def sample_open(sample_idxs, col_features, col_labels, p_dic):
tool_fn.progress_bar(0, 1, prefix='>>> Sample extracting...')
feature_samples = os.listdir(p_dic.get('sample_data_dir') + '/feature')
label_samples = os.listdir(p_dic.get('sample_data_dir') + '/label')
feature_samples.sort()
label_samples.sort()
feature = np.empty([0, len(col_features)], dtype=np.float32)
label = np.empty([0, len(col_labels)], dtype=np.float32)
loads = 0
for sample_idx in sample_idxs:
feature_csv_path = p_dic.get(
'sample_data_dir') + '/feature/' + feature_samples[sample_idx]
label_csv_path = p_dic.get(
'sample_data_dir') + '/label/' + label_samples[sample_idx]
feature_part = pd.read_csv(feature_csv_path,
dtype=np.float32,
delimiter=',',
error_bad_lines=False,
index_col=0).values
label_part = pd.read_csv(label_csv_path,
dtype=np.float32,
delimiter=',',
error_bad_lines=False,
index_col=0).values
feature = np.concatenate((feature, feature_part))
label = np.concatenate((label, label_part))
loads += 1
tool_fn.progress_bar(loads,
len(sample_idxs),
prefix='>>> Sample extracting...')
print("Sample extracting done")
return feature, label
def main():
p_dic = getattr(conf.dic.path_dic, FLAGS.env_name)
c_dic = getattr(conf.dic.col_dic, FLAGS.env_name)
s_dic = getattr(conf.dic.space_dic, FLAGS.env_name)
cols = {
'ORIGIN_F': getattr(conf.list.preproc.origin, FLAGS.env_name),
'PREPROC_F': getattr(conf.list.preproc.preproc, FLAGS.env_name)
}
origin_f = [c_dic[feature] for feature in cols.get('ORIGIN_F')]
origins = os.listdir(p_dic.get('origin_data_dir'))
origins.sort()
print('Extract original data from dir: ' + p_dic.get('origin_data_dir'))
tool_fn.progress_bar(0, 1, prefix='>>> Extracting...')
for origin in range(len(origins)):
csv_path = p_dic.get('origin_data_dir') + '/' + origins[origin]
origin_df_part = pd.read_csv(csv_path, delimiter=',', error_bad_lines=False, usecols=origin_f)[origin_f]
origin_df_part = origin_df_part.loc[origin_df_part.index % 4 == 3]
if origin == 0:
origin_df = origin_df_part
else:
origin_df = pd.concat([origin_df, origin_df_part])
tool_fn.progress_bar(origin + 1, len(origins), prefix='>>> Extracting...')
striped_date = pd.Series(origin_df['Date/Time'].str.strip())
splited_date = pd.DataFrame(striped_date.str.split(r'[/ :]').tolist(), columns = ['Month', 'Day', '', 'Hour', '', ''])
splited_date = splited_date.drop([''], axis=1)
splited_date.reset_index(drop=True, inplace=True)
origin_df = origin_df.drop(['Date/Time'], axis=1)
origin_df.reset_index(drop=True, inplace=True)
origin_df = pd.concat([splited_date, origin_df], axis=1)
if not os.path.exists(p_dic.get('preproc_data_dir')):
os.makedirs(p_dic.get('preproc_data_dir'))
print('Preprocess to dir: ' + p_dic.get('preproc_data_dir'))
tool_fn.progress_bar(0, 1, prefix='>>> Preprocessing...')
f_space = np.array([s_dic[feature] for feature in cols.get('PREPROC_F')], dtype=np.float32).T
origin_df_size = origin_df.shape[0]
for split in range(FLAGS.frac_num):
preprocessed_df = origin_df.head(np.ceil(origin_df_size / FLAGS.frac_num).astype(int)).apply(pd.to_numeric)
preprocessed_df = tool_fn.normalize(preprocessed_df, f_space[0], f_space[1])
origin_df.drop(preprocessed_df.index)
preprocessed_df.to_csv(p_dic.get('preproc_data_dir') + '/' + FLAGS.preproc_name + '_' + str(split) + '.csv')
tool_fn.progress_bar(split + 1, FLAGS.frac_num, prefix='>>> Preprocessing...')
