def excess_demand(self, signal, time):
"""Excess demand function for moving averages based heuristic (group 1)."""
m_window_average = moving_average(signal, time, self.m)
n_window_average = moving_average(signal, time, self.n)
price_change = percent_change(m_window_average, n_window_average)
price_change_strengths = self.ppc_signals.grade(price_change)
investor_signal_strengths = self.investor_signals.centers
result = np.dot(investor_signal_strengths, price_change_strengths)
result /= np.sum(price_change_strengths)
return result
def swa_train(model, swa_model, train_iter, valid_iter, optimizer, criterion, pretrain_epochs, swa_epochs, swa_lr, cycle_length, device, writer, cpt_filename):
swa_n = 1
swa_model.load_state_dict(copy.deepcopy(model.state_dict()))
utils.save_checkpoint(
cpt_directory,
1,
'{}-swa-{:2.4f}-{:03d}-{}'.format(date, swa_lr, cycle_length, cpt_filename),
state_dict=model.state_dict(),
swa_state_dict=swa_model.state_dict(),
swa_n=swa_n,
optimizer=optimizer.state_dict()
)
for e in range(swa_epochs):
epoch = e + pretrain_epochs
time_ep = time.time()
lr = utils.schedule(epoch, cycle_length, lr_init, swa_lr)
utils.adjust_learning_rate(optimizer, lr)
train_res = utils.train_epoch(model, train_iter, optimizer, criterion, device)
valid_res = utils.evaluate(model, valid_iter, criterion, device)
utils.moving_average(swa_model, model, swa_n)
swa_n += 1
utils.bn_update(train_iter, swa_model)
swa_res = utils.evaluate(swa_model, valid_iter, criterion, device)
time_ep = time.time() - time_ep
values = [epoch + 1, lr, swa_lr, cycle_length, train_res['loss'], valid_res['loss'], swa_res['loss'], None, None, time_ep]
writer.writerow(values)
table = tabulate.tabulate([values], columns, tablefmt='simple', floatfmt='8.4f')
if epoch % 20 == 0:
table = table.split('\n')
table = '\n'.join([table[1]] + table)
else:
table = table.split('\n')[2]
print(table)
utils.save_checkpoint(
cpt_directory,
epoch + 1,
'{}-swa-{:2.4f}-{:03d}-{}'.format(date, swa_lr, cycle_length, cpt_filename),
state_dict=model.state_dict(),
swa_state_dict=swa_model.state_dict(),
swa_n=swa_n,
optimizer=optimizer.state_dict()
)
def plot_scores(self, scores: list, window=50) -> None:
if window:
scores = moving_average(scores, window_size=window)
plt.plot(range(len(scores)), scores)
plt.xlabel("Episode")
plt.ylabel("Episode reward (smoothed {})".format(window))
plt.show()
def show_exec_usage(jobs, save_path):
# get job completion time
jct = int(np.ceil(np.max([job.finish_time for job in jobs])))
job_durations = [job.finish_time - job.start_time for job in jobs]
exec_occupation = np.zeros(jct)
exec_limit = np.ones(jct) * args.exec_cap
num_jobs_in_sys = np.zeros(jct)
for job in jobs:
for stage in job.stages:
for task in stage.tasks:
exec_occupation[int(task.start_time):int(task.finish_time
)] += 1
num_jobs_in_sys[int(job.start_time):int(job.finish_time)] += 1
exec_usage = np.sum(exec_occupation) / np.sum(exec_limit)
fig = plt.figure()
plt.subplot(2, 1, 1)
plt.plot(utils.moving_average(exec_occupation,
10000)) # TODO: why set as 10000?
plt.ylabel('Number of busy executors')
plt.title('Executor usage: ' + str(exec_usage) + '\n Average JCT: ' +
str(np.mean(job_durations)))
plt.subplot(2, 1, 2)
plt.plot(num_jobs_in_sys)
plt.xlabel('Time (ms)')
plt.ylabel('Number of jobs in the system')
fig.savefig(save_path, dpi=100)
plt.close(fig)
def divideTrajectory(self, points):
"""
Subdivide a list of points.
points: (list of Point objects)
"""
self.traj_pts = []
for index in range(len(points) - 1):
num_subsamples = math.hypot(points[index+1].x - points[index].x,\
points[index+1].y - points[index].y) / self.TRAJECTORY_SLICE_SIZE
xs = np.linspace(points[index].x,
points[index + 1].x,
num=num_subsamples)
ys = np.linspace(points[index].y,
points[index + 1].y,
num=num_subsamples)
# Don't add the last point, because it will be added as the first point of
# the next segment.
self.traj_pts.extend(zip(xs, ys)[0:-1])
# Add final trajectory point.
self.traj_pts.append((points[-1].x, points[-1].y))
# Smooth the trajectory with a moving average.
# path_as_numpy_array = np.array(self.trajectory.points)
self.traj_pts = utils.moving_average(np.array(self.traj_pts))
# self.trajectory.points = [tuple(point) for point in self.smoothed_path]
# self.trajectory.update_distances()
# print "Loaded:", len(self.trajectory.points), "smoothed points"
rospy.loginfo('Finished dividing trajectory (and smoothing).')
def __str__(self):
self.update()
underlying_price = utils.moving_average(self.obj, 3)
if self.put:
option = "P"
else:
option = "C"
if self.pct_otm() < 0:
otm_str = str(self.pct_otm()) + " (itm)"
else:
otm_str = str(self.pct_otm())
return ("""
Position: {expiry} {obj} {strike}{call}
Underlying price: {underlying}
Option price: {price}
DTE: {dte}
pct_otm = {otm}
yield = {yields}
""".format(expiry=self.expiry,
obj=self.obj.info['symbol'],
strike=self.strike,
call=option,
underlying=underlying_price,
price=self.price,
dte=self.dte,
otm=otm_str,
yields=self.pct_yield()))
def pulse_demodulation(y, args):
'''
描述：脉冲解调
参数：信号，参数列表
返回：返回：0-1序列
'''
framerate = args.framerate
frequency = args.frequency
volume = args.volume
start_place = args.start_place
pulse_length = args.pulse_length
interval_0 = args.interval_0
interval_1 = args.interval_1
duration_pulse = round(pulse_length * framerate)
duration_0 = round(interval_0 * framerate)
duration_1 = round(interval_1 * framerate)
window = duration_pulse
seq = []
y = bandpass(y, framerate, frequency - 500, frequency + 500) #先滤波
y = FFT(y, window, framerate, frequency) #fourier变换
y = moving_average(y)
max_list = get_pulse_max(y, window, volume)
previous = 0
for i in range(len(max_list)):
duration = max_list[i] - previous - duration_pulse
if duration >= duration_0 * 0.9 and duration <= duration_0 * 1.1:
seq.append(0)
else:
seq.append(1)
previous = max_list[i]
return seq
def subplot(
self, all_tracks: List[Track], artists: List[str], smoothing: int
) -> plot.SubPlot: # type: ignore # pylint: disable=arguments-differ
days, listens_per_day = utils.listens_per_day(all_tracks)
return plot.SubPlot(*(plot.Graph(
x_values=days,
y_values=utils.moving_average(listens_per_day[artist], smoothing),
legend_label=artist,
plot_type="-",
) for artist in artists))
def plot_from_file(file_name,
param_name,
last_N=100,
color='blue',
limit_x=None,
limit_x_range=None,
range_y=None,
y_ticks=None):
