class VisdomMonitor(Wrapper):
def __init__(self, env, cmdl):
super(VisdomMonitor, self).__init__(env)
self.freq = cmdl.report_freq # in steps
self.cmdl = cmdl
if self.cmdl.display_plots:
self.vis = Visdom()
self.plot = self.vis.line(
Y=np.array([0]), X=np.array([0]),
opts=dict(
title=cmdl.label,
caption="Episodic reward per 1200 steps.")
)
self.step_cnt = 0
self.ep_cnt = -1
self.ep_rw = []
self.last_reported_ep = 0
def _step(self, action):
# self._before_step(action)
observation, reward, done, info = self.env.step(action)
done = self._after_step(observation, reward, done, info)
return observation, reward, done, info
def _reset(self):
self._before_reset()
observation = self.env.reset()
self._after_reset(observation)
return observation
def _after_step(self, o, r, done, info):
self.ep_rw[self.ep_cnt] += r
self.step_cnt += 1
if self.step_cnt % self.freq == 0:
self._update_plot()
return done
def _before_reset(self):
self.ep_rw.append(0)
def _after_reset(self, observation):
self.ep_cnt += 1
# print("[%2d][%4d] RESET" % (self.ep_cnt, self.step_cnt))
def _update_plot(self):
# print(self.last_reported_ep, self.ep_cnt + 1)
completed_eps = self.ep_rw[self.last_reported_ep:self.ep_cnt + 1]
ep_mean_reward = sum(completed_eps) / len(completed_eps)
if self.cmdl.display_plots:
self.vis.line(
X=np.array([self.step_cnt]),
Y=np.array([ep_mean_reward]),
win=self.plot,
update='append'
)
self.last_reported_ep = self.ep_cnt + 1
def set_model(self, model):
"""Save the model instance and initialize any required variables which would be updated throughout training.
Parameters
----------
model : :class:`~gensim.models.basemodel.BaseTopicModel`
The model for which the training will be reported (logged or visualized) by the callback.
"""
self.model = model
self.previous = None
# check for any metric which need model state from previous epoch
if any(isinstance(metric, (DiffMetric, ConvergenceMetric)) for metric in self.metrics):
self.previous = copy.deepcopy(model)
# store diff diagonals of previous epochs
self.diff_mat = Queue()
if any(metric.logger == "visdom" for metric in self.metrics):
if not VISDOM_INSTALLED:
raise ImportError("Please install Visdom for visualization")
self.viz = Visdom()
# store initial plot windows of every metric (same window will be updated with increasing epochs)
self.windows = []
if any(metric.logger == "shell" for metric in self.metrics):
# set logger for current topic model
self.log_type = logging.getLogger('gensim.models.ldamodel')
def testCallbackUpdateGraph(self):
# Popen have no context-manager in 2.7, for this reason - try/finally.
try:
# spawn visdom.server
proc = subprocess.Popen(['python', '-m', 'visdom.server', '-port', str(self.port)])
# wait for visdom server startup (any better way?)
time.sleep(3)
viz = Visdom(server=self.host, port=self.port)
assert viz.check_connection()
# clear screen
viz.close()
self.model.update(self.corpus)
finally:
proc.kill()
def __init__(self, env, cmdl):
super(VisdomMonitor, self).__init__(env)
self.freq = cmdl.report_freq # in steps
self.cmdl = cmdl
if self.cmdl.display_plots:
self.vis = Visdom()
self.plot = self.vis.line(
Y=np.array([0]), X=np.array([0]),
opts=dict(
title=cmdl.label,
caption="Episodic reward per 1200 steps.")
)
self.step_cnt = 0
self.ep_cnt = -1
self.ep_rw = []
self.last_reported_ep = 0
def __init__(self, env, cmdl):
super(EvaluationMonitor, self).__init__(env)
self.freq = cmdl.eval_freq # in steps
self.eval_steps = cmdl.evaluator.eval_steps
self.cmdl = cmdl
if self.cmdl.display_plots:
self.vis = Visdom()
self.plot = self.vis.line(
Y=np.array([0]), X=np.array([0]),
opts=dict(
title=cmdl.label,
caption="Episodic reward per %d steps." % self.eval_steps)
)
self.crt_step = 0
self.step_cnt = 0
self.ep_cnt = 0
self.total_rw = 0
self.max_mean_rw = -100
def __init__(self,
env=None,
log_checkpoints=True,
losses=('loss', ),
phases=('train', 'valid'),
server='http://localhost',
port=8097,
max_y=None,
**opts):
self.viz = None
if Visdom is not None:
self.viz = Visdom(server=server, port=port, env=env)
self.legend = ['%s.%s' % (p, l) for p in phases for l in losses]
opts.update({'legend': self.legend})
self.opts = opts
self.env = env
self.max_y = max_y
self.log_checkpoints = log_checkpoints
self.losses = set(losses)
self.last = {p: {l: None for l in losses} for p in phases}
self.pane = self._init_pane()
def main():
args = parser.parse_args()
cf = ConfigParser.ConfigParser()
try:
cf.read(args.conf)
except:
print("conf file not exists")
logger = init_logger(os.path.join(args.log_dir, 'train_cnn_lstm_ctc.log'))
dataset = cf.get('Data', 'dataset')
data_dir = cf.get('Data', 'data_dir')
feature_type = cf.get('Data', 'feature_type')
out_type = cf.get('Data', 'out_type')
n_feats = cf.getint('Data', 'n_feats')
batch_size = cf.getint("Training", 'batch_size')
#Data Loader
train_dataset = myDataset(data_dir, data_set='train', feature_type=feature_type, out_type=out_type, n_feats=n_feats)
train_loader = myCNNDataLoader(train_dataset, batch_size=batch_size, shuffle=True,
num_workers=4, pin_memory=False)
dev_dataset = myDataset(data_dir, data_set="test", feature_type=feature_type, out_type=out_type, n_feats=n_feats)
dev_loader = myCNNDataLoader(dev_dataset, batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=False)
#decoder for dev set
decoder = GreedyDecoder(dev_dataset.int2phone, space_idx=-1, blank_index=0)
#Define Model
rnn_input_size = cf.getint('Model', 'rnn_input_size')
rnn_hidden_size = cf.getint('Model', 'rnn_hidden_size')
rnn_layers = cf.getint('Model', 'rnn_layers')
rnn_type = RNN[cf.get('Model', 'rnn_type')]
bidirectional = cf.getboolean('Model', 'bidirectional')
batch_norm = cf.getboolean('Model', 'batch_norm')
num_class = cf.getint('Model', 'num_class')
drop_out = cf.getfloat('Model', 'num_class')
model = CNN_LSTM_CTC(rnn_input_size=rnn_input_size, rnn_hidden_size=rnn_hidden_size, rnn_layers=rnn_layers,
rnn_type=rnn_type, bidirectional=bidirectional, batch_norm=batch_norm,
num_class=num_class, drop_out=drop_out)
#model.apply(xavier_uniform_init)
print(model.name)
#Training
init_lr = cf.getfloat('Training', 'init_lr')
num_epoches = cf.getint('Training', 'num_epoches')
end_adjust_acc = cf.getfloat('Training', 'end_adjust_acc')
decay = cf.getfloat("Training", 'lr_decay')
weight_decay = cf.getfloat("Training", 'weight_decay')
try:
seed = cf.getint('Training', 'seed')
except:
seed = torch.cuda.initial_seed()
params = { 'num_epoches':num_epoches, 'end_adjust_acc':end_adjust_acc, 'seed':seed,
'decay':decay, 'learning_rate':init_lr, 'weight_decay':weight_decay, 'batch_size':batch_size,
'feature_type':feature_type, 'n_feats': n_feats, 'out_type': out_type }
if USE_CUDA:
torch.cuda.manual_seed(seed)
model = model.cuda()
print(params)
loss_fn = CTCLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=init_lr, weight_decay=weight_decay)
#visualization for training
from visdom import Visdom
viz = Visdom(env='863_corpus')
title = dataset+' '+feature_type+str(n_feats)+' CNN_LSTM_CTC'
opts = [dict(title=title+" Loss", ylabel = 'Loss', xlabel = 'Epoch'),
dict(title=title+" CER on Train", ylabel = 'CER', xlabel = 'Epoch'),
dict(title=title+' CER on DEV', ylabel = 'DEV CER', xlabel = 'Epoch')]
viz_window = [None, None, None]
count = 0
learning_rate = init_lr
acc_best = -100
acc_best_true = -100
adjust_rate_flag = False
stop_train = False
adjust_time = 0
start_time = time.time()
loss_results = []
training_cer_results = []
dev_cer_results = []
while not stop_train:
if count >= num_epoches:
break
count += 1
if adjust_rate_flag:
learning_rate *= decay
adjust_rate_flag = False
for param in optimizer.param_groups:
param['lr'] *= decay
print("Start training epoch: %d, learning_rate: %.5f" % (count, learning_rate))
logger.info("Start training epoch: %d, learning_rate: %.5f" % (count, learning_rate))
loss = train(model, train_loader, loss_fn, optimizer, logger, print_every=20)
loss_results.append(loss)
cer = dev(model, train_loader, decoder, logger)
print("cer on training set is %.4f" % cer)
logger.info("cer on training set is %.4f" % cer)
training_cer_results.append(cer)
acc = dev(model, dev_loader, decoder, logger)
dev_cer_results.append(acc)
#model_path_accept = './log/epoch'+str(count)+'_lr'+str(learning_rate)+'_cv'+str(acc)+'.pkl'
#model_path_reject = './log/epoch'+str(count)+'_lr'+str(learning_rate)+'_cv'+str(acc)+'_rejected.pkl'
if acc > (acc_best + end_adjust_acc):
acc_best = acc
adjust_rate_count = 0
model_state = copy.deepcopy(model.state_dict())
op_state = copy.deepcopy(optimizer.state_dict())
elif (acc > acc_best - end_adjust_acc):
adjust_rate_count += 1
if acc > acc_best and acc > acc_best_true:
acc_best_true = acc
model_state = copy.deepcopy(model.state_dict())
op_state = copy.deepcopy(optimizer.state_dict())
else:
adjust_rate_count = 0
#torch.save(model.state_dict(), model_path_reject)
print("adjust_rate_count:"+str(adjust_rate_count))
print('adjust_time:'+str(adjust_time))
logger.info("adjust_rate_count:"+str(adjust_rate_count))
logger.info('adjust_time:'+str(adjust_time))
if adjust_rate_count == 10:
adjust_rate_flag = True
adjust_time += 1
adjust_rate_count = 0
acc_best = acc_best_true
model.load_state_dict(model_state)
optimizer.load_state_dict(op_state)
if adjust_time == 8:
stop_train = True
time_used = (time.time() - start_time) / 60
print("epoch %d done, cv acc is: %.4f, time_used: %.4f minutes" % (count, acc, time_used))
logger.info("epoch %d done, cv acc is: %.4f, time_used: %.4f minutes" % (count, acc, time_used))
x_axis = range(count)
y_axis = [loss_results[0:count], training_cer_results[0:count], dev_cer_results[0:count]]
for x in range(len(viz_window)):
if viz_window[x] is None:
viz_window[x] = viz.line(X = np.array(x_axis), Y = np.array(y_axis[x]), opts = opts[x],)
else:
viz.line(X = np.array(x_axis), Y = np.array(y_axis[x]), win = viz_window[x], update = 'replace',)
print("End training, best cv acc is: %.4f" % acc_best)
logger.info("End training, best cv acc is: %.4f" % acc_best)
best_path = os.path.join(args.log_dir, 'best_model'+'_cv'+str(acc_best)+'.pkl')
cf.set('Model', 'model_file', best_path)
cf.write(open(args.conf, 'w'))
params['epoch']=count
torch.save(CNN_LSTM_CTC.save_package(model, optimizer=optimizer, epoch=params, loss_results=loss_results, training_cer_results=training_cer_results, dev_cer_results=dev_cer_results), best_path)
import numpy as np import torch import torch.nn as nn import torch.optim as optim import torch.nn.init as init from torch.autograd import Variable from visdom import Visdom viz = Visdom() num_data=1000 num_epoch=400 x = init.uniform(torch.Tensor(num_data,1),-10,10) y = init.uniform(torch.Tensor(num_data,1),-10,10) z = x**2 + y**2 x_noise = x + init.normal(torch.FloatTensor(num_data,1),std=0.5) y_noise = y + init.normal(torch.FloatTensor(num_data,1),std=0.5) z_noise = x_noise**2 + y_noise**2 data_noise = torch.cat([x_noise,y_noise,z_noise],1) # visualize data win_1=viz.scatter( X=data_noise, opts=dict( markersize=5, markercolor=np.ndarray(shape=[num_data,3],dtype=float,buffer=[51,153,255]*np.ones(shape=[num_data,3])) ) )
from visdom import Visdom
import numpy as np
import math
vis = Visdom()
vis.heatmap(X=np.outer(np.arange(1, 6), np.arange(1, 11)),
opts=dict(
columnnames=['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'],
rownames=['y1', 'y2', 'y3', 'y4', 'y5'],
colormap='Electric',
))
nn.Linear(200, 10),
nn.LeakyReLU(inplace=True),
)
def forward(self, x):
x = self.model(x)
return x
device = torch.device('cuda:0')
net = MLP().to(device)
optimizer = optim.SGD(net.parameters(), lr=learning_rate, weight_decay=0.01)
criteon = nn.CrossEntropyLoss().to(device)
viz = Visdom()
viz.line([0.], [0.], win='train_loss', opts=dict(title='train loss'))
viz.line([[0.0, 0.0]], [0.],
win='test',
opts=dict(title='test loss&acc.', legend=['loss', 'acc.']))
