def train(config, expe_name, trial_id, beta_default):
"""
Main training script. Defines the cost function and epidemiological model to form the environment model.
Parameters
----------
config: str
Identifier of the configuration parameters.
expe_name: str
String to describe your experiment, will be used for the logging directory.
trial_id: int
Trial identifier for logging path.
beta_default: float
Additional parameter for our experiments.
"""
# Get the configuration
params = get_params(config_id=config, expe_name=expe_name)
params.update(trial_id=trial_id)
params['cost_params']['beta_default'] = beta_default
# Set seeds
set_seeds(params['seed'])
# Get the epidemiological model
model = get_model(model_id=params['model_id'],
params=params['model_params'])
# Update cost function params
params['cost_params']['N_region'] = int(
model.pop_sizes[params['model_params']['region']])
params['cost_params']['N_country'] = int(
sum(list(model.pop_sizes.values())))
# Get cost function
cost_function = get_cost_function(cost_function_id=params['cost_id'],
params=params['cost_params'])
# Create the optimization problem as a Gym-like environment
env = get_env(env_id=params['env_id'],
cost_function=cost_function,
model=model,
simulation_horizon=params['simulation_horizon'],
seed=params['seed'])
# Setup logdir
params = get_logdir(params=params)
# Get DQN algorithm parameterized by beta
algorithm = get_algorithm(algo_id=params['algo_id'],
env=env,
params=params)
# Run the training loop
algorithm.learn(num_train_steps=params['num_train_steps'])
def setup_for_replay(folder,
seed=np.random.randint(1e6),
deterministic_model=False):
from epidemioptim.environments.models import get_model
from epidemioptim.environments.cost_functions import get_cost_function
from epidemioptim.environments.gym_envs import get_env
from epidemioptim.optimization import get_algorithm
# print('Replaying: ', folder)
with open(folder + 'params.json', 'r') as f:
params = json.load(f)
if deterministic_model:
params['model_params']['stochastic'] = False
params[
'logdir'] = None #get_repo_path() + 'data/results/experiments' + params['logdir'].split('EpidemicDiscrete-v0')[1]
model = get_model(model_id=params['model_id'],
params=params['model_params'])
# update reward params
params['cost_params']['N_region'] = int(
model.pop_sizes[params['model_params']['region']])
params['cost_params']['N_country'] = int(
np.sum(list(model.pop_sizes.values())))
set_seeds(seed)
cost_function = get_cost_function(cost_function_id=params['cost_id'],
params=params['cost_params'])
# Form the Gym-like environment
env = get_env(env_id=params['env_id'],
cost_function=cost_function,
model=model,
simulation_horizon=params['simulation_horizon'],
seed=seed)
# Get DQN algorithm parameterized by beta
algorithm = get_algorithm(algo_id=params['algo_id'],
env=env,
params=params)
if params['algo_id'] == 'NSGAII':
algorithm.load_model(folder + 'res_eval.pk')
else:
algorithm.load_model(folder + 'models/best_model.cp')
return algorithm, cost_function, env, params
#model = get_model(model_id='prague_seirah', params=dict(region=region, stochastic=stochastic))
model = get_model(model_id='sqeir', params=dict(stochastic=stochastic))
#N_region = model.pop_sizes[region]
#N_country = np.sum(list(model.pop_sizes.values()))
#ratio_death_to_R = 0.005
ratio_death_to_R = 0.02
# cost_function = get_cost_function(cost_function_id='multi_cost_death_gdp_controllable', params=dict(N_region=N_region,
# N_country=N_country,
# ratio_death_to_R=ratio_death_to_R)
# )
cost_function = get_cost_function(
cost_function_id='korea_multi_cost_death_gdp_controllable',
params=dict(ratio_death_to_R=ratio_death_to_R))
env = get_env(env_id='KoreaEpidemicDiscrete-v0',
params=dict(
cost_function=cost_function,
model=model,
simulation_horizon=simulation_horizon)) #params 이거 맞나?