metadata = load_pickle(file_name)
score = metadata[param_name]
mean, std = moving_average(score, last_N=last_N)
if limit_x is not None:
episodes = range(limit_x)
mean = mean[:limit_x]
std = std[:limit_x]
elif limit_x_range is not None:
episodes = metadata[limit_x_range]
else:
episodes = range(len(score))
mean, std = moving_average(score, last_N=last_N)
lower_bound = [a_i - 0.5 * b_i for a_i, b_i in zip(mean, std)]
upper_bound = [a_i + 0.5 * b_i for a_i, b_i in zip(mean, std)]
# plt.plot(episodes, score)
plt.fill_between(episodes,
lower_bound,
upper_bound,
facecolor=color,
alpha=0.5)
plt.plot(episodes, mean, color=color)
if range_y is not None:
plt.ylim(range_y)
if y_ticks is not None:
plt.yticks(np.arange(range_y[0], range_y[1] + 2 * y_ticks, y_ticks))
if limit_x_range is not None:
plt.xlabel(limit_x_range)
else:
plt.xlabel("episodes")
plt.ylabel(param_name)
def nearest_strike_by_pct_otm(
ticker, pct_otm, expiry
): # nearest OTM strike on a particular expiry; pct_otm = 0.03 := 3% OTM, pct_otm < 0 := ITM put
current_price = utils.moving_average(ticker, 3)
otm_strike_price = current_price * (1 - pct_otm)
strikes = ticker.option_chain(expiry).puts[
"strike"] # all strikes for a given expiry
dist_to_strike = list(
map(lambda x: round(abs(x - otm_strike_price), 3), strikes))
index = np.where(dist_to_strike == min(dist_to_strike))
# print(current_price)
# print(otm_strike_price)
return strikes[index[0][0]]
def pct_otm(
self
): # Puts: stock is $40, strike is $32, pct_otm = 0.20; opposite for Call
self.update()
underlying_price = utils.moving_average(self.obj, 3)
if self.put:
delta = underlying_price - self.strike
pct_otm = round(delta / underlying_price, 4)
else:
delta = self.strike - underlying_price
pct_otm = round(delta / underlying_price, 4)
return pct_otm
def __init__(self, model, settings):
self.settings = settings
self.X_scaler = model['scaler']
self.clf = model['clf']
self.y_start_stop = settings['y_start_stop']
self.scale = settings['scale']
self.orient = settings['orient']
self.pix_per_cell = settings['pix_per_cell']
self.cell_per_block = settings['cell_per_block']
self.spatial_size = settings['spatial_size']
self.hist_bins = settings['hist_bins']
self.color_space = settings['color_space']
self.smoother = utils.moving_average(10)
self.ypos_scales = [([400,480], 0.7), ([400,550], 1.0), ([400,600], 1.5), ([400,660], 2), ([400,660], 2.5)]
self.vehicles = []
def exp_learning_rate(self, X, Y, lr_start, lr_end, n_iterations, batch_size=100):
if not self._is_built: raise Exception("Model not built yet, use load or build to build model")
with self.sess:
# make pre processing a part of loading the batches instead
print("preprocessing images...")
X = self.sess.run(self.pre_process, {self.x: X})
print("...preprocessing done!")
# record training data
losses = []
learning_rates = []
lr = lr_start
c = pow(lr_end/lr_start, 1/n_iterations) # constant for increasing learning rate
N = X.shape[0]
perm = np.random.permutation(N)
for i in range(n_iterations):
#self.sess.run(tf.global_variables_initializer()) # trying remove if not working
batch_inds = perm[i*batch_size:batch_size*(i+1)]
feed_dict = {self.x: X[batch_inds], self.y: Y[batch_inds],
self.learning_rate: lr}
feed_dict = {self.x: X[0:batch_size], self.y: Y[0:batch_size],
self.learning_rate: lr}
# performs gradient descent and updates weights
_, train_loss, train_cost, train_acc = self.sess.run([self.optimizer,
self.loss,
self.cost,
self.accuracy],
feed_dict)
losses.append(train_loss)
print(train_loss)
learning_rates.append(lr)
lr = lr_start * pow(c, i)
print("learning rate:", self.learning_rate.eval({self.learning_rate: lr}))
if train_loss > 100:
break
losses = utils.moving_average(losses, n=3)
return losses, learning_rates[0:len(losses)]
def data_by_ticker_and_pct_otm(
ticker, pct_otm
): # return list of (expiry, nearest strike, put option cost, DTE, ROI) for each expiry in ticker.options
recent_price = utils.moving_average(ticker, 3)
print(
str(ticker.info['symbol']) + " 3-day MA: %s; OTM %s pct" %
(str(recent_price), str(pct_otm * 100)))
today = datetime.date.today()
data = []
for date in ticker.options:
strike = nearest_strike_by_pct_otm(ticker, pct_otm, date)
option_chain = ticker.option_chain(date).puts
option_price = option_chain.loc[option_chain['strike'] ==
strike]['lastPrice'].iloc[0]
dte = (utils.str_to_date(date) - today).days
roi = round(option_price / recent_price, 4)
data.append((date, strike, option_price, dte, roi))
return data
def main(argv):
learn_env = parse_params(argv)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
env = UnityEnvironment(file_name=learn_env['banana_location'])
brain_name = env.brain_names[0]
brain = env.brains[brain_name]
env_info = env.reset(train_mode=True)[brain_name]
# update hyperparameters with Unity Environment
learn_env["brain_name"] = brain_name
learn_env["brain"] = brain
learn_env["action_size"] = brain.vector_action_space_size
learn_env["state_size"] = len(env_info.vector_observations[0])
agent = DDQNPrioritizedAgent(learn_env)
scores = train_dqn(env, learn_env, agent)
label = "10 episode moving average score"
ma_window_size = 10
ma_scores = moving_average(scores,ma_window_size)
torch_model_fname = learn_env.get('model_file')
if(torch_model_fname):
torch.save(agent.qnetwork_local.state_dict(), torch_model_fname)
#plot results
plt.figure()
plt.title('Mean Score')
plt.xlabel("Episodes")
plt.ylabel("Mean Score")
episodes=range(len(ma_scores))
plt.plot(episodes, ma_scores, label=label)
plt.legend(loc='lower right')
f_name = learn_env.get('plt_file')
if (f_name):
plt.savefig(f_name)
plt.show()
else:
plt.show()
env.close()
def create_mAP_line(self,
dirpath,
videoname,
mAP,
Converter,
pred_list,
smooth_step=10,
to_file=False):
#TODO: create smoothed mAP lines for IMAGENET dataset, obsolete since now YOUTUBE dataset is used instead
#INPUT: dirpath contains all video frame for a given video
# smooth_step to smooth the mAP results using 10 previous & 10 after frame
#OUTPUT: generate a smoothed mAP line
#test sample dirpath: './data/VID_data/ILSVRC2015_VID_train_0000/ILSVRC2015_train_00005004/*'
filepath_list = glob.glob(dirpath)
filepath_list = sorted(filepath_list)
y_raw = []
dtoken = self.vid_dirtoken_dict[videoname]
for idx, filepath in enumerate(filepath_list):
if idx % 100 == 0:
print(idx, filepath)
ftoken = Converter.get_ftoken_from_vidname_and_idx(