global_step = 0
for epoch in range(epochs):
for batch_idx, (data, target) in enumerate(train_loader):
data = data.view(-1, 28 * 28)
data, target = data.to(device), target.cuda()
logits = net(data)
loss = criteon(logits, target)
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from visdom import Visdom
import time
import numpy as np
try:
viz = Visdom()
startup_sec = 1
while not viz.check_connection() and startup_sec > 0:
time.sleep(0.1)
startup_sec -= 0.1
assert viz.check_connection(), 'No connection could be formed quickly'
# image callback demo
def show_color_image_window(color, win=None):
image = np.full([3, 256, 256], color, dtype=float)
return viz.image(
image,
opts=dict(title='Colors', caption='Press arrows to alter color.'),
win=win
)
import MyDataloader
from TumorNetwithoutSource import *
import csv
import pandas as pd
import SimpleITK as sitk
from medpy import metric
import numpy as np
import time
import shutil
import sys
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
from visdom import Visdom
viz = Visdom(env='PiaNet TumorNet without Source 134')
viz.line([0], [0], win='train')
viz.line([0], [0], win='valid')
viz.line([0], [0], win='tumor')
#################initialization network##############
def weights_init(model):
if isinstance(model, nn.Conv3d) or isinstance(model, nn.ConvTranspose3d):
nn.init.kaiming_uniform_(model.weight.data, 0.25)
nn.init.constant_(model.bias.data, 0)
# elif isinstance(model, nn.InstanceNorm3d):
# nn.init.constant_(model.weight.data,1.0)
# nn.init.constant_(model.bias.data, 0)
class Callback(object):
"""A class representing routines called reactively at specific phases during trained.
These can be used to log or visualize the training progress using any of the metric scores developed before.
The values are stored at the end of each training epoch. The following metric scores are currently available:
* :class:`~gensim.models.callbacks.CoherenceMetric`
* :class:`~gensim.models.callbacks.PerplexityMetric`
* :class:`~gensim.models.callbacks.DiffMetric`
* :class:`~gensim.models.callbacks.ConvergenceMetric`
"""
def __init__(self, metrics):
"""
Parameters
----------
metrics : list of :class:`~gensim.models.callbacks.Metric`
The list of metrics to be reported by the callback.
"""
self.metrics = metrics
def set_model(self, model):
"""Save the model instance and initialize any required variables which would be updated throughout training.
Parameters
----------
model : :class:`~gensim.models.basemodel.BaseTopicModel`
The model for which the training will be reported (logged or visualized) by the callback.
"""
self.model = model
self.previous = None
# check for any metric which need model state from previous epoch
if any(isinstance(metric, (DiffMetric, ConvergenceMetric)) for metric in self.metrics):
self.previous = copy.deepcopy(model)
# store diff diagonals of previous epochs
self.diff_mat = Queue()
if any(metric.logger == "visdom" for metric in self.metrics):
if not VISDOM_INSTALLED:
raise ImportError("Please install Visdom for visualization")
self.viz = Visdom()
# store initial plot windows of every metric (same window will be updated with increasing epochs)
self.windows = []
if any(metric.logger == "shell" for metric in self.metrics):
# set logger for current topic model
self.log_type = logging.getLogger('gensim.models.ldamodel')
def on_epoch_end(self, epoch, topics=None):
"""Report the current epoch's metric value.
Called at the end of each training iteration.
Parameters
----------
epoch : int
The epoch that just ended.
topics : list of list of str, optional
List of tokenized topics. This is required for the coherence metric.
Returns
-------
dict of (str, object)
Mapping from metric names to their values. The type of each value depends on the metric type,
for example :class:`~gensim.models.callbacks.DiffMetric` computes a matrix while
:class:`~gensim.models.callbacks.ConvergenceMetric` computes a float.
"""
# stores current epoch's metric values
current_metrics = {}
# plot all metrics in current epoch
for i, metric in enumerate(self.metrics):
label = str(metric)
value = metric.get_value(topics=topics, model=self.model, other_model=self.previous)
current_metrics[label] = value
if metric.logger == "visdom":
if epoch == 0:
if value.ndim > 0:
diff_mat = np.array([value])
viz_metric = self.viz.heatmap(
X=diff_mat.T, env=metric.viz_env, opts=dict(xlabel='Epochs', ylabel=label, title=label)
)
# store current epoch's diff diagonal
self.diff_mat.put(diff_mat)
# saving initial plot window
self.windows.append(copy.deepcopy(viz_metric))
else:
viz_metric = self.viz.line(
Y=np.array([value]), X=np.array([epoch]), env=metric.viz_env,
opts=dict(xlabel='Epochs', ylabel=label, title=label)
)
# saving initial plot window
self.windows.append(copy.deepcopy(viz_metric))
else:
if value.ndim > 0:
# concatenate with previous epoch's diff diagonals
diff_mat = np.concatenate((self.diff_mat.get(), np.array([value])))
self.viz.heatmap(
X=diff_mat.T, env=metric.viz_env, win=self.windows[i],
opts=dict(xlabel='Epochs', ylabel=label, title=label)
)
self.diff_mat.put(diff_mat)
else:
self.viz.line(
Y=np.array([value]),
X=np.array([epoch]),
env=metric.viz_env,
win=self.windows[i],
update='append'
)
if metric.logger == "shell":
statement = "".join(("Epoch ", str(epoch), ": ", label, " estimate: ", str(value)))
self.log_type.info(statement)
# check for any metric which need model state from previous epoch
if any(isinstance(metric, (DiffMetric, ConvergenceMetric)) for metric in self.metrics):
self.previous = copy.deepcopy(self.model)
return current_metrics
def get_c(inflection, b):
"""
c = log(b) / inflection
"""
return math.log(b) / inflection
if __name__ == '__main__':
from argparse import ArgumentParser
parser = ArgumentParser()
parser.add_argument('--env', default='sigmoid_schedule')
parser.add_argument('--inflection', default=5000, type=int)
args = parser.parse_args()
from visdom import Visdom
viz = Visdom(env=args.env)
import numpy as np
Y = np.linspace(1, args.inflection * 2, 1000)
for b in (10, 100, 1000):
c = get_c(args.inflection, b)
title = 'c=%.g;b=%d;inflection=%d' % (c, b, args.inflection)
viz.line(np.array([generic_sigmoid(b=b, c=c)(i) for i in Y]), Y,
opts={'title': title})
for c in (0.0001, 0.001, 0.005):
b = get_b(args.inflection, c)
title = 'c=%.g;b=%d;inflection=%d' % (c, b, args.inflection)
viz.line(np.array([generic_sigmoid(b=b, c=c)(i) for i in Y]), Y,
opts={'title': title})
def __init__(self, env_name='main'):
self.viz = Visdom()
self.env = env_name
self.scatters = {}
image = np.transpose(image, (2, 0, 1))
all_timesteps = tx
all_rewards = ty
# np.save('./results_temp/rewards.npy', all_rewards)
# np.save('./results_temp/timesteps.npy', all_timesteps)
# return viz.image(image, win=win)
return viz.image(image, win=win)
def get_stats(folder, smooth=1, bin_size=100):
tx, ty = load_data(folder, smooth, bin_size)
all_timesteps = tx
all_rewards = ty
return all_timesteps, all_rewards
if __name__ == "__main__":
from visdom import Visdom
viz = Visdom()
visdom_plot(viz,
None,
'./results_temp/',
'BreakOut',
'a2c',
bin_size=100,
smooth=1)
def __init__(self):
try:
from visdom import Visdom
viz = Visdom(port=8097, server='http://localhost')
except socket_error as e:
viz = None
def __init__(self, server, port, env_name='main'):
self.viz = Visdom(server=server, port=port)
self.env = env_name
self.plots = {}
self.opts = {}
def main():
print("#######")
print(
"WARNING: All rewards are clipped or normalized so you need to use a monitor (see envs.py) or visdom plot to get true rewards"
)
print("#######")
os.environ['OMP_NUM_THREADS'] = '1'
if args.vis:
from visdom import Visdom
viz = Visdom()
win = None
envs = [
make_env(args.env_name, args.seed, i, args.log_dir)
for i in range(args.num_processes)
]
if args.num_processes > 1:
envs = SubprocVecEnv(envs)
else:
envs = DummyVecEnv(envs)
if len(envs.observation_space.shape) == 1:
envs = VecNormalize(envs)
obs_shape = envs.observation_space.shape
obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:]
) # I guess the obs_shape[0] is channel number
if len(envs.observation_space.shape) == 3:
actor_critic = CNNPolicy(obs_shape[0], envs.action_space,
args.recurrent_policy)
else:
assert not args.recurrent_policy, \
"Recurrent policy is not implemented for the MLP controller"
actor_critic = MLPPolicy(obs_shape[0], envs.action_space)
if envs.action_space.__class__.__name__ == "Discrete":
action_shape = 1
else:
action_shape = envs.action_space.shape[0]
if args.cuda:
actor_critic.cuda()
if args.algo == 'a2c':
optimizer = optim.RMSprop(actor_critic.parameters(),
args.lr,
eps=args.eps,
alpha=args.alpha)
elif args.algo == 'ppo':
optimizer = optim.Adam(actor_critic.parameters(),
args.lr,
eps=args.eps)
elif args.algo == 'acktr':
optimizer = KFACOptimizer(actor_critic)
rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape,
envs.action_space, actor_critic.state_size)
current_obs = torch.zeros(args.num_processes, *obs_shape)
def update_current_obs(obs):
shape_dim0 = envs.observation_space.shape[0]
obs = torch.from_numpy(obs).float()
if args.num_stack > 1:
current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:]
current_obs[:, -shape_dim0:] = obs
obs = envs.reset()
update_current_obs(obs)
rollouts.observations[0].copy_(current_obs)
