals=[False],
possible_next_actions=[["R", "U", "D"]],
)
tdps = environment.preprocess_samples(samples, 1)
pre_export_q_values = trainer.q_network(tdps[0].states).detach().numpy()
predictor = trainer.predictor()
with tempfile.TemporaryDirectory() as tmpdirname:
tmp_path = os.path.join(tmpdirname, "model")
predictor.save(tmp_path, "minidb")
new_predictor = DQNPredictor.load(tmp_path, "minidb", False)
post_export_q_values = new_predictor.predict([samples.states[0]])
for i, action in enumerate(environment.ACTIONS):
self.assertAlmostEquals(
pre_export_q_values[0][i], post_export_q_values[0][action], places=4
)
def _test_dqn_workflow(self, use_gpu=False, use_all_avail_gpus=False):
"""Run DQN workflow to ensure no crashes, algorithm correctness
not tested here."""
with tempfile.TemporaryDirectory() as tmpdirname:
lockfile = os.path.join(tmpdirname, "multiprocess_lock")
Path(lockfile).touch()
params = {
"training_data_path": os.path.join(
curr_dir, "test_data/discrete_action/cartpole_training.json.bz2"
),
"eval_data_path": os.path.join(
curr_dir, "test_data/discrete_action/cartpole_eval.json.bz2"
),
"state_norm_data_path": os.path.join(
curr_dir, "test_data/discrete_action/cartpole_norm.json"
),
"model_output_path": tmpdirname,
"use_gpu": use_gpu,
"use_all_avail_gpus": use_all_avail_gpus,
"init_method": "file://" + lockfile,
"num_nodes": 1,
"node_index": 0,
"actions": ["0", "1"],
"epochs": 1,
"rl": {},
"rainbow": {"double_q_learning": False, "dueling_architecture": False},
"training": {"minibatch_size": 128},
}
dqn_workflow.main(params)
predictor_files = glob.glob(tmpdirname + "/predictor_*.c2")
assert len(predictor_files) == 1, "Somehow created two predictor files!"
predictor = DQNPredictor.load(predictor_files[0], "minidb")
test_float_state_features = [{"0": 1.0, "1": 1.0, "2": 1.0, "3": 1.0}]
q_values = predictor.predict(test_float_state_features)
assert len(q_values[0].keys()) == 2
def test_read_c2_model_from_file(self):
"""Test reading output caffe2 model from file and using it for inference."""
path = os.path.join(curr_dir, "test_data/discrete_action/example_predictor.c2")
predictor = DQNPredictor.load(path, "minidb")
test_float_state_features = [{"0": 1.0, "1": 1.0, "2": 1.0, "3": 1.0}]
q_values = predictor.predict(test_float_state_features)
assert len(q_values[0].keys()) == 2
def main(model_path):
predictor = DQNPredictor.load(model_path, "minidb", int_features=False)
env = OpenAIGymEnvironment(gymenv=ENV)
avg_rewards, avg_discounted_rewards = env.run_ep_n_times(AVG_OVER_NUM_EPS,
predictor,
test=True)
logger.info(
"Achieved an average reward score of {} over {} evaluations.".format(
avg_rewards, AVG_OVER_NUM_EPS))
def main(model_path):
predictor = DQNPredictor.load(model_path, "minidb", int_features=False)
env = OpenAIGymEnvironment(gymenv=ENV)
avg_rewards, avg_discounted_rewards = env.run_ep_n_times(
AVG_OVER_NUM_EPS, predictor, test=True
)
logger.info(
"Achieved an average reward score of {} over {} evaluations.".format(
avg_rewards, AVG_OVER_NUM_EPS
)
)