self.vid_dirtoken_dict, videoname, idx)
for pred in pred_list:
y_raw.append(
mAP.score_from_file(filepath, Converter, pred,
self.anno_dict[dtoken][ftoken]))
y = utils.moving_average(y_raw)
if to_file == True:
#ftoken = converter.Converter.get_token_from_filepath(filepath)
#dtoken = converter.Converter.get_dirtoken_from_filetoken(ftoken)
pickle.dump(
y,
open(
namespace.DIRPATH_VID_mAP_LINES +
namespace.FILETEMPLATE_mAP_LINES.format(dtoken), 'wb'))
return y
def time_series_daily(label, window_size=7):
df = df_dict['covid-us']
x = df['date']
daily = daily_increase(df[label])
moving_avg = moving_average(daily, window_size)
trace1 = go.Bar(x=x, y=daily, name=f'Daily new {label}',
marker = dict(color = colors_bar[label],
line=dict(color=colors_bar[label],width=1.5),
opacity=0.2),
hovertemplate='%{x|%b %d, %Y} <br>Daily: %{y:-.0f}'
)
trace2 = go.Scatter(x=x, y=moving_avg,
name=f'Moving average in {window_size} days',
line={'width':3, 'color': colors_line[label]},
hovertemplate='Moving average: %{y:-.0f}'
)
title = f'Daily reported new Covid {label.lower()} in U.S. over time'
layout = dict(title=title,
yaxis_title=f'# of {label} per day',
xaxis_title='Date/Time',
font=dict(family="Courier New, monospace",
size=16),
hoverlabel=dict(
bgcolor="white",
font_size=16,
font_family="Rockwell"),
hovermode='x Unified',
legend=dict(
yanchor="top",
y=0.99,
xanchor="left",
x=0.01)
)
data = [trace1, trace2]
fig = dict(data=data, layout=layout)
return fig
def run_qlearning():
# const
num_agents = 5
num_episodes = 10000
# env
env = envs.debug.OneRoundDeterministicRewardMultiagentEnv(num_agents)
# algo
actions = list(itertools.product(range(2), repeat=num_agents))
discount = 1
explorationProb = .3
stepSize = .3
algo = algos.tabular.QLearningAlgorithm(actions, discount, explorationProb,
stepSize)
# train
rewards = []
for episode in range(num_episodes):
obs = env.reset()
action = algo.getAction(tuple(obs))
next_obs, reward, done, _ = env.step(action)
rewards.append(reward)
algo.incorporateFeedback(tuple(obs), action, reward, None)
# final weights
lines = sorted(algo.weights.items(), key=lambda (k, v): v, reverse=True)
print("state action value")
for line in lines:
print(line)
avg_rewards = utils.moving_average(rewards, num_episodes / 100)
plt.plot(range(len(avg_rewards)),
avg_rewards,
alpha=.5,
label="average rewards")
plt.legend(loc=8)
plt.show()
def time_series_state(plot_type='daily', state_name='Rhode Island', label='cases',):
# print(label, plot_type, state_name)
df = df_dict['covid-us-state']
df['state_code'] = df['state'].apply(lambda x: state_code_dict[x])
state_code = state_code_dict[state_name]
df_state = df[df.state_code == state_code]
state = state_name
df_state = df_state.sort_values(by='date')
df_state = pd.DataFrame(df_state, columns=df_state.columns)
x = df_state.date
y = df_state[label].values
if plot_type == 'daily':
window_size = 7
daily_cases = daily_increase(y)
moving_avg = moving_average(daily_cases, window_size)
trace_bar = go.Bar(x=x, y=daily_cases, name=f'Daily new {label}',
marker = dict(color = colors_bar[label],
line=dict(color=colors_bar[label],width=1.5),
opacity=0.2),
hovertemplate='Date: %{x|%A, %b %d, %Y} <br> Daily increase : %{y:.0f}'
)
trace_line = go.Scatter(
x=x,
y=moving_avg,
name=f'Moving average in {window_size} days',
line={'width':1.5, 'color': colors_line[label]},
hovertemplate='7 Day Avg. : %{y:.0f}')
title = f'Daily reported new Covid {label.lower()} in {state}'
layout = dict(title=title,
yaxis_title=f'# of {label} per day',
xaxis_title='Date/Time',
font=dict(family="Courier New, monospace",
size=16),
hoverlabel=dict(
bgcolor="white",
font_size=16,
font_family="Rockwell"),
hovermode='x Unified',
legend=dict(
yanchor="top",
y=0.99,
xanchor="left",
x=0.01))
fig = dict(data=[trace_bar, trace_line], layout=layout)
return fig
elif plot_type == 'cumulative':
trace = go.Scatter(x=x, y=y, mode='lines', name=label, fill='tozeroy',
fillcolor=colors[label],
line={'width': 2, 'color': colors[label]},
hovertemplate='%{x|%b %d, %Y} <br> %{y:-.0f}'
)
title = f'Cumulative Covid {label.lower()} in {state}'
layout = dict(title=title,
yaxis_title=f'Confirmed # of {label}',
xaxis_title='Date/Time',
font=dict(family="Courier New, monospace",
size=16))
data = [trace]
fig = dict(data=data, layout=layout)
return fig
# check_point
saver = tf.compat.v1.train.Saver(tf.compat.v1.global_variables(),
max_to_keep=100)
#################################################
# load check point (load weights)
if 0:
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
saver.restore(sess, r'')
#################################################
sess.run(tf.initialize_all_variables())
##################################################
# start training
##################################################
loss_ema = moving_average(init_avg=0.009)
loss_hist = []
val_loss_hist = []
for e in range(settings['epoch']):
cur_lr = settings['init_learning_rate'] * settings['learning_rate_decay']**e
loss_sum = 0
for step in range(settings['step_per_epoch']):
c_stft, n_stft, c_wav, _ = data_generator.get_batch_data(
settings['batch_size'],
clean_percentage=0.,
is_noise=True,
is_reverb=settings['use_reverb'],
mode='train',
sess=sess)
loss, _ = sess.run(
total_reward = 0
deltas = []
for episode in range(NUM_EPISODES + 1):
eps = START_EPSILON / (1.0 + episode * EPSILON_TAPER)
if episode%10000 == 0:
cp_file = checkpoint(Q, CHECKPOINT_DIR, GAME, episode)
print('Saved Checkpoint to: ', cp_file)
biggest_change = 0
curr_state = env.reset()
for step in range(MAX_STEPS):
prev_state = curr_state
state_visit_counts[prev_state] = state_visit_counts.get(prev_state,0)+1
action = epsilon_action(curr_state, eps)
curr_state, reward, done, info = env.step(action)
total_reward += reward
old_qsa = Q[prev_state][action]
update_Q(prev_state, action, reward, curr_state)
biggest_change = max(biggest_change, np.abs(old_qsa - Q[prev_state][action]))
if done:
break
deltas.append(biggest_change)
mean_state_visits = np.mean(list(state_visit_counts.values()))
print('Each state was visited on average: ', mean_state_visits, ' times')
plt.plot(moving_average(deltas, n=1000))
plt.show()
str(resolution[1]) + '.bmp', img * max_val)
test = get_peano_index(img.shape[0]) # Parcours de peano
# test = [a.flatten() for a in np.indices((256, 256))] #Parcours ligne par ligne
hidden = img[test[0], test[1]]