# These variables are used to compute average rewards for all processes.
episode_rewards = torch.zeros([args.num_processes, 1])
final_rewards = torch.zeros([args.num_processes, 1])
if args.cuda:
current_obs = current_obs.cuda()
rollouts.cuda()
start = time.time()
for j in range(num_updates):
for step in range(args.num_steps):
# args.num_steps should be the length of interactions before each updating/training
# Sample actions
value, action, action_log_prob, states = actor_critic.act(
Variable(rollouts.observations[step], volatile=True),
Variable(rollouts.states[step], volatile=True),
Variable(rollouts.masks[step], volatile=True))
cpu_actions = action.data.squeeze(1).cpu().numpy(
) # returns are state value, sampled action, act_log_prob, hidden states
# Obser reward and next obs
obs, reward, done, info = envs.step(cpu_actions)
reward = torch.from_numpy(np.expand_dims(np.stack(reward),
1)).float()
episode_rewards += reward
# If done then clean the history of observations.
masks = torch.FloatTensor([[0.0] if done_ else [1.0]
for done_ in done])
final_rewards *= masks
final_rewards += (1 - masks) * episode_rewards
episode_rewards *= masks
if args.cuda:
masks = masks.cuda()
if current_obs.dim() == 4:
current_obs *= masks.unsqueeze(2).unsqueeze(2)
else:
current_obs *= masks
update_current_obs(obs)
rollouts.insert(
step, current_obs, states.data, action.data,
action_log_prob.data, value.data, reward, masks
) # so the rollout stores one batch of interaction sequences, each sequence has length of args.num_steps
next_value = actor_critic(
Variable(rollouts.observations[-1], volatile=True),
Variable(rollouts.states[-1], volatile=True),
Variable(rollouts.masks[-1], volatile=True))[0].data
rollouts.compute_returns(next_value, args.use_gae, args.gamma,
args.tau)
if args.algo in ['a2c', 'acktr']:
values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(
Variable(rollouts.observations[:-1].view(-1, *obs_shape)),
Variable(rollouts.states[0].view(-1, actor_critic.state_size)),
Variable(rollouts.masks[:-1].view(-1, 1)),
Variable(rollouts.actions.view(-1, action_shape)))
# values should be values of observations, states are the hidden states used in rnn module, by pwang8
values = values.view(
args.num_steps, args.num_processes,
1) # values are estimated current state values
action_log_probs = action_log_probs.view(args.num_steps,
args.num_processes, 1)
# rollouts.returns are current "Action" value calculted following Bellmans' eqaution gamma * State_value(t+1) + reward(t)
advantages = Variable(
rollouts.returns[:-1]
) - values # This is also the definition of advantage value (action_value - state_value).
value_loss = advantages.pow(
2).mean() # values are estimated current state_value(t)
action_loss = -(Variable(advantages.data) *
action_log_probs).mean()
# If ACKTR is utilized, it is not only a different optimizer is used, they also added some new loss source
if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0:
# Sampled fisher, see Martens 2014
actor_critic.zero_grad()
pg_fisher_loss = -action_log_probs.mean()
value_noise = Variable(torch.randn(values.size()))
if args.cuda:
value_noise = value_noise.cuda()
sample_values = values + value_noise
vf_fisher_loss = -(
values - Variable(sample_values.data)
).pow(2).mean(
) # don't know what is the difference between this and just randomly sample some noise
fisher_loss = pg_fisher_loss + vf_fisher_loss
optimizer.acc_stats = True
fisher_loss.backward(retain_graph=True)
optimizer.acc_stats = False
optimizer.zero_grad()
(value_loss * args.value_loss_coef + action_loss -
dist_entropy * args.entropy_coef).backward()
if args.algo == 'a2c':
nn.utils.clip_grad_norm(actor_critic.parameters(),
args.max_grad_norm)
optimizer.step()
elif args.algo == 'ppo':
advantages = rollouts.returns[:-1] - rollouts.value_preds[:
-1] # calculating the advantage value of an action
advantages = (advantages - advantages.mean()) / (advantages.std() +
1e-5)
# The difference from this ppo optimization to the optimization above is that: it updates params for
# multiple epochs in ppo optimization. Because of this, it samples from the rollouts storage a minibatch
# every time to calculate gradient. Sampling is conducted for optimization purpose.
for e in range(args.ppo_epoch):
if args.recurrent_policy:
data_generator = rollouts.recurrent_generator(
advantages, args.num_mini_batch)
else:
data_generator = rollouts.feed_forward_generator(
advantages, args.num_mini_batch)
for sample in data_generator:
observations_batch, states_batch, actions_batch, \
return_batch, masks_batch, old_action_log_probs_batch, \
adv_targ = sample
# Reshape to do in a single forward pass for all steps
values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(
Variable(observations_batch), Variable(states_batch),
Variable(masks_batch), Variable(actions_batch))
# For the 1st epoch of updating, I guess the action_log_probls is the same as old_action_log_probs_batch
# because params of the NN have not been updated at that time. But later, in other updating epochs,
# this ratio will generate some error. The old_action_log_probs_batch will not be updated during
# these param updating epochs.
# action_log_probs is the log prob of that action taken by the agent. So it's one value here, not
# log_prob for all actions with certain input observation/state. By pwang8, Dec 31, 2017
adv_targ = Variable(adv_targ)
ratio = torch.exp(action_log_probs -
Variable(old_action_log_probs_batch))
surr1 = ratio * adv_targ
surr2 = torch.clamp(ratio, 1.0 - args.clip_param,
1.0 + args.clip_param) * adv_targ
action_loss = -torch.min(
surr1,
surr2).mean() # PPO's pessimistic surrogate (L^CLIP)
# compared to a2c, the major difference for ppo is that action_loss is calculated in controlled way
value_loss = (Variable(return_batch) -
values).pow(2).mean()
optimizer.zero_grad()
(value_loss + action_loss -
dist_entropy * args.entropy_coef).backward()
nn.utils.clip_grad_norm(actor_critic.parameters(),
args.max_grad_norm)
optimizer.step()
rollouts.after_update()
if j % args.save_interval == 0 and args.save_dir != "":
save_path = os.path.join(args.save_dir, args.algo)
try:
os.makedirs(save_path)
except OSError:
pass
# A really ugly way to save a model to CPU
save_model = actor_critic
if args.cuda:
save_model = copy.deepcopy(actor_critic).cpu()
save_model = [
save_model,
hasattr(envs, 'ob_rms') and envs.ob_rms or None
]
torch.save(save_model,
os.path.join(save_path, args.env_name + ".pt"))
if j % args.log_interval == 0:
end = time.time()
total_num_steps = (j + 1) * args.num_processes * args.num_steps
print(
"Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}"
.format(j, total_num_steps,
int(total_num_steps / (end - start)),
final_rewards.mean(), final_rewards.median(),
final_rewards.min(), final_rewards.max(),
dist_entropy.data[0], value_loss.data[0],
action_loss.data[0]))
if args.vis and j % args.vis_interval == 0:
try:
# Sometimes monitor doesn't properly flush the outputs
win = visdom_plot(viz, win, args.log_dir, args.env_name,
args.algo)
except IOError:
pass
# -*- coding: utf-8 -*- """ Created on Mon Jul 23 10:43:59 2018 @author: Administrator """ from visdom import Visdom vis = Visdom() svgstr = """ <svg height="300" width="300"> <ellipse cx="80" cy="80" rx="50" ry="30" style="fill:red;stroke:purple;stroke-width:2" /> 抱歉，你的浏览器不支持在线显示SVG对象. </svg> """ vis.svg(svgstr=svgstr, opts=dict(title='SVG图像'))
class VisdomLinePlotter(object):
"""Plots to Visdom"""
def __init__(self, env_name='main'):
self.viz = Visdom()
self.env = env_name
self.plots = {}
self.moving = {}
self.moving_counter = {}
self.rotation = torch.Tensor(
R.from_euler('zyz', [-90, 0, 0], degrees=True).as_dcm())
print("Rotation", self.rotation)
def plot_point_cloud(self, points, var_name):
if (len(list(points.shape)) == 3):
points = points.view(3, -1)
points = points.transpose(1, 0)
points = torch.mm(points, self.rotation)
# Rotate matrix to plot better
print(points.shape)
if var_name not in self.plots:
self.plots[var_name] = self.viz.scatter(X=points,
env=self.env,
name="pointplot",
opts=dict(markersize=1))
# print("YOU HAVE 10s to rotate the graph")
# time.sleep(10)
else:
self.viz.scatter(X=points,
env=self.env,
name="pointplot",
win=self.plots[var_name],
opts=dict(markersize=1))
def plot(self, var_name, split_name, title_name, x, y):
if var_name not in self.plots:
self.plots[var_name] = self.viz.line(X=np.array([x, x]),
Y=np.array([y, y]),
env=self.env,
opts=dict(legend=[split_name],
title=title_name,
xlabel='Epochs',
ylabel=var_name))
else:
self.viz.line(X=np.array([x]),
Y=np.array([y]),
env=self.env,
win=self.plots[var_name],
name=split_name,
update='append')
def plot_moving_avg(self,
var_name,
split_name,
title_name,
x,
y,
window=10):
if var_name + split_name not in self.moving:
self.moving[var_name + split_name] = np.zeros(window)
self.moving_counter[var_name + split_name] = 0
self.moving_counter[var_name + split_name] += 1
index = self.moving_counter[var_name + split_name]
self.moving[var_name + split_name][index % window] = y
if self.moving_counter[var_name + split_name] >= window:
avg = self.moving[var_name + split_name].mean()
if var_name not in self.plots:
self.plots[var_name] = self.viz.line(X=np.array([avg, avg]),
Y=np.array([y, y]),
env=self.env,
opts=dict(
legend=[split_name],
title=title_name,
xlabel='Epochs',
ylabel=var_name))
else:
self.viz.line(X=np.array([x]),
Y=np.array([avg]),
env=self.env,
win=self.plots[var_name],
name=split_name,
update='append')
def __init__(self, env_name='main', heatsize=None):
self.vis = Visdom()
self.env = env_name
self.hmaps = {}
self.heatsize = heatsize
class Visualizer:
def __init__(self, env="main",
server="http://localhost",
port=8097,
base_url="/",
http_proxy_host=None,
http_proxy_port=None,
log_to_filename=None):
self._viz = Visdom(env=env,
server=server,
port=port,
http_proxy_host=http_proxy_host,
http_proxy_port=http_proxy_port,
log_to_filename=log_to_filename,
use_incoming_socket=False)
self._viz.close(env=env)
def plot_line(self, values, steps, name, legend=None):
if legend is None:
opts = dict(title=name)
else:
opts = dict(title=name, legend=legend)
self._viz.line(
X=numpy.column_stack(steps),
Y=numpy.column_stack(values),
win=name,
update='append',
opts=opts
)
def plot_text(self, text, title, pre=True):
_width = max([len(x) for x in text.split("\n")]) * 10
_heigth = len(text.split("\n")) * 20
_heigth = max(_heigth, 120)
if pre:
text = "<pre>{}</pre>".format(text)
self._viz.text(text, win=title, opts=dict(title=title,
width=min(_width, 400),
height=min(_heigth, 400)))
def plot_bar(self, data, labels, title):
self._viz.bar(win=title, X=data,
opts=dict(legend=labels, stacked=False, title=title))
def plot_scatter(self, data, labels, title):
X = numpy.concatenate(data, axis=0)
Y = numpy.concatenate([numpy.full(len(d), i)
for i, d in enumerate(data, 1)], axis=0)
self._viz.scatter(win=title, X=X, Y=Y,
opts=dict(legend=labels, title=title,
markersize=5,
webgl=True,
width=400,
height=400,
markeropacity=0.5))
def plot_heatmap(self, data, labels, title):
self._viz.heatmap(win=title,
X=data,
opts=dict(
title=title,
columnnames=labels[1],
rownames=labels[0],
width=700,
height=700,
layoutopts={'plotly': {
'xaxis': {
'side': 'top',
'tickangle': -60,
# 'autorange': "reversed"
},
'yaxis': {
'autorange': "reversed"
},
}
}
))
"_" + Configure.dataDistribution + \
"_" + dataParameter
if Configure.standard == 'max':
name += "_max"
# if Configure.cases!=None:
# name+=Configure.cases
if Configure.relate:
name += "_relate"
pool = None
if Configure.mulThread > 0:
pool = threadpool.ThreadPool(Configure.mulThread)
id = 0
viz = Visdom(server="114.212.82.243", env=name)
class experimentThread(threading.Thread):
def __init__(self, _allDeviceNum, _ratio, errorDistribution,
errorParameter, dataDistribution, dataParameter, id):
threading.Thread.__init__(self)
self._allDeviceNum = _allDeviceNum
self._ratio = _ratio
self.errorDistribution = errorDistribution
self.errorParameter = errorParameter
self.dataDistribution = dataDistribution
self.dataParameter = dataParameter
self.id = id
self.success = False
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
import torchvision
import torchvision.transforms as transforms
from models import nin
from visdom import Visdom
import thop
import time
from thop import profile
from models import nin_gc
from models import standard_dw
viz = Visdom()
line = viz.line(np.arange(10))
# 随机种子——训练结果可复现
def setup_seed(seed):
# 为CPU设置种子用于生成随机数,以使得结果是确定的
torch.manual_seed(seed)
#torch.cuda.manual_seed(seed)
# 为GPU设置种子用于生成随机数,以使得结果是确定的
torch.cuda.manual_seed_all(seed)
# 为numpy设置种子用于生成随机数,以使得结果是确定的
np.random.seed(seed)
# 将会让程序在开始时花费一点额外时间，为整个网络的每个卷积层
# 搜索最适合它的卷积实现算法，进而实现网络的加速。适用场景是
def get_scheduler(optimizer, opt):
"""Return a learning rate scheduler
Parameters:
optimizer -- the optimizer of the network
opt (option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions锛庛€€
opt.lr_policy is the name of learning rate policy: linear | step | plateau | cosine
For 'linear', we keep the same learning rate for the first <opt.niter> epochs
and linearly decay the rate to zero over the next <opt.niter_decay> epochs.
