def track(env, terminal, done, policy):
# episode done
if terminal or done:
# save all data
utils.save_as_pickle(env.world.episode_info,
None,
'rp_ep{}_{}'.format(env.world.position_index,
policy),
force_save=True)
# reset
env.world.episode_info = utils.reset_tracking_info()
env.world.episode_info['a1'].append(
env.world.positions[env.world.position_index + 1][0].tolist())
env.world.episode_info['a2'].append(
env.world.positions[env.world.position_index + 1][1].tolist())
env.world.episode_info['goal'] = env.world.positions[
env.world.position_index + 1][-1].tolist()
env.world.episode_info['random_positions'] = env.world.positions[
env.world.position_index + 1]
return
a1_pos = utils.get_position(env, 'a1').tolist()
a2_pos = utils.get_position(env, 'a2').tolist()
goal_pos = utils.get_position(env, 'goal').tolist()
env.world.episode_info['a1'].append(a1_pos)
env.world.episode_info['a2'].append(a2_pos)
env.world.episode_info['goal'] = goal_pos
def pre_process(leaf_size=40,
processed_data=r'./proc_data.tree',
lookup=r'./id_star.lkp'):
"""
Function to pre process the stars data into a kd tree and save the data
Args:
leaf_size: The number of points at which the algorithm switches over to brute-force. Has to be positive
processed_data: The location where the processed data is to be stored as a pickle file
lookup: The location where the data id to object lookup is stored
Returns:
No return
"""
utils.skip_line(1)
stars = []
idx = 0
id_star_dct = {}
for line in sys.stdin:
star_name, x, y, z = utils.parse_line(line)
stars.append([x, y, z])
id_star_dct[idx] = star(x, y, z, star_name)
idx += 1
# construct a kd-tree
tree = spatial.KDTree(stars, leafsize=leaf_size)
utils.save_as_pickle(data=tree, pickle_file=processed_data)
utils.save_as_pickle(data=id_star_dct, pickle_file=lookup)
def calculate_user_preferences(user_id_list: str=None, start: str=Config.train_start, end:str=Config.train_end, save_path: str='./dev_recommendation_sources/'):
if user_id_list == 'dev':
user_id_list = load_pickle(Config.dev_user_list)
elif user_id_list == 'test':
user_id_list = load_pickle(Config.test_user_list)
elif isinstance(user_id_list, str): # 파일 경로를 입력할 경우
user_id_list = load_pickle(user_id_list)
print('Load calculating tools...', end='\t')
try:
user_time_read = load_user_time_read(Config.user_time_read)
post_id_encoder = PostIdEncoder(Config.encodings_root)
tfidf_generator = TFIDFGenerator(Config.tfidf_root)
except: # When import in Jupyter Notebook
user_time_read = load_user_time_read("../preprocessed/user_time_read.json")
post_id_encoder = PostIdEncoder("../encodings/")
tfidf_generator = TFIDFGenerator("../tfidf")
print('loaded!')
post_meta_id = [] # user_id_list의 유저들이 조회한 모든 글
posts_raw = [] # 유저들의 로그에서 등장한 모든 글을 담을 리스트
user_preferences_raw = [] # 유저들의 feature 벡터를 담을 리스트
for user_id in tqdm(user_id_list, desc=f'User Preference Extraction ({start}-{end})'):
# 설정한 구간에 대한 해당 유저의 로그
history = filter_read_by_time(user_time_read[user_id], start, end)
history = squeeze(list(map(lambda x: x[-1], history)))
# 유저 로그로부터 TF-IDF 행렬 생성
user_tfidf = tfidf_generator.generate(post_id_encoder.transform(history), drop_id=False) #
# TF-IDF 행렬로부터 유저 feature 벡터를 생성
preference = sparse.csr_matrix(user_tfidf.iloc[:, 1:].values.sum(axis=0)[:, np.newaxis]) # post_meta_id 컬럼을 제외한 뒤 summation
user_tfidf = user_tfidf.groupby('post_meta_id').first().reset_index() # faster than drop_duplicates()
user_tfidf = user_tfidf.loc[~user_tfidf['post_meta_id'].isin(post_meta_id), :]
if len(user_tfidf) > 0:
post_meta_id.extend(user_tfidf['post_meta_id'].tolist())
posts_raw.append(sparse.csr_matrix(user_tfidf.iloc[:, 1:])) # post_meta_id 컬럼을 제외하고 append -> post_meta_id 리스트를 개별적으로 생성하므로 불필요
user_preferences_raw.append(preference)
print('Postprocessing...', end='\t')
posts = sparse.vstack(posts_raw)
user_preferences = sparse.hstack(user_preferences_raw)
idf = np.array(np.log(tfidf_generator.DF.values.squeeze()) - np.log((posts != 0).sum(axis=0) + 1e-4)) # 1e-4: to prevent ZeroDivisionError
recommend_output = (posts.multiply(idf)).dot(user_preferences)
print('finished!')
if save_path:
interval = f'{start}-{end}'
if interval not in os.listdir(save_path):
os.mkdir(os.path.join(save_path, interval))
save_path = os.path.join(save_path, interval)
sparse.save_npz(os.path.join(save_path, f'recommend_output.npz'), recommend_output)
sparse.save_npz(os.path.join(save_path, f'user_preferences.npz'), user_preferences)
np.save(os.path.join(save_path, f'idf.npy'), idf)
sparse.save_npz(os.path.join(save_path, f'posts.npz'), posts)
save_as_pickle(os.path.join(save_path, f'post_meta_id.pkl'), post_meta_id)
print(f'Saved successfully in {save_path}😎')
else:
return recommend_output, user_preferences, idf, posts, post_meta_id
def run_switch_ssm_fitness_simulations(input_fnames, output_fname,
sim_params, sim_label):
"""
Run switch SSM fitness simulations.