for idx, noise in enumerate(gauss_noise):
if not noise['corr']:
img_noisy = (img == 0) * np.random.normal(
noise['mu1'], noise['sig1'],
img.shape) + (img == 1) * np.random.normal(
noise['mu2'], noise['sig2'], img.shape)
corr = ''
corr_param = ''
else:
img_noisy = moving_average(
(img == 0) *
np.random.normal(noise['mu1'], noise['sig1'], img.shape),
noise['corr_param'][0],
noise['corr_param'][1]) + moving_average(
(img == 1) * np.random.normal(
noise['mu2'], noise['sig2'], img.shape),
noise['corr_param'][0], noise['corr_param'][1])
corr = 'corr'
corr_param = str(noise['corr_param'][0]) + '_' + str(
noise['corr_param'][1])
noise_param = '(' + str(noise['mu1']) + ',' + str(
noise['sig1']) + ')' + '_' + '(' + str(
noise['mu2']) + ',' + str(noise['sig2']) + ')'
cv.imwrite(
resfolder + '/' + imgf + '/' + str(resolution[0]) + '_' +
str(resolution[1]) + '_' + corr + '_' + corr_param +
noise_param + '.bmp',
from agents import PermanentAgentStorage
# Environment setup:
n_agents = 2
randomness = 0.05
n_tests = 5000
env = MultiAgentNavigation(size=5, n_agents=n_agents, draw=False)
scores_Q_online_decentralized = np.zeros(n_tests)
agent_storage = PermanentAgentStorage(env,
DynaAgent,
k=0,
randomness=randomness,
lr=0.02)
for i in range(n_tests):
# Generate Data
steps, observations, actions = run_episode(env,
agent_storage.get_next_agent,
return_details=True)
scores_Q_online_decentralized[i] = steps
if i % 10 == 0:
print(i)
print("Mean Q_initialized - constant: " +
str(np.mean(scores_Q_online_decentralized)) + " --- S.E.: " +
str(np.std(scores_Q_online_decentralized) / np.sqrt(n_tests)))
plt.plot(moving_average(scores_Q_online_decentralized))
plt.show()
# train the DQN
if T % train_freq == 0:
train_on_batch(memory, min(batch_size, T), df)
observation = next_observation
if T % target_update_freq == 0:
Q_network_target.load_state_dict(Q_network.state_dict())
if done:
if (i_episode + 1) % 100 == 0:
print(
"Episode {} finished after {} timesteps, T: {}".format(
i_episode, t + 1, T))
break
rewards.append(total_r)
env.close()
return rewards
N_EPS = 500
# rewards_DQN_dueling = learn_episodic_DQN(N_EPS, 500, use_dueling=True)
rewards_DQN = learn_episodic_DQN(N_EPS, 500)
plt.plot(moving_average(rewards_DQN, 100), label="DQN")
plt.legend()
plt.show()
def main():
distance_functions = [euclidian_distance]
clustering_classes = [PerfectClustering, OnlineClustering]
network_config = 'sp=True_tm=True_tp=False_SDRClassifier'
exp_names = ['binary_ampl=10.0_mean=0.0_noise=0.0',
'binary_ampl=10.0_mean=0.0_noise=1.0',
'sensortag_z']
# Exp params
moving_average_window = 1 # for all moving averages of the experiment
ClusteringClass = clustering_classes[0]
distance_func = distance_functions[0]
exp_name = exp_names[0]
start_idx = 0
end_idx = 100
input_width = 2048 * 32
active_cells_weight = 0
predicted_active_cells_weight = 1
max_num_clusters = 3
num_cluster_snapshots = 2
show_plots = False
distance_matrix_ignore_noise = True # whether to ignore label 0 (noise)
# Clean an create output directory for the graphs
plots_output_dir = 'plots/%s' % exp_name
if os.path.exists(plots_output_dir):
shutil.rmtree(plots_output_dir)
os.makedirs(plots_output_dir)
# load traces
file_name = get_file_name(exp_name, network_config)
traces = loadTraces(file_name)
sensor_values = traces['sensorValue'][start_idx:end_idx]
categories = traces['actualCategory'][start_idx:end_idx]
raw_anomaly_scores = traces['rawAnomalyScore'][start_idx:end_idx]
anomaly_scores = []
anomaly_score_ma = 0.0
for raw_anomaly_score in raw_anomaly_scores:
anomaly_score_ma = moving_average(anomaly_score_ma,
raw_anomaly_score,
moving_average_window)
anomaly_scores.append(anomaly_score_ma)
active_cells = traces['tmActiveCells'][start_idx:end_idx]
predicted_active_cells = traces['tmPredictedActiveCells'][start_idx:end_idx]
# generate sdrs to cluster
active_cells_sdrs = convert_to_sdrs(active_cells, input_width)
predicted_activeCells_sdrs = np.array(convert_to_sdrs(predicted_active_cells,
input_width))
sdrs = (active_cells_weight * np.array(active_cells_sdrs) +
predicted_active_cells_weight * predicted_activeCells_sdrs)
# start and end for the x axis of the graphs
start = start_idx
if end_idx < 0:
end = len(sdrs) - end_idx - 1
else:
end = end_idx
xlim = [start, end]
# list of timesteps specifying when a snapshot of the clusters will be taken
step = (end - start) / num_cluster_snapshots - 1
cluster_snapshot_indices = range(start + step, end, step)
# run clustering
(clustering_accuracies,
cluster_snapshots,
closest_cluster_history) = run(sdrs,
categories,
distance_func,
moving_average_window,
max_num_clusters,
ClusteringClass,
cluster_snapshot_indices)
# plot cluster assignments over time
for i in range(num_cluster_snapshots):
clusters = cluster_snapshots[i]
plot_cluster_assignments(plots_output_dir, clusters, cluster_snapshot_indices[i])
# plot inter-cluster distance matrix
cluster_ids = [c.id for c in closest_cluster_history if c is not None]
plot_id = 'inter-cluster_t=%s' % cluster_snapshot_indices[i]
plot_inter_sequence_distances(plots_output_dir,
plot_id,
distance_func,
sdrs[:cluster_snapshot_indices[i]],
cluster_ids[:cluster_snapshot_indices[i]],
distance_matrix_ignore_noise)
# plot inter-category distance matrix
plot_id = 'inter-category_t=%s ' % cluster_snapshot_indices[i]
plot_inter_sequence_distances(plots_output_dir,
plot_id,
distance_func,
sdrs[:cluster_snapshot_indices[i]],
categories[:cluster_snapshot_indices[i]],
distance_matrix_ignore_noise)
# plot clustering accuracy over time
plot_id = 'file=%s | moving_average_window=%s' % (exp_name,
moving_average_window)
plot_accuracy(plots_output_dir,
plot_id,
sensor_values,
categories,
anomaly_scores,
clustering_accuracies,
xlim)
if show_plots:
plt.show()
def main():
distance_functions = [euclidian_distance]
clustering_classes = [PerfectClustering, OnlineClusteringV2]
# Exp params
moving_average_window = 2 # for all moving averages of the experiment
ClusteringClass = clustering_classes[1]
distance_func = distance_functions[0]
merge_threshold = 30 # Cutoff distance to merge clusters. 'None' to ignore.
start_idx = 0
end_idx = -1
input_width = 2048 * 32
active_cells_weight = 0
predicted_active_cells_weight = 10
max_num_clusters = 3
num_cluster_snapshots = 1
show_plots = True
distance_matrix_ignore_noise = False # ignore label 0 if used to label noise.