For other schedulers (step, plateau, and cosine), we use the default PyTorch schedulers.
See https://pytorch.org/docs/stable/optim.html for more details.
"""
if opt.lr_policy == 'linear':
def lambda_rule(step):
lr_l = 1.0 - max(0, step - opt.niter) / float(opt.niter_decay + 1)
return lr_l
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda_rule)
elif opt.lr_policy == 'step':
scheduler = lr_scheduler.StepLR(optimizer,
step_size=opt.lr_decay_iters,
gamma=0.1)
elif opt.lr_policy == 'plateau':
#scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.2, threshold=0.01, patience=5)
scheduler = lr_scheduler.ReduceLROnPlateau(
optimizer,
mode='min',
factor=opt.lr_reduce_factor,
threshold=opt.lr_reduce_threshold,
patience=opt.step_patience,
min_lr=opt.min_lr)
elif opt.lr_policy == 'cosine':
scheduler = lr_scheduler.CosineAnnealingLR(optimizer,
T_max=opt.niter,
eta_min=0)
elif opt.lr_policy == 'warmup':
# opt.factor = 2
# opt.warmup = 4000
# optimizer, factor, d_model, warmup_steps=4000, init_steps = 0
scheduler = WarmupLR(optimizer=optimizer,
factor=opt.lr_factor,
d_model=opt.d_model,
warmup_steps=opt.warmup_step,
init_steps=opt.steps,
visdom_freq_steps=opt.visdom_freq_steps)
from visdom import Visdom
vis = Visdom(server=opt.display_server,
port=opt.display_port,
env=opt.visdom_id + 'Learning Rate',
raise_exceptions=True)
scheduler.set_visdom(visdom_lr=opt.visdom_lr, vis=vis)
else:
return NotImplementedError(
'learning rate policy [%s] is not implemented', opt.lr_policy)
return scheduler
def __init__(self, env_name='main'):
self.viz = Visdom()
self.env = env_name
self.plots = {}
help='IP of the visdom server')
parser.add_argument('--visdom_port',
type=int,
default=DEFAULT_VISDOM_PORT,
help='IP port of the visdom server')
parser.add_argument('--value',
type=str,
default=DEFAULT_DATA_VALUE,
help='Y value for the line plot')
args = parser.parse_args()
print("Connecting to visdom server on ", args.visdom_host, ":",
args.visdom_port)
value = args.value
viz = Visdom(server="http://" + args.visdom_host, port=args.visdom_port)
assert viz.check_connection()
#if not viz.win_exists(WIN):
# viz.text("Bitcoin variation notifications:\n", win=WIN)
if os.path.exists(store_file):
f = open(store_file, 'rb')
iteration = int(pickle.load(f))
f.close()
else:
iteration = 0
print(iteration, value)
if iteration == 0:
def main():
os.environ['OMP_NUM_THREADS'] = '1'
if args.vis:
from visdom import Visdom
viz = Visdom()
win = None
envs = [
make_env(args.env_name, args.seed, i, args.log_dir,
args.start_container) for i in range(args.num_processes)
]
if args.num_processes > 1:
envs = SubprocVecEnv(envs)
else:
envs = DummyVecEnv(envs)
obs_shape = envs.observation_space.shape
obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:])
obs_numel = reduce(operator.mul, obs_shape, 1)
if len(obs_shape) == 3 and obs_numel > 1024:
actor_critic = CNNPolicy(obs_shape[0], envs.action_space,
args.recurrent_policy)
else:
assert not args.recurrent_policy, \
"Recurrent policy is not implemented for the MLP controller"
actor_critic = MLPPolicy(obs_numel, envs.action_space)
modelSize = 0
for p in actor_critic.parameters():
pSize = reduce(operator.mul, p.size(), 1)
modelSize += pSize
print(str(actor_critic))
print('Total model size: %d' % modelSize)
if envs.action_space.__class__.__name__ == "Discrete":
action_shape = 1
else:
action_shape = envs.action_space.shape[0]
if args.cuda:
actor_critic.cuda()
if args.algo == 'a2c':
optimizer = optim.RMSprop(actor_critic.parameters(),
args.lr,
eps=args.eps,
alpha=args.alpha)
elif args.algo == 'ppo':
optimizer = optim.Adam(actor_critic.parameters(),
args.lr,
eps=args.eps)
elif args.algo == 'acktr':
optimizer = KFACOptimizer(actor_critic)
rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape,
envs.action_space, actor_critic.state_size)
current_obs = torch.zeros(args.num_processes, *obs_shape)
def update_current_obs(obs):
shape_dim0 = envs.observation_space.shape[0]
obs = torch.from_numpy(obs).float()
if args.num_stack > 1:
current_obs[:, :-shape_dim0] = current_obs[:, shape_dim0:]
current_obs[:, -shape_dim0:] = obs
obs = envs.reset()
update_current_obs(obs)
rollouts.observations[0].copy_(current_obs)
# These variables are used to compute average rewards for all processes.
episode_rewards = torch.zeros([args.num_processes, 1])
final_rewards = torch.zeros([args.num_processes, 1])
reward_avg = 0
if args.cuda:
current_obs = current_obs.cuda()
rollouts.cuda()
start = time.time()
for j in range(num_updates):
for step in range(args.num_steps):
# Sample actions
value, action, action_log_prob, states = actor_critic.act(
Variable(rollouts.observations[step]),
Variable(rollouts.states[step]),
Variable(rollouts.masks[step]))
cpu_actions = action.data.squeeze(1).cpu().numpy()
# Observation, reward and next obs
obs, reward, done, info = envs.find_action(cpu_actions)
# Maxime: clip the reward within [0,1] for more reliable training
# This code deals poorly with large reward values
reward = np.clip(reward, a_min=0, a_max=None) / 400
reward = torch.from_numpy(np.expand_dims(np.stack(reward),
1)).float()