Compare the results of different growth policies.
"""
print "running switch ssm fitness simulations..."
fitness_params_fname = input_fnames[0]
ssm_params_fname = input_fnames[1]
params = simulation.load_params(fitness_params_fname)
model_params = simulation.load_params(ssm_params_fname)
params.update(model_params)
all_policies = OrderedDict()
all_policies["Posterior predictive"] = policies.posterior_pred_policy
all_policies["Plastic"] = policies.plastic_growth_policy
all_policies["Random"] = policies.rand_growth_policy
all_policies["Glucose-only"] = policies.glu_only_growth_policy
all_policies["Posterior pred. (BH)"] = policies.bh_particle_filter_policy
all_policies["Random (BH)"] = policies.bh_rand_growth_policy
# fixed parameters for all simulations
p_init_output = params["p_init_output"]
time_obj = time_unit.Time(params["t_start"],
params["t_end"],
step_size=params["step_size"])
# include model parameters in into simulation parameters set
# so that the policies that need to run the model (like the
# posterior predictive policy) can access it
params.update(sim_params)
# include list of policies we ran
params["policies"] = all_policies.keys()
# setting of probabilities for data-generating SSM model
p_switch_to_switch = params["p_switch_to_switch"]
p_noswitch_to_switch = params["p_noswitch_to_switch"]
nutr_labels = params["nutr_labels"]
nutr_growth_rates = params["nutr_growth_rates"]
out_trans_mat1 = params["out_trans_mat1"]
out_trans_mat2 = params["out_trans_mat2"]
def nutrient_simulator(time_obj):
return nutrient_simulators.ssm_nutrient_simulator(time_obj,
out_trans_mat1=out_trans_mat1,
out_trans_mat2=out_trans_mat2,
p_switch_to_switch=p_switch_to_switch,
p_noswitch_to_switch=p_noswitch_to_switch,
p_init_output=p_init_output)
# simulate mixed sugar environment
env_obj = \
env.MixedDiscEnvironment(nutr_labels,
nutrient_simulator,
nutr_growth_rates=nutr_growth_rates)
fitness_obj = fitness.FitnessSim(all_policies, env_obj, params)
#sim_results = {}
sim_results = fitness_obj.simulate(time_obj)
final_results = {"sim_params": sim_params,
"sim_results": sim_results}
utils.save_as_pickle(output_fname, sim_results,
extra={"params": params})
def merge_switch_ssm_fitness_sims(input_fnames, output_fname):
"""
Combine all the switch SSM fitness simulations into a single pickle file.
"""
### combine all the simulations into one pickle file.
all_sim_data = OrderedDict()
for fname in input_fnames:
sim_name = os.path.basename(fname).split(".data")[0]
curr_sim_data = utils.load_pickle(fname)
all_sim_data[sim_name] = curr_sim_data
extra = {}
utils.save_as_pickle(output_fname, all_sim_data, extra)
def run_bet_hedge_sims_2_nutr(param_fname, output_fname, label):
"""
Running bet hedging simulations for two nutrients.
"""
print "running bet hedge simulations (2 nutrients)..."
# simulate two nutrient model
params = simulation.load_params(param_fname)
print "params: ", params
data = {}
extra = {"params": params}
# run bet hedging simulation
###
### need skeleton code here
###
# save results as a pickle file
utils.save_as_pickle(output_fname, data, extra)
def _save_info(self):
self.data = [
self.pos_queries, self.test_pos, self.obs, self.actions,
self.q_vals, self.argmaxs, self.swapped_obs, self.swapped_actions,
self.swapped_q_vals, self.swapped_argmaxs
]
self.names = [
"pos_queries", "test_positions", "obs", "actions", "q_vals",
"argmaxs", "swapped_obs", "swapped_actions", "swapped_q_vals",
"swapped_argmaxs"
]
assert len(self.data) == len(
self.names
), "Number of data to be saved must be matched with the number of names to save data under!"
print("Saving under filename: {} this information: {}".format(
self.filename, self.names))
# Creating directory if it doesn't exist
os.makedirs(os.path.dirname(self.filename), exist_ok=True)
# plt.show()
self.info = utils.save_as_pickle(self.data, self.names, self.filename)
def load_train_labels():
"""
loads the train labels as a numpy array
:return:
"""
config = get_config()
with open(config['image_paths']['train_labels'], newline='') as csv_file:
reader = csv.DictReader(
csv_file,
delimiter=',',
)
labels = np.asarray([row['label'] for row in reader])
label_dict = get_label_dict()
return np.asarray(
[label_dict[label_string] for label_string in labels])
if __name__ == '__main__':
config = get_config()
train_images_np = transform_images(
config['image_paths']['train_images_raw'])
save_as_pickle(train_images_np,
config['image_paths']['train_images_pickle'])
test_images_np = transform_images(config['image_paths']['test_images_raw'])
save_as_pickle(test_images_np, config['image_paths']['test_images_pickle'])
# load_train_labels()
# loaded = load_pickle( config['image_paths']['train_images_pickle'] )
# print(loaded.shape)
def get_save_df(root_dir: str = Config.raw_dir, vocab_path: str = VOCAB_PATH):
metadata = load_raw('metadata', root_dir)
vocab = load_pickle(vocab_path)
df = get_df(metadata, vocab)
save_as_pickle(df.values.squeeze().tolist(),
'../tfidf/df_vocab7000_aggregation.pkl')