exp_name = 'body_acc_x_inertial_signals_train'
# Clean an create output directory for the graphs
plots_output_dir = 'plots/%s' % exp_name
if os.path.exists(plots_output_dir):
shutil.rmtree(plots_output_dir)
os.makedirs(plots_output_dir)
# load traces
file_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
os.pardir, 'htm', 'traces',
'trace_%s.csv' % exp_name)
traces = loadTraces(file_path)
num_records = len(traces['scalarValue'])
# start and end for the x axis of the graphs
if start_idx < 0:
start = num_records + start_idx
else:
start = start_idx
if end_idx < 0:
end = num_records + end_idx
else:
end = end_idx
xlim = [0, end - start]
# input data
sensor_values = traces['scalarValue'][start:end]
categories = traces['label'][start:end]
active_cells = traces['tmActiveCells'][start:end]
predicted_active_cells = traces['tmPredictedActiveCells'][start:end]
raw_anomaly_scores = traces['rawAnomalyScore'][start:end]
anomaly_scores = []
anomaly_score_ma = 0.0
for raw_anomaly_score in raw_anomaly_scores:
anomaly_score_ma = moving_average(anomaly_score_ma, raw_anomaly_score,
moving_average_window)
anomaly_scores.append(anomaly_score_ma)
# generate sdrs to cluster
active_cells_sdrs = convert_to_sdrs(active_cells, input_width)
predicted_active_cells_sdrs = np.array(
convert_to_sdrs(predicted_active_cells, input_width))
sdrs = (float(active_cells_weight) * np.array(active_cells_sdrs) +
float(predicted_active_cells_weight) * predicted_active_cells_sdrs)
# list of timesteps specifying when a snapshot of the clusters will be taken
step = (end - start) / num_cluster_snapshots - 1
cluster_snapshot_indices = range(step, end - start, step)
# run clustering
(clustering_accuracies, cluster_snapshots,
closest_cluster_history) = run(sdrs, categories, anomaly_scores,
distance_func, moving_average_window,
max_num_clusters, ClusteringClass,
merge_threshold, cluster_snapshot_indices)
# cluster_categories = []
# for c in closest_cluster_history:
# if c is not None:
# cluster_categories.append(c.label_distribution()[0]['label'])
# plot cluster assignments over time
for i in range(num_cluster_snapshots):
clusters = cluster_snapshots[i]
snapshot_index = cluster_snapshot_indices[i]
plot_cluster_assignments(plots_output_dir, clusters, snapshot_index)
# plot inter-cluster distance matrix
# plot_id = 'inter-cluster_t=%s' % snapshot_index
# plot_inter_sequence_distances(plots_output_dir,
# plot_id,
# distance_func,
# sdrs[:snapshot_index],
# cluster_categories[:snapshot_index],
# distance_matrix_ignore_noise)
# plot inter-category distance matrix
plot_id = 'inter-category_t=%s ' % snapshot_index
plot_inter_sequence_distances(plots_output_dir, plot_id, distance_func,
sdrs[:snapshot_index],
categories[:snapshot_index],
distance_matrix_ignore_noise)
# plot clustering accuracy over time
plot_id = 'file=%s | moving_average_window=%s' % (exp_name,
moving_average_window)
plot_accuracy(plots_output_dir, plot_id, sensor_values, categories,
anomaly_scores, clustering_accuracies, xlim)
if show_plots:
plt.show()
for epoch in range(start_epoch, args.epochs):
time_ep = time.time()
lr = schedule(epoch)
utils.adjust_learning_rate(optimizer, lr)
train_res = utils.train_epoch(loaders['train'], model, criterion,
optimizer)
if epoch == 0 or epoch % args.eval_freq == args.eval_freq - 1 or epoch == args.epochs - 1:
test_res = utils.eval(loaders['test'], model, criterion)
else:
test_res = {'loss': None, 'accuracy': None}
if args.swa and (epoch + 1) >= args.swa_start and (
epoch + 1 - args.swa_start) % args.swa_c_epochs == 0:
utils.moving_average(swa_model, model, 1.0 / (swa_n + 1))
swa_n += 1
if epoch == 0 or epoch % args.eval_freq == args.eval_freq - 1 or epoch == args.epochs - 1:
utils.bn_update(loaders['train'], swa_model)
swa_res = utils.eval(loaders['test'], swa_model, criterion)
else:
swa_res = {'loss': None, 'accuracy': None}
if (epoch + 1) % args.save_freq == 0:
utils.save_checkpoint(
args.dir,
epoch + 1,
state_dict=model.state_dict(),
swa_state_dict=swa_model.state_dict() if args.swa else None,
swa_n=swa_n if args.swa else None,
optimizer=optimizer.state_dict())
def distance(self):
""" Returns the moving average of the _sensor_data """
return moving_average(list(self._sensor_data), 3)
def main():
script_dir = os.path.dirname(__file__)
module_path = os.path.abspath(os.path.join(script_dir, '..', '..'))
global msglogger
# Parse arguments
args = parser.get_parser().parse_args()
if args.epochs is None:
args.epochs = 90
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
msglogger = apputils.config_pylogger(
os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir)
# Log various details about the execution environment. It is sometimes useful
# to refer to past experiment executions and this information may be useful.
apputils.log_execution_env_state(args.compress,
msglogger.logdir,
gitroot=module_path)
msglogger.debug("Distiller: %s", distiller.__version__)
start_epoch = 0
ending_epoch = args.epochs
perf_scores_history = []
if args.evaluate:
args.deterministic = True
if args.deterministic:
# Experiment reproducibility is sometimes important. Pete Warden expounded about this
# in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/
distiller.set_deterministic(
) # Use a well-known seed, for repeatability of experiments
else:
# Turn on CUDNN benchmark mode for best performance. This is usually "safe" for image
# classification models, as the input sizes don't change during the run
# See here: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/3
cudnn.benchmark = True
if args.cpu or not torch.cuda.is_available():
# Set GPU index to -1 if using CPU
args.device = 'cpu'
args.gpus = -1
else:
args.device = 'cuda'
if args.gpus is not None:
try:
args.gpus = [int(s) for s in args.gpus.split(',')]
except ValueError:
raise ValueError(
'ERROR: Argument --gpus must be a comma-separated list of integers only'
)
available_gpus = torch.cuda.device_count()
for dev_id in args.gpus:
if dev_id >= available_gpus:
raise ValueError(
'ERROR: GPU device ID {0} requested, but only {1} devices available'
.format(dev_id, available_gpus))
# Set default device in case the first one on the list != 0
torch.cuda.set_device(args.gpus[0])
# Infer the dataset from the model name
args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet'
args.num_classes = 10 if args.dataset == 'cifar10' else 1000
# Create the model
model = create_model(args.pretrained,
args.dataset,
args.arch,
parallel=not args.load_serialized,
device_ids=args.gpus)
if args.swa:
swa_model = create_model(args.pretrained,
args.dataset,
args.arch,
parallel=not args.load_serialized,
device_ids=args.gpus)
swa_n = 0
compression_scheduler = None
# Create a couple of logging backends. TensorBoardLogger writes log files in a format
# that can be read by Google's Tensor Board. PythonLogger writes to the Python logger.
tflogger = TensorBoardLogger(msglogger.logdir)
pylogger = PythonLogger(msglogger)
# TODO(barrh): args.deprecated_resume is deprecated since v0.3.1
if args.deprecated_resume:
msglogger.warning(
'The "--resume" flag is deprecated. Please use "--resume-from=YOUR_PATH" instead.'
)
if not args.reset_optimizer:
msglogger.warning(
'If you wish to also reset the optimizer, call with: --reset-optimizer'
)
args.reset_optimizer = True
args.resumed_checkpoint_path = args.deprecated_resume
# We can optionally resume from a checkpoint
optimizer = None
# TODO: resume from swa mode
if args.resumed_checkpoint_path:
if args.swa:
model, swa_model, swa_n, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint(
model,
args.resumed_checkpoint_path,
swa_model=swa_model,
swa_n=swa_n,
model_device=args.device)
else:
model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint(
model, args.resumed_checkpoint_path, model_device=args.device)
elif args.load_model_path:
model = apputils.load_lean_checkpoint(model,
args.load_model_path,
model_device=args.device)
if args.reset_optimizer:
start_epoch = 0
if optimizer is not None:
optimizer = None
msglogger.info(
'\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0'
)
# Define loss function (criterion)
criterion = nn.CrossEntropyLoss().to(args.device)
if optimizer is None:
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
msglogger.info('Optimizer Type: %s', type(optimizer))
msglogger.info('Optimizer Args: %s', optimizer.defaults)
# This sample application can be invoked to produce various summary reports.
if args.summary:
return summarize_model(model, args.dataset, which_summary=args.summary)
activations_collectors = create_activation_stats_collectors(
model, *args.activation_stats)