episode_rewards += reward
# If done then clean the history of observations.
masks = torch.FloatTensor([[0.0] if done_ else [1.0]
for done_ in done])
final_rewards *= masks
final_rewards += (1 - masks) * episode_rewards
episode_rewards *= masks
if args.cuda:
masks = masks.cuda()
if current_obs.dim() == 4:
current_obs *= masks.unsqueeze(2).unsqueeze(2)
else:
current_obs *= masks
update_current_obs(obs)
rollouts.insert(step, current_obs, states.data, action.data,
action_log_prob.data, value.data, reward, masks)
next_value = actor_critic(Variable(rollouts.observations[-1]),
Variable(rollouts.states[-1]),
Variable(rollouts.masks[-1]))[0].data
rollouts.compute_returns(next_value, args.use_gae, args.gamma,
args.tau)
if args.algo in ['a2c', 'acktr']:
values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(
Variable(rollouts.observations[:-1].view(-1, *obs_shape)),
Variable(rollouts.states[0].view(-1, actor_critic.state_size)),
Variable(rollouts.masks[:-1].view(-1, 1)),
Variable(rollouts.actions.view(-1, action_shape)))
values = values.view(args.num_steps, args.num_processes, 1)
action_log_probs = action_log_probs.view(args.num_steps,
args.num_processes, 1)
advantages = Variable(rollouts.returns[:-1]) - values
value_loss = advantages.pow(2).mean()
action_loss = -(Variable(advantages.data) *
action_log_probs).mean()
if args.algo == 'acktr' and optimizer.steps % optimizer.Ts == 0:
# Sampled fisher, see Martens 2014
actor_critic.zero_grad()
pg_fisher_loss = -action_log_probs.mean()
value_noise = Variable(torch.randn(values.size()))
if args.cuda:
value_noise = value_noise.cuda()
sample_values = values + value_noise
vf_fisher_loss = -(values -
Variable(sample_values.data)).pow(2).mean()
fisher_loss = pg_fisher_loss + vf_fisher_loss
optimizer.acc_stats = True
fisher_loss.backward(retain_graph=True)
optimizer.acc_stats = False
optimizer.zero_grad()
(value_loss * args.value_loss_coef + action_loss -
dist_entropy * args.entropy_coef).backward()
if args.algo == 'a2c':
nn.utils.clip_grad_norm(actor_critic.parameters(),
args.max_grad_norm)
optimizer.step()
elif args.algo == 'ppo':
advantages = rollouts.returns[:-1] - rollouts.value_preds[:-1]
advantages = (advantages - advantages.mean()) / (advantages.std() +
1e-5)
for e in range(args.ppo_epoch):
if args.recurrent_policy:
data_generator = rollouts.recurrent_generator(
advantages, args.num_mini_batch)
else:
data_generator = rollouts.feed_forward_generator(
advantages, args.num_mini_batch)
for sample in data_generator:
observations_batch, states_batch, actions_batch, \
return_batch, masks_batch, old_action_log_probs_batch, \
adv_targ = sample
# Reshape to do in a single forward pass for all steps
values, action_log_probs, dist_entropy, states = actor_critic.evaluate_actions(
Variable(observations_batch), Variable(states_batch),
Variable(masks_batch), Variable(actions_batch))
adv_targ = Variable(adv_targ)
ratio = torch.exp(action_log_probs -
Variable(old_action_log_probs_batch))
surr1 = ratio * adv_targ
surr2 = torch.clamp(ratio, 1.0 - args.clip_param,
1.0 + args.clip_param) * adv_targ
action_loss = -torch.min(
surr1,
surr2).mean() # PPO's pessimistic surrogate (L^CLIP)
value_loss = (Variable(return_batch) -
values).pow(2).mean()
optimizer.zero_grad()
(value_loss + action_loss -
dist_entropy * args.entropy_coef).backward()
nn.utils.clip_grad_norm(actor_critic.parameters(),
args.max_grad_norm)
optimizer.step()
rollouts.after_update()
if j % args.save_interval == 0 and args.save_dir != "":
save_path = os.path.join(args.save_dir, args.algo)
try:
os.makedirs(save_path)
except OSError:
pass
# A really ugly way to save a model to CPU
save_model = actor_critic
if args.cuda:
save_model = copy.deepcopy(actor_critic).cpu()
save_model = [
save_model,
hasattr(envs, 'ob_rms') and envs.ob_rms or None
]
torch.save(save_model,
os.path.join(save_path, args.env_name + ".pt"))
if j % args.log_interval == 0:
reward_avg = 0.99 * reward_avg + 0.01 * final_rewards.mean()
end = time.time()
total_num_steps = (j + 1) * args.num_processes * args.num_steps
print(
"Updates {}, num timesteps {}, FPS {}, running avg reward {:.3f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}"
.format(j, total_num_steps,
int(total_num_steps / (end - start)), reward_avg,
dist_entropy.data[0], value_loss.data[0],
action_loss.data[0]))
"""
print("Updates {}, num timesteps {}, FPS {}, mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}, entropy {:.5f}, value loss {:.5f}, policy loss {:.5f}".
format(
j,
total_num_steps,
int(total_num_steps / (end - start)),
final_rewards.mean(),
final_rewards.median(),
final_rewards.min(),
final_rewards.max(), dist_entropy.data[0],
value_loss.data[0], action_loss.data[0])
)
"""
if args.vis and j % args.vis_interval == 0:
try:
# Sometimes monitor doesn't properly flush the outputs
win = visdom_plot(viz, win, args.log_dir, args.env_name,
args.algo)
except IOError:
pass
main_proc = True
if args.distributed:
if args.gpu_rank:
torch.cuda.set_device(int(args.gpu_rank))
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
world_size=args.world_size, rank=args.rank)
main_proc = args.rank == 0 # Only the first proc should save models
save_folder = args.save_folder
loss_results, cer_results, wer_results = torch.Tensor(args.epochs), torch.Tensor(args.epochs), torch.Tensor(
args.epochs)
best_wer = None
if args.visdom and main_proc:
from visdom import Visdom
viz = Visdom()
opts = dict(title=args.id, ylabel='', xlabel='Epoch', legend=['Loss', 'WER', 'CER'])
viz_window = None
epochs = torch.arange(1, args.epochs + 1)
if args.tensorboard and main_proc:
os.makedirs(args.log_dir, exist_ok=True)
from tensorboardX import SummaryWriter
tensorboard_writer = SummaryWriter(args.log_dir)
os.makedirs(save_folder, exist_ok=True)
avg_loss, start_epoch, start_iter = 0, 0, 0
if args.continue_from: # Starting from previous model
print("Loading checkpoint model %s" % args.continue_from)
package = torch.load(args.continue_from, map_location=lambda storage, loc: storage)
model = DeepSpeech.load_model_package(package)
class Visualizer(object):
def __init__(self, name, save=True, output_dir="."):
self.visdom = Visdom()
self.name = name
self.plots = dict()
self.save = save
if save:
if not os.path.exists(output_dir):
raise ValueError("output_dir does not exists")
# output directory for reconstructions
self.recon_dir = os.path.join(output_dir, "reconstructions")
if not os.path.exists(self.recon_dir):
os.mkdir(self.recon_dir)
# output directory for traversals
self.trav_dir = os.path.join(output_dir, "traversals")
if not os.path.exists(self.trav_dir):
os.mkdir(self.trav_dir)
def traverse(self,
decoder,
latent_vector,
dims=None,
num_traversals=None,
iter_n=""):
""" Traverses a latent vector along a given dimension(s).
Args:
decoder: (torch.nn.Module) decoder model that generates
the reconstructions from a latent vector
latent_vector: (torch.tensor) latent vector representation to be traversed
of shape (z_dim)
dims: (list, range or torch.tensor) list of dimensions to traverse in the latent vector
(optional)
num_traversals: (int) how many reconstructions to generate for each dimension.
The image grid will be of shape: len(dims) x num_traversals
iter_n: (str) iteration at which plotting and/or image index (OPTIONAL)
"""
if dims is None:
dims = torch.arange(latent_vector.size(0))
elif not (isinstance(dims, list) or isinstance(dims, range)
or isinstance(dims, torch.tensor)):
raise ValueError(
f"dims must either be a list or a torch.tensor, received {type(dims)}"
)
if num_traversals is None:
num_traversals = latent_vector.size(0)
elif not isinstance(num_traversals, int):
raise ValueError(
f"num_traversals must either be an int, received {type(num_traversals)}"
)
traversals = torch.linspace(-3., 3., steps=num_traversals).to(
latent_vector.device)
reconstructions = []
for dim in dims:
tiles = latent_vector.repeat(num_traversals, 1)
tiles[:, dim] = traversals
dim_recon = decoder(tiles)
reconstructions.append(dim_recon)
reconstructions = torch.sigmoid(torch.cat(reconstructions, dim=0))
reconstructed = torchvision.utils.make_grid(reconstructions,
normalize=True,
nrow=len(dims))
self.visdom.images(reconstructed.cpu(),
env=self.name + "-traversals",
opts={"title": iter_n},
nrow=len(dims))
if self.save:
torchvision.utils.save_image(reconstructions,
os.path.join(
self.trav_dir,
f"traversals-{iter_n}.png"),
normalize=True,
nrow=len(dims))
def show_reconstructions(self, images, reconstructions, iter_n=""):
""" Plots the ELBO loss, reconstruction loss and KL divergence
Args:
images: (torch.tensor) of shape batch_size x 3 x size x size
reconstructions: (torch.Tensor) of shape batch_size x 3 x size x size
iter_n: (str) iteration at which plotting (OPTIONAL)
"""
original = torchvision.utils.make_grid(images, normalize=True)
reconstructed = torchvision.utils.make_grid(
torch.sigmoid(reconstructions), normalize=True)
self.visdom.images(torch.stack([original, reconstructed], dim=0).cpu(),
env=self.name + "-reconstructed",
opts={"title": iter_n},
nrow=8)
if self.save:
torchvision.utils.save_image(original,
os.path.join(
self.recon_dir,
f"original-{iter_n}.png"),
normalize=True)
torchvision.utils.save_image(torch.sigmoid(reconstructions),
os.path.join(
self.recon_dir,
f"reconstructed-{iter_n}.png"),
normalize=True)
def __init_plots(self, iter_n, elbo, reconstruction_loss, kl_loss):
self.plots["elbo"] = self.visdom.line(torch.tensor([elbo]),
X=torch.tensor([iter_n]),
env=self.name + "-stats",
opts={
"title": "ELBO",
"width": 600,
"height": 500
})
self.plots["reconstruction_loss"] = self.visdom.line(
torch.tensor([reconstruction_loss]),
X=torch.tensor([iter_n]),
env=self.name + "-stats",
opts={
"title": "Reconstruction loss",
"width": 600,
"height": 500
})
self.plots["kl_loss"] = self.visdom.line(torch.tensor([kl_loss]),
X=torch.tensor([iter_n]),
env=self.name + "-stats",
opts={
"title": "KL divergence",
"width": 600,
"height": 500
})
def plot_stats(self, iter_n, elbo, reconstruction_loss, kl_loss):
""" Plots the ELBO loss, reconstruction loss and KL divergence
Args:
iter_n: (int) iteration at which plotting
elbo: (int)
reconstruction_loss: (int)
kl_loss: (int)
"""
# Initialize the plots
if not self.plots:
self.__init_plots(iter_n, elbo, reconstruction_loss, kl_loss)
return
self.plots["elbo"] = self.visdom.line(torch.tensor([elbo]),
X=torch.tensor([iter_n]),
win=self.plots["elbo"],
update="append",
env=self.name + "-stats",
opts={
"title": "ELBO",
"width": 600,
"height": 500
})
self.plots["reconstruction_loss"] = self.visdom.line(
torch.tensor([reconstruction_loss]),
X=torch.tensor([iter_n]),
win=self.plots["reconstruction_loss"],
update="append",
env=self.name + "-stats",
opts={
"title": "Reconstruction Loss",
"width": 600,
"height": 500
})
self.plots["kl_loss"] = self.visdom.line(torch.tensor([kl_loss]),
X=torch.tensor([iter_n]),
win=self.plots["kl_loss"],
update="append",
env=self.name + "-stats",
opts={
"title": "KL Divergence",
"width": 600,
"height": 500
})
def plot_means(self, z):
""" Plots dimension-wise boxplot distribution
Args:
z: (torch.tensor) single batch of latent vector representations
"""
if not self.plots.get("latent", False):
self.plots["latent"] = self.visdom.boxplot(
X=z,
env=self.name + "-stats",
opts={
"title": "Latent stats",
"width": 1200,
"height": 600,
"legend": [f"z_{i}" for i in range(1,
z.size(1) + 1)]
})
else:
self.plots["latent"] = self.visdom.boxplot(
X=z,
win=self.plots["latent"],
env=self.name + "-stats",
opts={
"title": "Latent stats",
"width": 1200,
"height": 600,
"legend": [f"z_{i}" for i in range(1,
z.size(1) + 1)]
})
def __init__(self, env='main'):
self.viz = Visdom()
self.env = env
self.meter = AverageMeter()
self.windows = defaultdict(lambda: None)
# LICENSE file in the root directory of this source tree.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from visdom import Visdom
import numpy as np
import math
import os.path
import getpass
from sys import platform as _platform
from six.moves import urllib
viz = Visdom()
assert viz.check_connection()
textwindow = viz.text('Hello World!')