def learn(FLAGS, env, agent):
render = False
sd = 1
print('\n\n>> Agent training start')
for e in range(FLAGS.num_episodes):
obs = env.reset()
ep_reward = 0
# rss : 추후 env 쪽으로 넘김
# env에 epsode_render 메서드 만들어서 거기에 포함
rss = []
start = time.time()
for j in range(FLAGS.num_steps):
if render:
env.render()
action = agent.choose_action(obs)
action = np.random.normal(loc=action, scale=sd)
action = np.clip(action, env.action_space.low,
env.action_space.high)
next_obs, reward, done, info = env.step(action[0])
next_obs = np.expand_dims(next_obs, axis=0)
reward = np.expand_dims(reward, axis=0)
# rss : 추후 env 쪽으로 넘김
if hasattr(env, 'envfn'):
rss.append(env.rs.reshape(-1))
ep_reward += np.sum(reward)
end = time.time()
tool_fn.progress_bar(
j + 1,
FLAGS.num_steps,
prefix='>> EP:%i' % e,
suffix='(STEP:%i, R:%.2f, FPS:%.0f, SD:%.2f)' %
(j, ep_reward / (j + 1), (j + 1) / (end - start), sd))
if j == FLAGS.num_steps - 1:
agent.perceive(obs,
action,
reward,
next_obs,
episode=e,
ep_reward=ep_reward)
if ep_reward > 10000:
render = True
break
else:
agent.perceive(obs, action, reward, next_obs)
obs = next_obs
if sd >= 0.1:
# sd *= .995
sd *= .9
# 플로팅 코드, 추후 rss 변수와 함께 env episode_render 메서드로 옮길 것 #
if hasattr(env, 'envfn'):
if not os.path.exists(
env.p_dic.get('agent_graph_dir') + '/' + str(e)):
os.makedirs(env.p_dic.get('agent_graph_dir') + '/' + str(e))
rss = np.array(rss)
# indexing = np.arange(FLAGS.num_steps)
# interval = indexing % (FLAGS.num_steps // 100) == 0
# for rs in range(len(env.cols.get('agent_rs'))):
# plt.title(env.cols.get('agent_rs')[rs] + ' (instant)')
# plt.plot(indexing[interval], rss[interval][:, rs])
# plt.figure(num=1, figsize=(6, 6), dpi=2000, facecolor='white')
# plt.savefig(env.p_dic.get('agent_graph_dir') + '/' + str(e) + '/' + env.cols.get('agent_rs')[rs] + '_inst.png', bbox_inches='tight')
# plt.close()
# 데이터센터만, 일단위 플로팅 #
rss_ = []
for i in range(rss.shape[0]):
if ((i % 24) == 0 and (i + 24) <= rss.shape[0]):
sum = 0
for j in range(24):
sum += rss[i + j]
rss_.append(sum / 24)
rss = np.array(rss_)
indexing = np.arange(rss.shape[0])
for rs in range(len(env.cols.get('agent_rs'))):
plt.title(env.cols.get('agent_rs')[rs] + ' (instant)_day')
plt.plot(indexing, rss[:, rs])
plt.figure(num=1, figsize=(6, 6), dpi=2000, facecolor='white')
plt.savefig(env.p_dic.get('agent_graph_dir') + '/' + str(e) +
'/' + env.cols.get('agent_rs')[rs] +
'_inst_day.png',
bbox_inches='tight')
plt.close()
# 일단위 플로팅 끝 #
if hasattr(env, 'envfn'):
env.sess.close()
agent.sess.close()
def main():
p_dic = getattr(conf.dic.path_dic, FLAGS.env_name)
c_dic = getattr(conf.dic.col_dic, FLAGS.env_name)
cols = {
'ENV_F': getattr(conf.list.envfn.feature, FLAGS.env_name),
'ENV_L': getattr(conf.list.envfn.label, FLAGS.env_name),
'AGENT_IVS': getattr(conf.list.agent.iv_state, FLAGS.env_name)
}
features = [c_dic[feature] for feature in cols.get('ENV_F')]
labels = [c_dic[label] for label in cols.get('ENV_L')]
independents = [c_dic[ivs] for ivs in cols.get('AGENT_IVS')]
origins = os.listdir(p_dic.get('origin_data_dir'))
origins.sort()
print('Extract original data from dir: ' + p_dic.get('origin_data_dir'))
tool_fn.progress_bar(0, 1, prefix='>>> Extracting...')