# Load the datasets: the dataset to load is inferred from the model name passed
# in args.arch. The default dataset is ImageNet, but if args.arch contains the
# substring "_cifar", then cifar10 is used.
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size,
args.effective_test_size)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler),
len(test_loader.sampler))
if args.sensitivity is not None:
sensitivities = np.arange(args.sensitivity_range[0],
args.sensitivity_range[1],
args.sensitivity_range[2])
return sensitivity_analysis(model, criterion, test_loader, pylogger,
args, sensitivities)
if args.evaluate:
return evaluate_model(model, criterion, test_loader, pylogger,
activations_collectors, args,
compression_scheduler)
if args.compress:
# The main use-case for this sample application is CNN compression. Compression
# requires a compression schedule configuration file in YAML.
compression_scheduler = distiller.file_config(
model, optimizer, args.compress, compression_scheduler,
(start_epoch - 1) if args.resumed_checkpoint_path else None)
# Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer)
model.to(args.device)
elif compression_scheduler is None:
compression_scheduler = distiller.CompressionScheduler(model)
if args.thinnify:
#zeros_mask_dict = distiller.create_model_masks_dict(model)
assert args.resumed_checkpoint_path is not None, \
"You must use --resume-from to provide a checkpoint file to thinnify"
distiller.remove_filters(model,
compression_scheduler.zeros_mask_dict,
args.arch,
args.dataset,
optimizer=None)
apputils.save_checkpoint(0,
args.arch,
model,
optimizer=None,
scheduler=compression_scheduler,
name="{}_thinned".format(
args.resumed_checkpoint_path.replace(
".pth.tar", "")),
dir=msglogger.logdir)
print(
"Note: your model may have collapsed to random inference, so you may want to fine-tune"
)
return
if args.lr_find:
lr_finder = distiller.LRFinder(model,
optimizer,
criterion,
device=args.device)
lr_finder.range_test(train_loader, end_lr=10, num_iter=100)
lr_finder.plot()
return
if start_epoch >= ending_epoch:
msglogger.error(
'epoch count is too low, starting epoch is {} but total epochs set to {}'
.format(start_epoch, ending_epoch))
raise ValueError('Epochs parameter is too low. Nothing to do.')
for epoch in range(start_epoch, ending_epoch):
# This is the main training loop.
msglogger.info('\n')
if compression_scheduler:
compression_scheduler.on_epoch_begin(
epoch, metrics=(vloss if (epoch != start_epoch) else 10**6))
# Train for one epoch
with collectors_context(activations_collectors["train"]) as collectors:
train(train_loader,
model,
criterion,
optimizer,
epoch,
compression_scheduler,
loggers=[tflogger, pylogger],
args=args)
# distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger])
# distiller.log_activation_statsitics(epoch, "train", loggers=[tflogger],
# collector=collectors["sparsity"])
if args.masks_sparsity:
msglogger.info(
distiller.masks_sparsity_tbl_summary(
model, compression_scheduler))
# evaluate on validation set
with collectors_context(activations_collectors["valid"]) as collectors:
top1, top5, vloss = validate(val_loader, model, criterion,
[pylogger], args, epoch)
msglogger.info('==> Top1: %.3f Top5: %.3f Loss: %.3f\n',
top1, top5, vloss)
distiller.log_activation_statsitics(
epoch,
"valid",
loggers=[tflogger],
collector=collectors["sparsity"])
save_collectors_data(collectors, msglogger.logdir)
stats = ('Performance/Validation/',
OrderedDict([('Loss', vloss), ('Top1', top1),
('Top5', top5)]))
if args.swa and (epoch + 1) >= args.swa_start and (
epoch + 1 - args.swa_start
) % args.swa_freq == 0 or epoch == ending_epoch - 1:
utils.moving_average(swa_model, model, 1. / (swa_n + 1))
swa_n += 1
utils.bn_update(train_loader, swa_model, args)
swa_top1, swa_top5, swa_loss = validate(val_loader, swa_model,
criterion, [pylogger],
args, epoch)
msglogger.info(
'==> SWA_Top1: %.3f SWA_Top5: %.3f SWA_Loss: %.3f\n',
swa_top1, swa_top5, swa_loss)
swa_res = OrderedDict([('SWA_Loss', swa_loss),
('SWA_Top1', swa_top1),
('SWA_Top5', swa_top5)])
stats[1].update(swa_res)
distiller.log_training_progress(stats,
None,
epoch,
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[tflogger])
if compression_scheduler:
compression_scheduler.on_epoch_end(epoch, optimizer)
# Update the list of top scores achieved so far, and save the checkpoint
update_training_scores_history(perf_scores_history, model, top1, top5,
epoch, args.num_best_scores)
is_best = epoch == perf_scores_history[0].epoch
checkpoint_extras = {
'current_top1': top1,
'best_top1': perf_scores_history[0].top1,
'best_epoch': perf_scores_history[0].epoch
}
if args.swa:
apputils.save_checkpoint(epoch,
args.arch,
model,
swa_model,
swa_n,
optimizer=optimizer,
scheduler=compression_scheduler,
extras=checkpoint_extras,
is_best=is_best,
name=args.name,
dir=msglogger.logdir)
else:
apputils.save_checkpoint(epoch,
args.arch,
model,
optimizer=optimizer,
scheduler=compression_scheduler,
extras=checkpoint_extras,
is_best=is_best,
name=args.name,
dir=msglogger.logdir)
# Finally run results on the test set
test(test_loader,
model,
criterion, [pylogger],
activations_collectors,
args=args)
if args.swa:
test(test_loader,
swa_model,
criterion, [pylogger],
activations_collectors,
args=args)
def train_main(cfg):
'''
训练的主函数
:param cfg: 配置
:return:
'''
# config
train_cfg = cfg.train_cfg
dataset_cfg = cfg.dataset_cfg
model_cfg = cfg.model_cfg
is_parallel = cfg.setdefault(key='is_parallel', default=False)
device = cfg.device
is_online_train = cfg.setdefault(key='is_online_train', default=False)
# 配置logger
logging.basicConfig(filename=cfg.logfile,
filemode='a',
level=logging.INFO,
format='%(asctime)s\n%(message)s',
datefmt='%Y-%m-%d %H:%M:%S')
logger = logging.getLogger()
#
# 构建数据集
train_dataset = LandDataset(DIR_list=dataset_cfg.train_dir_list,
mode='train',
input_channel=dataset_cfg.input_channel,
transform=dataset_cfg.train_transform)
split_val_from_train_ratio = dataset_cfg.setdefault(
key='split_val_from_train_ratio', default=None)
if split_val_from_train_ratio is None:
val_dataset = LandDataset(DIR_list=dataset_cfg.val_dir_list,
mode='val',
input_channel=dataset_cfg.input_channel,
transform=dataset_cfg.val_transform)
else:
val_size = int(len(train_dataset) * split_val_from_train_ratio)
train_size = len(train_dataset) - val_size
train_dataset, val_dataset = random_split(
train_dataset, [train_size, val_size],
generator=torch.manual_seed(cfg.random_seed))
# val_dataset.dataset.transform = dataset_cfg.val_transform # 要配置一下val的transform
print(f"按照{split_val_from_train_ratio}切分训练集...")