updatetextwindow = viz.text('Hello World! More text should be here')
assert updatetextwindow is not None, 'Window was none'
viz.text('And here it is', win=updatetextwindow, append=True)
# video demo:
try:
video = np.empty([256, 250, 250, 3], dtype=np.uint8)
for n in range(256):
video[n, :, :, :].fill(n)
viz.video(tensor=video)
class visdomer(object):
def __init__(self, port, hostname, model_name, compare_phase,
init_elements, init_params):
self.port = port
self.hostname = hostname
self.env_name = model_name
self.viz = Visdom(server=hostname, port=port, env=model_name)
assert self.viz.check_connection(
timeout_seconds=3), 'No connection could be formed quickly'
self.compare_phase = compare_phase
self.init_params = init_params
self.windows = {}
for element in init_elements:
self.windows[element] = self.visdom_init(element)
def get_viz(self):
return self.viz
def get_window(self, metric):
# metric :: loss, accuracy,
return self.windows[metric]
def visdom_init(self, element):
if element == 'lr':
init_lr = self.init_params.get('lr')
assert init_lr is not None, 'init_lr is None when init visom of learning rate'
window = self.viz.line(X=torch.ones(1),
Y=torch.tensor([init_lr]),
opts=dict(title='{}'.format(element),
showlegend=True,
legend=['{}'.format(element)],
xtype='linear',
xlabel='epoch',
xtickmin=0,
xtick=True,
xtickstep=10,
ytype='linear',
ytickmin=0,
ylabel='{}'.format(element),
ytick=True))
return window
elif element in ['loss', 'accuracy', 'miou', 'f1score']:
assert isinstance(self.compare_phase, list) and len(
self.compare_phase
) == 2, 'compare_phase must be list and length with 2'
window = self.viz.line(
X=torch.stack((torch.ones(1), torch.ones(1)), 1),
Y=torch.stack((torch.ones(1), torch.ones(1)), 1),
opts=dict(title='{}-{}-{}'.format(self.compare_phase[0],
self.compare_phase[1],
element),
showlegend=True,
legend=[
'{}-{}'.format(self.compare_phase[0], element),
'{}-{}'.format(self.compare_phase[1], element)
],
xtype='linear',
label='epoch',
xtickmin=0,
xtick=True,
xtickstep=10,
ytype='linear',
ylabel='{}'.format(element),
ytickmin=0,
ytick=True))
elif element in ['cell_entropy', 'network_entropy', 'entropy']:
init_entropy = self.init_params['{:}'.format(element)]
window = self.viz.line(X=torch.ones(1),
Y=torch.tensor([init_entropy]),
opts=dict(title='{}'.format(element),
showlegend=True,
legend=['{}'.format(element)],
xtype='linear',
xlabel='epoch',
xtickmin=0,
xtick=True,
xtickstep=10,
ytype='linear',
ytickmin=0,
ylabel='{}'.format(element),
ytick=True))
elif element in ['warmup_loss', 'warmup_miou']:
window = self.viz.line(X=torch.ones(1),
Y=torch.ones(1),
opts=dict(title='{}'.format(element),
showlegend=True,
legend=['{}'.format(element)],
xtype='linear',
xlabel='epoch',
xtickmin=0,
xtick=True,
xtickstep=10,
ytype='linear',
ytickmin=0,
ylabel='{}'.format(element),
ytick=True))
return window
else:
raise NotImplementedError(
'do not support metric {}'.format(element))
return window
def visdom_update(self, epoch, update_element, update_value):
if update_element in ['loss', 'accuracy', 'miou', 'f1score']:
#print(update_value)
assert isinstance(update_value, list) and len(
update_value
) == 2, 'update_value should be list and with length 2, but got {:} with length {:}'.format(
type(update_value), len(update_value))
train_log = update_value[0]
valid_log = update_value[1]
window = self.get_window(update_element)
self.viz.line(
X=torch.stack((torch.ones(1) * epoch, torch.ones(1) * epoch),
1),
Y=torch.stack(
(torch.tensor([train_log]), torch.tensor([valid_log])), 1),
win=window,
update='append' if epoch != 1 else 'insert')
elif update_element == 'lr':
current_lr = update_value[0]
window = self.get_window(update_element)
self.viz.line(X=torch.ones(1) * epoch,
Y=torch.tensor([current_lr]),
win=window,
update='append' if epoch != 1 else 'insert')
elif update_element in ['cell_entropy', 'network_entropy', 'entropy']:
current_entropy = update_value[0]
window = self.get_window(update_element)
self.viz.line(X=torch.ones(1) * epoch,
Y=torch.tensor([current_entropy]),
win=window,
update='append' if epoch != 1 else 'insert')
elif update_element in ['warmup_loss', 'warmup_miou']:
current_value = update_value[0]
window = self.get_window(update_element)
self.viz.line(X=torch.ones(1) * epoch,
Y=torch.tensor([current_value]),
win=window,
update='append' if epoch != 1 else 'insert')
class VisdomLogger(Logger):
"""
Logger that uses visdom to create learning curves
Parameters:
===========
- env: str, name of the visdom environment
- log_checkpoints: bool, whether to use checkpoints or epoch averages
for training loss
- legend: tuple, names of the different losses that will be plotted.
"""
def __init__(self,
env=None,
log_checkpoints=True,
losses=('loss', ),
phases=('train', 'valid'),
server='http://localhost',
port=8097,
max_y=None,
**opts):
self.viz = None
if Visdom is not None:
self.viz = Visdom(server=server, port=port, env=env)
self.legend = ['%s.%s' % (p, l) for p in phases for l in losses]
opts.update({'legend': self.legend})
self.opts = opts
self.env = env
self.max_y = max_y
self.log_checkpoints = log_checkpoints
self.losses = set(losses)
self.last = {p: {l: None for l in losses} for p in phases}
self.pane = self._init_pane()
@skip_on_import_error(Visdom)
def _init_pane(self):
nan = np.array([np.NAN, np.NAN])
X = np.column_stack([nan] * len(self.legend))
Y = np.column_stack([nan] * len(self.legend))
return self.viz.line(
X=X, Y=Y, env=self.env, opts=self.opts)
def _update_last(self, epoch, loss, phase, loss_label):
self.last[phase][loss_label] = {'X': epoch, 'Y': loss}
def _plot_line(self, X, Y, phase, loss_label):
name = "%s.%s" % (phase, loss_label)
X = np.array([self.last[phase][loss_label]['X'], X])
Y = np.array([self.last[phase][loss_label]['Y'], Y])
if self.max_y:
Y = np.clip(Y, Y.min(), self.max_y)
self.viz.updateTrace(
X=X, Y=Y, name=name, append=True, win=self.pane, env=self.env)
def _plot_payload(self, epoch, losses, phase):
for label, loss in losses.items():
if label not in self.losses:
continue
if self.last[phase][label] is not None:
self._plot_line(epoch, loss, phase=phase, loss_label=label)
self._update_last(epoch, loss, phase, label)
@skip_on_import_error(Visdom)
def epoch_end(self, payload):
if self.log_checkpoints:
# only use epoch end if checkpoint isn't being used
return
losses, epoch = payload['loss'], payload['epoch'] + 1
self._plot_payload(epoch, losses, 'train')
@skip_on_import_error(Visdom)
def validation_end(self, payload):
losses, epoch = payload['loss'], payload['epoch'] + 1
self._plot_payload(epoch, losses, 'valid')
@skip_on_import_error(Visdom)
def checkpoint(self, payload):
epoch = payload['epoch'] + payload["batch"] / payload["total_batches"]
losses = payload['loss']
self._plot_payload(epoch, losses, 'train')
@skip_on_import_error(Visdom)
def attention(self, payload):
title = "epoch {epoch}/ batch {batch_num}".format(**payload)
if 'title' in self.opts:
title = self.opts['title'] + ": " + title
self.viz.heatmap(
X=np.array(payload["att"]),
env=self.env,
opts={'rownames': payload["hyp"],
'columnnames': payload["target"],
'title': title})
config['weight_decay'] = 1e-4
config['start_epoch'] = 0
config['batch_size'] = 26
config['valid_batch_size'] = 4
config['save_freq'] = 100
config['sigma'] = 5.
config['root_dir'] = r"./tcdata/round2train"
config['label_dir'] = r"./tcdata/round2train_checked.json"
config['valid_root_dir'] = r"./tcdata/round2_valid53/valid"
config['valid_label_dir'] = r"./tcdata/round2_valid53/round2_valid53_annotation51.json"
if config['showFlag'] == 1:
#生成一个viz的环境
viz = Visdom()
# python -m visdom.server
#初始化两个小的窗格，来分别绘制train,test的情况
# 绘制初始点，原点
viz.line([0.], [0.], win='train_loss', opts=dict(title='train loss')) #single-line
viz.line([0.], [0.], win='coor_loss', opts=dict(title='coor loss')) #single-line
viz.line([0.], [0.], win='valid_train_loss', opts=dict(title='valid_train loss')) #single-line
viz.line([0.], [0.], win='acc_coor', opts=dict(title='average acc_coor')) #single-line
# 数据预处理设置
# normMean = [0.16783774]
# normStd = [0.18892017]
# normMean = [0.168036]
# normStd = [0.177935]
normMean = [0.168036]
normStd = [0.177935]
def main():
args = parser.parse_args()
cf = ConfigParser.ConfigParser()
try:
cf.read(args.conf)
except:
print("conf file not exists")
sys.exit(1)
USE_CUDA = cf.getboolean('Training', 'use_cuda')
try:
seed = long(cf.get('Training', 'seed'))
except:
seed = torch.cuda.initial_seed()
cf.set('Training', 'seed', seed)
cf.write(open(args.conf, 'w'))
torch.manual_seed(seed)
if USE_CUDA:
torch.cuda.manual_seed(seed)
log_dir = cf.get('Data', 'log_dir')
log_file = os.path.join(log_dir, cf.get('Data', 'log_file'))
logger = init_logger(log_file)
#Define Model
rnn_input_size = cf.getint('Model', 'rnn_input_size')
rnn_hidden_size = cf.getint('Model', 'rnn_hidden_size')
rnn_layers = cf.getint('Model', 'rnn_layers')
rnn_type = RNN[cf.get('Model', 'rnn_type')]
bidirectional = cf.getboolean('Model', 'bidirectional')
batch_norm = cf.getboolean('Model', 'batch_norm')
rnn_param = {"rnn_input_size":rnn_input_size, "rnn_hidden_size":rnn_hidden_size, "rnn_layers":rnn_layers,
"rnn_type":rnn_type, "bidirectional":bidirectional, "batch_norm":batch_norm}
num_class = cf.getint('Model', 'num_class')
drop_out = cf.getfloat('Model', 'drop_out')
model = CTC_Model(rnn_param=rnn_param, num_class=num_class, drop_out=drop_out)
print("Model Structure:")
logger.info("Model Structure:")
for idx, m in enumerate(model.children()):
print(idx, m)
logger.info(str(idx) + "->" + str(m))
data_dir = cf.get('Data', 'data_dir')
batch_size = cf.getint("Training", 'batch_size')
#Data Loader
train_dataset = SpeechDataset(data_dir, data_set='train')
dev_dataset = SpeechDataset(data_dir, data_set="dev")
train_loader = SpeechDataLoader(train_dataset, batch_size=batch_size, shuffle=True,