env.reset()
actions = np.random.choice(
[0, 1, 2, 3, 4, 5],
size=53) # need to change to have 6 actions 0, 1, 2, 3, 4, 5
t = 0
r = 0
done = False
from epidemioptim.environments.gym_envs import get_env
if __name__ == '__main__':
simulation_horizon = 364
stochastic = True
region = 'IDF'
model = get_model(model_id='prague_seirah', params=dict(region=region,
stochastic=stochastic))
N_region = model.pop_sizes[region]
N_country = np.sum(list(model.pop_sizes.values()))
ratio_death_to_R = 0.005
cost_function = get_cost_function(cost_function_id='multi_cost_death_gdp_controllable', params=dict(N_region=N_region,
N_country=N_country,
ratio_death_to_R=ratio_death_to_R)
)
env = get_env(env_id='EpidemicDiscrete-v0',
cost_function=cost_function,
model=model,
simulation_horizon=simulation_horizon)
all_stats = []
for i_loop in range(10):
env.reset()
# STEP 4: Initialize the scripted sequence of actions
# This is useful only here for simulation, learning algorithm will replace this.
# * use a action_type corresponding to your env (on/off, sticky_one_week, switch or commit)
def setup_for_replay(folder,
seed=np.random.randint(1e6),
deterministic_model=False):
from epidemioptim.environments.models import get_model
from epidemioptim.environments.cost_functions import get_cost_function
from epidemioptim.environments.gym_envs import get_env
from epidemioptim.optimization import get_algorithm
print('Replaying: ', folder)
with open(folder + 'params.json', 'r') as f:
params = json.load(f)
if deterministic_model:
params['model_params']['stochastic'] = False
params[
'logdir'] = None #get_repo_path() + 'data/results/experiments' + params['logdir'].split('EpidemicDiscrete-v0')[1]
model = get_model(model_id=params['model_id'],
params=params['model_params'])
set_seeds(seed)
cost_function = get_cost_function(cost_function_id=params['cost_id'],
params=params['cost_params'])
# Form the Gym-like environment
env = get_env(env_id=params['env_id'],
cost_function=cost_function,
model=model,
simulation_horizon=params['simulation_horizon'],
seed=seed)
# Get DQN algorithm parameterized by beta
algorithm = get_algorithm(algo_id=params['algo_id'],
env=env,
params=params)
# if params['algo_id'] == 'NGSA':
# algorithm.load_model(folder + 'res_eval.pk')
# else:
algorithm.load_model(folder + 'models/best_model.cp')
return algorithm, cost_function, env, params
# Compute the real SIDEP incidence from a CSV file
# def get_incidence():
# PATH_TO_DATA = get_repo_path() + '/data/jane_model_data/inci.csv'
# data = pd.read_csv(PATH_TO_DATA, delimiter=";")
# true_data = pd.DataFrame(data['jour'])
# true_data['T']=data['P']
# week_grouped = true_data.groupby(['jour'])['T'].sum()/2
# data = week_grouped.values.tolist()
# #pre = np.zeros(132).tolist()
# #post = np.zeros(214).tolist()
# #pre.extend(data)
# #pre.extend(post)
# yhat = scipy.signal.savgol_filter(data, 53, 3)
# # for i in range(455, 731):
# # yhat[i] = 0
# return yhat
data['stats_run'] = stats_run
data['title'] = 'Death Cost: {}, Aggregated Cost: {:.2f}'.format(
int(cumulative_death[-1]),
np.sum(self.history['aggregated_costs']))
return data
if __name__ == '__main__':
from epidemioptim.utils import plot_stats, plot_preds
from epidemioptim.environments.cost_functions import get_cost_function
from epidemioptim.environments.models import get_model
simulation_horizon = 360
model = get_model(model_id='heffernan_model',
params=dict(stochastic=False))
cost_func = get_cost_function(cost_function_id='one_cost_death_toll',
params=dict(ratio_death_to_R=0.0001))
env = gym.make('EpidemicVaccination-v0',
cost_function=cost_func,
model=model,
simulation_horizon=simulation_horizon)
env.reset()
model_states = []
env.model.current_state, env.model.current_internal_params = env.initialize_model_for_vaccine(
)
# Actions
actions = ([0, 0, 1], [0, 0, 1], [0, 1, 1], [0, 1, 1], [0, 0, 1],
[0, 1, 1], [0, 1, 1], [0, 1, 1], [0, 1, 1], [0, 1,
1], [0, 1,
1], [0, 1, 1])
#actions = ([0,0,1], [0,0,1], [0,0,1], [0,1,1],[1,0,1], [1,1,1], [1,1,1], [1,1,1], [1,1,1], [0,1,1], [1,1,1], [1,1,0])