# 构建dataloader
def _init_fn():
np.random.seed(cfg.random_seed)
train_dataloader = DataLoader(train_dataset,
batch_size=train_cfg.batch_size,
shuffle=True,
num_workers=train_cfg.num_workers,
drop_last=True,
worker_init_fn=_init_fn())
val_dataloader = DataLoader(val_dataset,
batch_size=train_cfg.batch_size,
num_workers=train_cfg.num_workers,
shuffle=False,
drop_last=True,
worker_init_fn=_init_fn())
# 构建模型
if train_cfg.is_swa:
model = torch.load(train_cfg.check_point_file, map_location=device).to(
device) # device参数传在里面，不然默认是先加载到cuda:0，to之后再加载到相应的device上
swa_model = torch.load(
train_cfg.check_point_file, map_location=device).to(
device) # device参数传在里面，不然默认是先加载到cuda:0，to之后再加载到相应的device上
if is_parallel:
model = torch.nn.DataParallel(model)
swa_model = torch.nn.DataParallel(swa_model)
swa_n = 0
parameters = swa_model.parameters()
else:
model = build_model(model_cfg).to(device)
if is_parallel:
model = torch.nn.DataParallel(model)
parameters = model.parameters()
# 定义优化器
optimizer_cfg = train_cfg.optimizer_cfg
lr_scheduler_cfg = train_cfg.lr_scheduler_cfg
if optimizer_cfg.type == 'adam':
optimizer = optim.Adam(params=parameters,
lr=optimizer_cfg.lr,
weight_decay=optimizer_cfg.weight_decay)
elif optimizer_cfg.type == 'adamw':
optimizer = optim.AdamW(params=parameters,
lr=optimizer_cfg.lr,
weight_decay=optimizer_cfg.weight_decay)
elif optimizer_cfg.type == 'sgd':
optimizer = optim.SGD(params=parameters,
lr=optimizer_cfg.lr,
momentum=optimizer_cfg.momentum,
weight_decay=optimizer_cfg.weight_decay)
elif optimizer_cfg.type == 'RMS':
optimizer = optim.RMSprop(params=parameters,
lr=optimizer_cfg.lr,
weight_decay=optimizer_cfg.weight_decay)
else:
raise Exception('没有该优化器！')
if not lr_scheduler_cfg:
lr_scheduler = None
elif lr_scheduler_cfg.policy == 'cos':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
optimizer,
lr_scheduler_cfg.T_0,
lr_scheduler_cfg.T_mult,
lr_scheduler_cfg.eta_min,
last_epoch=lr_scheduler_cfg.last_epoch)
elif lr_scheduler_cfg.policy == 'LambdaLR':
import math
lf = lambda x: (((1 + math.cos(x * math.pi / train_cfg.num_epochs)) / 2
)**1.0) * 0.95 + 0.05 # cosine
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer,
lr_lambda=lf)
lr_scheduler.last_epoch = 0
else:
lr_scheduler = None
# 定义损失函数
DiceLoss_fn = DiceLoss(mode='multiclass')
SoftCrossEntropy_fn = SoftCrossEntropyLoss(smooth_factor=0.1)
loss_func = L.JointLoss(first=DiceLoss_fn,
second=SoftCrossEntropy_fn,
first_weight=0.5,
second_weight=0.5).cuda()
# loss_cls_func = torch.nn.BCEWithLogitsLoss()
# 创建保存模型的文件夹
check_point_dir = '/'.join(model_cfg.check_point_file.split('/')[:-1])
if not os.path.exists(check_point_dir): # 如果文件夹不存在就创建
os.mkdir(check_point_dir)
# 开始训练
auto_save_epoch_list = train_cfg.setdefault(key='auto_save_epoch_list',
default=5) # 每隔几轮保存一次模型，默认为5
train_loss_list = []
val_loss_list = []
val_loss_min = 999999
best_epoch = 0
best_miou = 0
train_loss = 10 # 设置一个初始值
logger.info('开始在{}上训练{}模型...'.format(device, model_cfg.type))
logger.info('补充信息：{}\n'.format(cfg.setdefault(key='info', default='None')))
for epoch in range(train_cfg.num_epochs):
print()
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
start_time = time.time()
print(f"正在进行第{epoch}轮训练...")
logger.info('*' * 10 + f"第{epoch}轮" + '*' * 10)
#
# 训练一轮
if train_cfg.is_swa: # swa训练方式
train_loss = train_epoch(swa_model, optimizer, lr_scheduler,
loss_func, train_dataloader, epoch,
device)
moving_average(model, swa_model, 1.0 / (swa_n + 1))
swa_n += 1
bn_update(train_dataloader, model, device)
else:
train_loss = train_epoch(model, optimizer, lr_scheduler, loss_func,
train_dataloader, epoch, device)
# train_loss = train_unet3p_epoch(model, optimizer, lr_scheduler, loss_func, train_dataloader, epoch, device)
#
# 在训练集上评估模型
# val_loss, val_miou = evaluate_unet3p_model(model, val_dataset, loss_func, device,
# cfg.num_classes, train_cfg.num_workers, batch_size=train_cfg.batch_size)
if not is_online_train: # 只有在线下训练的时候才需要评估模型
val_loss, val_miou = evaluate_model(model, val_dataloader,
loss_func, device,
cfg.num_classes)
else:
val_loss = 0
val_miou = 0
train_loss_list.append(train_loss)
val_loss_list.append(val_loss)
# 保存模型
if not is_online_train: # 非线上训练时需要保存best model
if val_loss < val_loss_min:
val_loss_min = val_loss
best_epoch = epoch
best_miou = val_miou
if is_parallel:
torch.save(model.module, model_cfg.check_point_file)
else:
torch.save(model, model_cfg.check_point_file)
if epoch in auto_save_epoch_list: # 如果再需要保存的轮次中，则保存
model_file = model_cfg.check_point_file.split(
'.pth')[0] + '-epoch{}.pth'.format(epoch)
if is_parallel:
torch.save(model.module, model_file)
else:
torch.save(model, model_file)
# 打印中间结果
end_time = time.time()
run_time = int(end_time - start_time)
m, s = divmod(run_time, 60)
time_str = "{:02d}分{:02d}秒".format(m, s)
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
out_str = "第{}轮训练完成，耗时{}，\t训练集上的loss={:.6f}；\t验证集上的loss={:.4f}，mIoU={:.6f}\t最好的结果是第{}轮，mIoU={:.6f}" \
.format(epoch, time_str, train_loss, val_loss, val_miou, best_epoch, best_miou)
# out_str = "第{}轮训练完成，耗时{}，\n训练集上的segm_loss={:.6f},cls_loss{:.6f}\n验证集上的segm_loss={:.4f},cls_loss={:.4f},mIoU={:.6f}\n最好的结果是第{}轮，mIoU={:.6f}" \
# .format(epoch, time_str, train_loss, train_cls_loss, val_loss, val_cls_loss, val_miou, best_epoch,
# best_miou)
print(out_str)
logger.info(out_str + '\n')
def main():
distance_functions = [euclidian_distance]
clustering_classes = [PerfectClustering, OnlineClustering]
network_config = 'sp=True_tm=True_tp=False_SDRClassifier'
exp_names = [
'binary_ampl=10.0_mean=0.0_noise=0.0',
'binary_ampl=10.0_mean=0.0_noise=1.0', 'sensortag_z'
]
# Exp params
moving_average_window = 1 # for all moving averages of the experiment
ClusteringClass = clustering_classes[0]
distance_func = distance_functions[0]
exp_name = exp_names[0]
start_idx = 0
end_idx = 100
input_width = 2048 * 32
active_cells_weight = 0
predicted_active_cells_weight = 1
max_num_clusters = 3
num_cluster_snapshots = 2
show_plots = False
distance_matrix_ignore_noise = True # whether to ignore label 0 (noise)
# Clean an create output directory for the graphs
plots_output_dir = 'plots/%s' % exp_name
if os.path.exists(plots_output_dir):
shutil.rmtree(plots_output_dir)
os.makedirs(plots_output_dir)
# load traces
file_name = get_file_name(exp_name, network_config)
traces = loadTraces(file_name)
sensor_values = traces['sensorValue'][start_idx:end_idx]
categories = traces['actualCategory'][start_idx:end_idx]
raw_anomaly_scores = traces['rawAnomalyScore'][start_idx:end_idx]
anomaly_scores = []
anomaly_score_ma = 0.0
for raw_anomaly_score in raw_anomaly_scores:
anomaly_score_ma = moving_average(anomaly_score_ma, raw_anomaly_score,
moving_average_window)
anomaly_scores.append(anomaly_score_ma)
active_cells = traces['tmActiveCells'][start_idx:end_idx]
predicted_active_cells = traces['tmPredictedActiveCells'][
start_idx:end_idx]
# generate sdrs to cluster
active_cells_sdrs = convert_to_sdrs(active_cells, input_width)
predicted_activeCells_sdrs = np.array(
convert_to_sdrs(predicted_active_cells, input_width))
sdrs = (active_cells_weight * np.array(active_cells_sdrs) +
predicted_active_cells_weight * predicted_activeCells_sdrs)
# start and end for the x axis of the graphs
start = start_idx
if end_idx < 0:
end = len(sdrs) - end_idx - 1
else:
end = end_idx
xlim = [start, end]
# list of timesteps specifying when a snapshot of the clusters will be taken
step = (end - start) / num_cluster_snapshots - 1
cluster_snapshot_indices = range(start + step, end, step)
# run clustering
(clustering_accuracies, cluster_snapshots,
closest_cluster_history) = run(sdrs, categories, distance_func,
moving_average_window, max_num_clusters,
ClusteringClass, cluster_snapshot_indices)
# plot cluster assignments over time
for i in range(num_cluster_snapshots):
clusters = cluster_snapshots[i]
plot_cluster_assignments(plots_output_dir, clusters,
cluster_snapshot_indices[i])
# plot inter-cluster distance matrix
cluster_ids = [c.id for c in closest_cluster_history if c is not None]
plot_id = 'inter-cluster_t=%s' % cluster_snapshot_indices[i]
plot_inter_sequence_distances(
plots_output_dir, plot_id, distance_func,
sdrs[:cluster_snapshot_indices[i]],
cluster_ids[:cluster_snapshot_indices[i]],
distance_matrix_ignore_noise)
# plot inter-category distance matrix
plot_id = 'inter-category_t=%s ' % cluster_snapshot_indices[i]
plot_inter_sequence_distances(plots_output_dir, plot_id, distance_func,
sdrs[:cluster_snapshot_indices[i]],
categories[:cluster_snapshot_indices[i]],
distance_matrix_ignore_noise)
# plot clustering accuracy over time
plot_id = 'file=%s | moving_average_window=%s' % (exp_name,
moving_average_window)
plot_accuracy(plots_output_dir, plot_id, sensor_values, categories,
anomaly_scores, clustering_accuracies, xlim)
if show_plots:
plt.show()
def learn_episodic_A2C(N_eps=500, max_ep_steps=500):
df = 0.99
rewards = []
env = gym.make('CartPole-v0')
env._max_episode_steps = max_ep_steps
for i_episode in range(N_eps):
observation = env.reset()
total_r = 0
for t in range(100000):
action = select_action(observation)
observation, reward, done, info = env.step(action)
policy.rewards.append(reward)
total_r += reward
if done:
train_on_rollout(df)
if (i_episode + 1) % 100 == 0:
print("Episode {} finished after {} timesteps".format(
i_episode, t + 1))
break
rewards.append(total_r)
env.close()
return rewards
N_EPS = 500
rewards_A2C = learn_episodic_A2C(N_EPS, 500)
plt.plot(moving_average(rewards_A2C, 100), label="A2C")
plt.legend()
plt.show()
def main():
distance_functions = [euclidian_distance]
clustering_classes = [PerfectClustering, OnlineClusteringV2]
network_config = "sp=True_tm=True_tp=False_SDRClassifier"
exp_names = [
"body_acc_x",
"binary_ampl=10.0_mean=0.0_noise=0.0",
"binary_ampl=10.0_mean=0.0_noise=1.0",
"sensortag_z",
]
# Exp params
moving_average_window = 2 # for all moving averages of the experiment
ClusteringClass = clustering_classes[1]
distance_func = distance_functions[0]
exp_name = exp_names[0]
start_idx = 1000
end_idx = 12000
input_width = 2048 * 32
active_cells_weight = 0
predicted_active_cells_weight = 10
max_num_clusters = 3
num_cluster_snapshots = 1
show_plots = True
distance_matrix_ignore_noise = True # whether to ignore label 0 (noise)
# Clean an create output directory for the graphs
plots_output_dir = "plots/%s" % exp_name
if os.path.exists(plots_output_dir):
shutil.rmtree(plots_output_dir)
os.makedirs(plots_output_dir)
# load traces
file_name = get_file_name(exp_name, network_config)
traces = loadTraces(file_name)
num_records = len(traces["sensorValue"])
# start and end for the x axis of the graphs
if start_idx < 0:
start = num_records + start_idx
else:
start = start_idx
if end_idx < 0:
end = num_records + end_idx
else:
end = end_idx
xlim = [0, end - start]
# input data
sensor_values = traces["sensorValue"][start:end]
categories = traces["actualCategory"][start:end]
active_cells = traces["tmActiveCells"][start:end]
predicted_active_cells = traces["tmPredictedActiveCells"][start:end]
raw_anomaly_scores = traces["rawAnomalyScore"][start:end]
anomaly_scores = []
anomaly_score_ma = 0.0
for raw_anomaly_score in raw_anomaly_scores:
anomaly_score_ma = moving_average(anomaly_score_ma, raw_anomaly_score, moving_average_window)
anomaly_scores.append(anomaly_score_ma)
# generate sdrs to cluster
active_cells_sdrs = convert_to_sdrs(active_cells, input_width)
predicted_active_cells_sdrs = np.array(convert_to_sdrs(predicted_active_cells, input_width))
sdrs = (
float(active_cells_weight) * np.array(active_cells_sdrs)
+ float(predicted_active_cells_weight) * predicted_active_cells_sdrs
)
# list of timesteps specifying when a snapshot of the clusters will be taken
step = (end - start) / num_cluster_snapshots - 1
cluster_snapshot_indices = range(step, end - start, step)
# run clustering
(clustering_accuracies, cluster_snapshots, closest_cluster_history) = run(
sdrs,
categories,
anomaly_scores,
distance_func,
moving_average_window,
max_num_clusters,
ClusteringClass,
cluster_snapshot_indices,
)
# cluster_categories = []
# for c in closest_cluster_history:
# if c is not None:
# cluster_categories.append(c.label_distribution()[0]['label'])
# plot cluster assignments over time
for i in range(num_cluster_snapshots):
clusters = cluster_snapshots[i]
snapshot_index = cluster_snapshot_indices[i]
plot_cluster_assignments(plots_output_dir, clusters, snapshot_index)
# plot inter-cluster distance matrix
# plot_id = 'inter-cluster_t=%s' % snapshot_index
# plot_inter_sequence_distances(plots_output_dir,
# plot_id,
# distance_func,
# sdrs[:snapshot_index],
# cluster_categories[:snapshot_index],
# distance_matrix_ignore_noise)
# plot inter-category distance matrix
plot_id = "inter-category_t=%s " % snapshot_index
plot_inter_sequence_distances(
plots_output_dir,
plot_id,
distance_func,
sdrs[:snapshot_index],
categories[:snapshot_index],
distance_matrix_ignore_noise,
)
# plot clustering accuracy over time
plot_id = "file=%s | moving_average_window=%s" % (exp_name, moving_average_window)
plot_accuracy(plots_output_dir, plot_id, sensor_values, categories, anomaly_scores, clustering_accuracies, xlim)
if show_plots:
plt.show()