num_workers=4, pin_memory=False)
dev_loader = SpeechDataLoader(dev_dataset, batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=False)
#ensure the feats is equal to the rnn_input_Size
assert train_dataset.n_feats == rnn_input_size
#decoder for dev set
decoder = GreedyDecoder(int2char, space_idx=len(int2char) - 1, blank_index=0)
#Training
init_lr = cf.getfloat('Training', 'init_lr')
num_epoches = cf.getint('Training', 'num_epoches')
end_adjust_acc = cf.getfloat('Training', 'end_adjust_acc')
decay = cf.getfloat("Training", 'lr_decay')
weight_decay = cf.getfloat("Training", 'weight_decay')
params = { 'num_epoches':num_epoches, 'end_adjust_acc':end_adjust_acc, 'seed':seed,
'decay':decay, 'learning_rate':init_lr, 'weight_decay':weight_decay, 'batch_size':batch_size, 'n_feats':train_dataset.n_feats }
print(params)
if USE_CUDA:
model = model.cuda()
loss_fn = CTCLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=init_lr, weight_decay=weight_decay)
#visualization for training
from visdom import Visdom
viz = Visdom()
title = 'TIMIT LSTM_CTC Acoustic Model'
opts = [dict(title=title+" Loss", ylabel = 'Loss', xlabel = 'Epoch'),
dict(title=title+" Loss on Dev", ylabel = 'DEV Loss', xlabel = 'Epoch'),
dict(title=title+' CER on DEV', ylabel = 'DEV CER', xlabel = 'Epoch')]
viz_window = [None, None, None]
count = 0
learning_rate = init_lr
loss_best = 1000
loss_best_true = 1000
adjust_rate_flag = False
stop_train = False
adjust_time = 0
acc_best = 0
start_time = time.time()
loss_results = []
dev_loss_results = []
dev_cer_results = []
while not stop_train:
if count >= num_epoches:
break
count += 1
if adjust_rate_flag:
learning_rate *= decay
adjust_rate_flag = False
for param in optimizer.param_groups:
param['lr'] *= decay
print("Start training epoch: %d, learning_rate: %.5f" % (count, learning_rate))
logger.info("Start training epoch: %d, learning_rate: %.5f" % (count, learning_rate))
loss = train(model, train_loader, loss_fn, optimizer, logger, print_every=20, USE_CUDA=USE_CUDA)
loss_results.append(loss)
acc, dev_loss = dev(model, dev_loader, loss_fn, decoder, logger, USE_CUDA=USE_CUDA)
print("loss on dev set is %.4f" % dev_loss)
logger.info("loss on dev set is %.4f" % dev_loss)
dev_loss_results.append(dev_loss)
dev_cer_results.append(acc)
#adjust learning rate by dev_loss
#adjust_rate_count : 表示连续超过count个epoch的loss在end_adjust_acc区间内认为稳定
if dev_loss < (loss_best - end_adjust_acc):
loss_best = dev_loss
loss_best_true = dev_loss
adjust_rate_count = 0
acc_best = acc
best_model_state = copy.deepcopy(model.state_dict())
best_op_state = copy.deepcopy(optimizer.state_dict())
elif (dev_loss < loss_best + end_adjust_acc):
adjust_rate_count += 1
if dev_loss < loss_best and dev_loss < loss_best_true:
loss_best_true = dev_loss
acc_best = acc
best_model_state = copy.deepcopy(model.state_dict())
best_op_state = copy.deepcopy(optimizer.state_dict())
else:
adjust_rate_count = 10
print("adjust_rate_count: %d" % adjust_rate_count)
print('adjust_time: %d' % adjust_time)
logger.info("adjust_rate_count: %d" % adjust_rate_count)
logger.info('adjust_time: %d' % adjust_time)
if adjust_rate_count == 10:
adjust_rate_flag = True
adjust_time += 1
adjust_rate_count = 0
if loss_best > loss_best_true:
loss_best = loss_best_true
model.load_state_dict(best_model_state)
optimizer.load_state_dict(best_op_state)
if adjust_time == 8:
stop_train = True
time_used = (time.time() - start_time) / 60
print("epoch %d done, dev acc is: %.4f, time_used: %.4f minutes" % (count, acc, time_used))
logger.info("epoch %d done, dev acc is: %.4f, time_used: %.4f minutes" % (count, acc, time_used))
x_axis = range(count)
y_axis = [loss_results[0:count], dev_loss_results[0:count], dev_cer_results[0:count]]
for x in range(len(viz_window)):
if viz_window[x] is None:
viz_window[x] = viz.line(X = np.array(x_axis), Y = np.array(y_axis[x]), opts = opts[x],)
else:
viz.line(X = np.array(x_axis), Y = np.array(y_axis[x]), win = viz_window[x], update = 'replace',)
print("End training, best dev loss is: %.4f, acc is: %.4f" % (loss_best_true, acc_best))
logger.info("End training, best dev loss acc is: %.4f, acc is: %.4f" % (loss_best_true, acc_best))
model.load_state_dict(best_model_state)
optimizer.load_state_dict(best_op_state)
best_path = os.path.join(log_dir, 'best_model'+'_dev'+str(acc_best)+'.pkl')
cf.set('Model', 'model_file', best_path)
cf.write(open(args.conf, 'w'))
params['epoch']=count
torch.save(CTC_Model.save_package(model, optimizer=optimizer, epoch=params, loss_results=loss_results, dev_loss_results=dev_loss_results, dev_cer_results=dev_cer_results), best_path)
class Visualizer(object):
"""
封装visdom的基本操作
"""
def __init__(self, env='default', **kwargs):
self.vis = Visdom(env=env,use_incoming_socket=False, **kwargs)
# 以文字和图的形式展示
# {'loss1':23,'loss2':24}
self.index = {}
self.log_text = ''
self.log_index = {} # {win:[log_text]}
def reinit(self, env='default', **kwargs):
self.vis = Visdom(env=env, **kwargs)
def img(self, name ,img_, **kwargs):
# img_: batchsize*channels*H*W 0-1 cuda tensor
self.vis.images(img_.cpu().numpy(),
win=name,
opts=dict(title=name),
**kwargs
)
def log(self, info='', win='log_text', append=False ):
"""
self.log({'loss':1,'lr':0.0001})
:param info:
:param win:
:return:
"""
log_text = self.log_index.get(win,'')
if append:
log_text +=(
'[{time}]{info}<br>'.format(
time = time.strftime('%m%d_%H:%M:%S'),
info = info
)
)
else:
log_text =(
'[{time}]{info}<br>'.format(
time = time.strftime('%m%d_%H:%M:%S'),
info = info
)
)
self.vis.text(log_text, win=win, opts= dict(title=win))
self.log_index[win] = log_text
def plot(self, win, y, **kwargs):
"""
plot('loss',2)
:param win:
:param y:
:param kwargs:
:return:
"""
x = self.index.get(win,0)
self.vis.line(
X=np.array([x]), Y=np.array([y]),
win = win,
opts= dict(title=win),
update = None if x ==0 else 'append',
**kwargs
)
self.index[win] = x+1
def img_many(self,d):
# d: {'1.jpg':b*c*H*W,'2.jpg':b*c*H*W}
for k,v in d.items():
self.img(k,v)
def plot_many(self,d):
# d:{'loss1':2,'loass2':4}
for k,v in d.items():
self.plot(k,v)
def __getattr__(self,name):
# self,function->self.vis.funtion
return getattr(self.vis, name)
def main():
args = parser.parse_args()
cf = ConfigParser.ConfigParser()
try:
cf.read(args.conf)
except:
print("conf file not exists")
sys.exit(1)
try:
seed = cf.get('Training', 'seed')
seed = long(seed)
except:
seed = torch.cuda.initial_seed()
cf.set('Training', 'seed', seed)
cf.write(open(args.conf, 'w'))
USE_CUDA = cf.getboolean("Training", "use_cuda")
torch.manual_seed(seed)
if USE_CUDA:
torch.cuda.manual_seed(seed)
logger = init_logger(os.path.join(args.log_dir, 'train_ctc_model.log'))
#Define Model
rnn_input_size = cf.getint('Model', 'rnn_input_size')
rnn_hidden_size = cf.getint('Model', 'rnn_hidden_size')
rnn_layers = cf.getint('Model', 'rnn_layers')
rnn_type = supported_rnn[cf.get('Model', 'rnn_type')]
bidirectional = cf.getboolean('Model', 'bidirectional')
batch_norm = cf.getboolean('Model', 'batch_norm')
rnn_param = {"rnn_input_size":rnn_input_size, "rnn_hidden_size":rnn_hidden_size, "rnn_layers":rnn_layers,
"rnn_type":rnn_type, "bidirectional":bidirectional, "batch_norm":batch_norm}
num_class = cf.getint('Model', 'num_class')
drop_out = cf.getfloat('Model', 'drop_out')
add_cnn = cf.getboolean('Model', 'add_cnn')
cnn_param = {}
layers = cf.getint('CNN', 'layers')
channel = eval(cf.get('CNN', 'channel'))
kernel_size = eval(cf.get('CNN', 'kernel_size'))
stride = eval(cf.get('CNN', 'stride'))
padding = eval(cf.get('CNN', 'padding'))
pooling = eval(cf.get('CNN', 'pooling'))
batch_norm = cf.getboolean('CNN', 'batch_norm')
activation_function = supported_activate[cf.get('CNN', 'activation_function')]
cnn_param['batch_norm'] = batch_norm
cnn_param['activate_function'] = activation_function
cnn_param["layer"] = []
for layer in range(layers):
layer_param = [channel[layer], kernel_size[layer], stride[layer], padding[layer]]
if pooling is not None:
layer_param.append(pooling[layer])
else:
layer_param.append(None)
cnn_param["layer"].append(layer_param)
model = CTC_Model(rnn_param=rnn_param, add_cnn=add_cnn, cnn_param=cnn_param, num_class=num_class, drop_out=drop_out)
for idx, m in enumerate(model.children()):
print(idx, m)
logger.info(str(idx) + "->" + str(m))
dataset = cf.get('Data', 'dataset')
data_dir = cf.get('Data', 'data_dir')
feature_type = cf.get('Data', 'feature_type')
out_type = cf.get('Data', 'out_type')
n_feats = cf.getint('Data', 'n_feats')
mel = cf.getboolean('Data', 'mel')
batch_size = cf.getint("Training", 'batch_size')
#Data Loader
train_dataset = SpeechDataset(data_dir, data_set='train', feature_type=feature_type, out_type=out_type, n_feats=n_feats, mel=mel)
dev_dataset = SpeechDataset(data_dir, data_set="dev", feature_type=feature_type, out_type=out_type, n_feats=n_feats, mel=mel)
if add_cnn:
train_loader = SpeechCNNDataLoader(train_dataset, batch_size=batch_size, shuffle=True,
num_workers=4, pin_memory=False)
dev_loader = SpeechCNNDataLoader(dev_dataset, batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=False)
else:
train_loader = SpeechDataLoader(train_dataset, batch_size=batch_size, shuffle=True,
num_workers=4, pin_memory=False)
dev_loader = SpeechDataLoader(dev_dataset, batch_size=batch_size, shuffle=False,
num_workers=4, pin_memory=False)
#decoder for dev set
decoder = GreedyDecoder(dev_dataset.int2class, space_idx=-1, blank_index=0)
#Training
init_lr = cf.getfloat('Training', 'init_lr')
num_epoches = cf.getint('Training', 'num_epoches')
end_adjust_acc = cf.getfloat('Training', 'end_adjust_acc')
decay = cf.getfloat("Training", 'lr_decay')
weight_decay = cf.getfloat("Training", 'weight_decay')
params = { 'num_epoches':num_epoches, 'end_adjust_acc':end_adjust_acc, 'mel': mel, 'seed':seed,