for origin in range(len(origins)):
csv_path = p_dic.get('origin_data_dir') + '/' + origins[origin]
origin_feature_df_part = pd.read_csv(csv_path,
dtype=np.float32,
delimiter=',',
error_bad_lines=False,
usecols=features)[features]
origin_label_df_part = pd.read_csv(csv_path,
dtype=np.float32,
delimiter=',',
error_bad_lines=False,
usecols=labels)[labels]
origin_ivs_df_part = pd.read_csv(csv_path,
dtype=np.float32,
delimiter=',',
error_bad_lines=False,
usecols=independents)[independents]
origin_feature_df_part = origin_feature_df_part.loc[
origin_feature_df_part.index % 4 == 3].iloc[:-1]
origin_label_df_part = origin_label_df_part.loc[
origin_label_df_part.index % 4 == 3].iloc[1:]
origin_ivs_df_part = origin_ivs_df_part.loc[origin_ivs_df_part.index %
4 == 3].iloc[:-1]
if origin == 0:
origin_feature_df = origin_feature_df_part
origin_label_df = origin_label_df_part
origin_ivs_df = origin_ivs_df_part
else:
origin_feature_df = pd.concat(
[origin_feature_df, origin_feature_df_part])
origin_label_df = pd.concat(
[origin_label_df, origin_label_df_part])
origin_ivs_df = pd.concat([origin_ivs_df, origin_ivs_df_part])
tool_fn.progress_bar(origin + 1,
len(origins),
prefix='>>> Extracting...')
origin_feature_df = origin_feature_df.reset_index(drop=True)
origin_label_df = origin_label_df.reset_index(drop=True)
if not os.path.exists(p_dic.get('sample_data_dir') + '/feature'):
os.makedirs(p_dic.get('sample_data_dir') + '/feature')
if not os.path.exists(p_dic.get('sample_data_dir') + '/label'):
os.makedirs(p_dic.get('sample_data_dir') + '/label')
if not os.path.exists(p_dic.get('sample_data_dir') + '/ivs'):
os.makedirs(p_dic.get('sample_data_dir') + '/ivs')
print('Sample to dir: ' + p_dic.get('sample_data_dir'))
tool_fn.progress_bar(0, 1, prefix='>>> Sampling...')
for split in range(FLAGS.frac_num):
sampled_feature_df = origin_feature_df.sample(frac=1. / FLAGS.frac_num,
random_state=sample_seed)
origin_feature_df.drop(sampled_feature_df.index)
sampled_feature_df.to_csv(
p_dic.get('sample_data_dir') + '/feature/' + FLAGS.sample_name +
'_feature_' + str(split) + '.csv')
sampled_label_df = origin_label_df.iloc[sampled_feature_df.index]
sampled_label_df.to_csv(
p_dic.get('sample_data_dir') + '/label/' + FLAGS.sample_name +
'_label_' + str(split) + '.csv')
tool_fn.progress_bar(split + 1,
FLAGS.frac_num,
prefix='>>> Sampling...')
origin_ivs_df.to_csv(
p_dic.get('sample_data_dir') + '/ivs/' + 'extracted_ivs' + '.csv')
def shuffle_n_batch(data_tuple, data_length):
tool_fn.progress_bar(0, 1, prefix='>>> Shuffling and batch making...')
dataset = tf.data.Dataset.from_tensor_slices(data_tuple)
batch = dataset.shuffle(data_length).repeat().batch(FLAGS.batch_size)
tool_fn.progress_bar(1, 1, prefix='>>> Shuffling and batch making...')
return batch