'decay':decay, 'learning_rate':init_lr, 'weight_decay':weight_decay, 'batch_size':batch_size,
'feature_type':feature_type, 'n_feats': n_feats, 'out_type': out_type }
print(params)
if USE_CUDA:
model = model.cuda()
loss_fn = CTCLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=init_lr, weight_decay=weight_decay)
#visualization for training
from visdom import Visdom
viz = Visdom()
if add_cnn:
title = dataset+' '+feature_type+str(n_feats)+' CNN_LSTM_CTC'
else:
title = dataset+' '+feature_type+str(n_feats)+' LSTM_CTC'
opts = [dict(title=title+" Loss", ylabel = 'Loss', xlabel = 'Epoch'),
dict(title=title+" Loss on Dev", ylabel = 'DEV Loss', xlabel = 'Epoch'),
dict(title=title+' CER on DEV', ylabel = 'DEV CER', xlabel = 'Epoch')]
viz_window = [None, None, None]
count = 0
learning_rate = init_lr
loss_best = 1000
loss_best_true = 1000
adjust_rate_flag = False
stop_train = False
adjust_time = 0
acc_best = 0
start_time = time.time()
loss_results = []
dev_loss_results = []
dev_cer_results = []
while not stop_train:
if count >= num_epoches:
break
count += 1
if adjust_rate_flag:
learning_rate *= decay
adjust_rate_flag = False
for param in optimizer.param_groups:
param['lr'] *= decay
print("Start training epoch: %d, learning_rate: %.5f" % (count, learning_rate))
logger.info("Start training epoch: %d, learning_rate: %.5f" % (count, learning_rate))
loss = train(model, train_loader, loss_fn, optimizer, logger, add_cnn=add_cnn, print_every=20, USE_CUDA=USE_CUDA)
loss_results.append(loss)
acc, dev_loss = dev(model, dev_loader, loss_fn, decoder, logger, add_cnn=add_cnn, USE_CUDA=USE_CUDA)
print("loss on dev set is %.4f" % dev_loss)
logger.info("loss on dev set is %.4f" % dev_loss)
dev_loss_results.append(dev_loss)
dev_cer_results.append(acc)
#adjust learning rate by dev_loss
if dev_loss < (loss_best - end_adjust_acc):
loss_best = dev_loss
loss_best_true = dev_loss
#acc_best = acc
adjust_rate_count = 0
model_state = copy.deepcopy(model.state_dict())
op_state = copy.deepcopy(optimizer.state_dict())
elif (dev_loss < loss_best + end_adjust_acc):
adjust_rate_count += 1
if dev_loss < loss_best and dev_loss < loss_best_true:
loss_best_true = dev_loss
#acc_best = acc
model_state = copy.deepcopy(model.state_dict())
op_state = copy.deepcopy(optimizer.state_dict())
else:
adjust_rate_count = 10
if acc > acc_best:
acc_best = acc
best_model_state = copy.deepcopy(model.state_dict())
best_op_state = copy.deepcopy(optimizer.state_dict())
print("adjust_rate_count:"+str(adjust_rate_count))
print('adjust_time:'+str(adjust_time))
logger.info("adjust_rate_count:"+str(adjust_rate_count))
logger.info('adjust_time:'+str(adjust_time))
if adjust_rate_count == 10:
adjust_rate_flag = True
adjust_time += 1
adjust_rate_count = 0
if loss_best > loss_best_true:
loss_best = loss_best_true
model.load_state_dict(model_state)
optimizer.load_state_dict(op_state)
if adjust_time == 8:
stop_train = True
time_used = (time.time() - start_time) / 60
print("epoch %d done, cv acc is: %.4f, time_used: %.4f minutes" % (count, acc, time_used))
logger.info("epoch %d done, cv acc is: %.4f, time_used: %.4f minutes" % (count, acc, time_used))
x_axis = range(count)
y_axis = [loss_results[0:count], dev_loss_results[0:count], dev_cer_results[0:count]]
for x in range(len(viz_window)):
if viz_window[x] is None:
viz_window[x] = viz.line(X = np.array(x_axis), Y = np.array(y_axis[x]), opts = opts[x],)
else:
viz.line(X = np.array(x_axis), Y = np.array(y_axis[x]), win = viz_window[x], update = 'replace',)
print("End training, best dev loss is: %.4f, acc is: %.4f" % (loss_best, acc_best))
logger.info("End training, best dev loss acc is: %.4f, acc is: %.4f" % (loss_best, acc_best))
model.load_state_dict(best_model_state)
optimizer.load_state_dict(best_op_state)
best_path = os.path.join(args.log_dir, 'best_model'+'_dev'+str(acc_best)+'.pkl')
cf.set('Model', 'model_file', best_path)
cf.write(open(args.conf, 'w'))
params['epoch']=count
torch.save(CTC_Model.save_package(model, optimizer=optimizer, epoch=params, loss_results=loss_results, dev_loss_results=dev_loss_results, dev_cer_results=dev_cer_results), best_path)
def reinit(self, env='default', **kwargs):
self.vis = Visdom(env=env, **kwargs)
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
from visdom import Visdom
import numpy as np
import math
import os.path
import getpass
viz = Visdom()
textwindow = viz.text('Hello World!')
# video demo:
try:
video = np.empty([256, 250, 250, 3], dtype=np.uint8)
for n in range(256):
video[n, :, :, :].fill(n)
viz.video(tensor=video)
# video demo: download video from http://media.w3.org/2010/05/sintel/trailer.ogv
videofile = '/home/%s/trailer.ogv' % getpass.getuser()
if os.path.isfile(videofile):
viz.video(videofile=videofile)
except ImportError:
import gym
import numpy as np
import matplotlib.pyplot as plt
from visdom import Visdom
viz = Visdom(env="mean_reward")
env = gym.make("Taxi-v2")
state_space = 500
action_space = 6
epsilon = 0.1
num_episode = 30000
q = np.zeros((state_space, action_space))
q1 = np.zeros((state_space, action_space))
q2 = np.zeros((state_space, action_space))
r_normal = []
r_double = []
def choose_action(state, epsilon, flag=True):
if flag:
if np.random.rand() > epsilon:
action = np.argmax(q[state])
else:
action = np.random.randint(action_space)
return action
else:
if np.random.rand() > epsilon:
action = np.argmax((q1[state] + q2[state]) / 2)
else:
action = np.random.randint(action_space)
return action
from visdom import Visdom
import numpy as np
import math
import os.path
import getpass
from sys import platform as _platform
from six.moves import urllib
import re
viz = Visdom()
f=open('log_purerl_1to0_5class11').read()
print f[:100]
# ttt0,ttt1,ttt2,ttt3,ttt4,ttt5=[],[],[],[],[],[]
ttt=[]
ttt.append([float(i) for i in re.findall('ttt0 is.* (.*)',f)])
ttt.append([float(i) for i in re.findall('acc:(.*?)\t',f)])
ttt.append([float(i) for i in re.findall('ttt2 is.* (.*)',f)])
ttt.append([float(i) for i in re.findall('ttt3 is.* (.*)',f)])
ttt.append([float(i) for i in re.findall('ttt4 is.* (.*)',f)])
ttt.append([float(i) for i in re.findall('ttt5 is.* (.*)',f)])
times=range(0,306,3)
# times=range(0,60,3)
# tt=[]
# for i in ttt:
# tt.append(i[:20])
# ttt=tt
# textwindow = viz.text('Hello World!')
win = viz.line(
import numpy as np import torch import torch.nn as nn import torch.optim as optim import torch.nn.init as init from torch.autograd import Variable # visdom is a visualization tool from facebook from visdom import Visdom viz = Visdom() num_data = 1000 num_epoch = 1000 noise = init.normal(torch.FloatTensor(num_data, 1), std=1) x = init.uniform(torch.Tensor(num_data, 1), -10, 10) y = 2 * x + 3 y_noise = 2 * x + 3 + noise # visualize data with visdom input_data = torch.cat([x, y_noise], 1) # win = viz.scatter( # X=input_data, # opts=dict( # xtickmin=-10, # xtickmax=10, # xtickstep=1, # ytickmin=-20, # ytickmax=20, # ytickstep=1,
class EvaluationMonitor(Wrapper):
def __init__(self, env, cmdl):
super(EvaluationMonitor, self).__init__(env)
self.freq = cmdl.eval_freq # in steps
self.eval_steps = cmdl.evaluator.eval_steps
self.cmdl = cmdl
if self.cmdl.display_plots:
self.vis = Visdom()
self.plot = self.vis.line(
Y=np.array([0]), X=np.array([0]),
opts=dict(
title=cmdl.label,
caption="Episodic reward per %d steps." % self.eval_steps)
)
self.crt_step = 0
self.step_cnt = 0
self.ep_cnt = 0
self.total_rw = 0
self.max_mean_rw = -100
def get_crt_step(self, crt_step):
self.crt_step = crt_step
def _reset_monitor(self):
self.step_cnt, self.ep_cnt, self.total_rw = 0, 0, 0
def _step(self, action):
# self._before_step(action)
observation, reward, done, info = self.env.step(action)
done = self._after_step(observation, reward, done, info)
return observation, reward, done, info
def _reset(self):
observation = self.env.reset()
self._after_reset(observation)
return observation
def _after_step(self, o, r, done, info):
self.total_rw += r
self.step_cnt += 1
if self.step_cnt == self.eval_steps:
self._update_plot()
self._reset_monitor()
return done
def _after_reset(self, observation):
self.ep_cnt += 1
# print("[%2d][%4d] RESET" % (self.ep_cnt, self.step_cnt))
def _update_plot(self):
mean_rw = self.total_rw / self.ep_cnt
max_mean_rw = self.max_mean_rw
bg_color = 'on_blue'
bg_color = 'on_magenta' if mean_rw > max_mean_rw else bg_color
self.max_mean_rw = mean_rw if mean_rw > max_mean_rw else max_mean_rw
if self.cmdl.display_plots:
self.vis.line(
X=np.array([self.crt_step]),
Y=np.array([mean_rw]),
win=self.plot,
update='append'
)
print(clr("[Evaluator] done in %5d steps. " % self.step_cnt,
'grey', 'on_white')
+ clr(" rw/ep=%3.2f " % mean_rw, 'white', bg_color))
args = parser.parse_args()
return args
def print_args(args):
print("##############################################")
for k, v in vars(args).items():
print(k + ': ' + str(v))
print("##############################################")
print()
# pip install visdom
# python -m visdom.server
# open web page --> http://localhost:8097
vis = Visdom()
episode_reward_plt = None
avg_episode_reward_plt = None
episode_loss_plt = None
epsilon_plt = None
def vis_plt(method, args):
global episode_reward_plt
global avg_episode_reward_plt
global episode_loss_plt
global epsilon_plt
layout_opts_dict = {
'plotly': {
GPU_ID = None
epochs = 100
batch_size = 32
start_epoch = 1
save_snapshot_interval_epoch = 1
peek_interval_epoch = 1
save_train_hr_interval_epoch = 1
loss_average_win_size = 2
validate_interval_epoch = 1
validate_batch_size = 4
plot_loss_start_epoch = 1
only_validate = False #
from visdom import Visdom
vis = Visdom(server='http://127.0.0.1', port=8097)
# =================== config for model and dataset =====================================================================
from squid.data import Photo2PhotoData
from squid.data import RandomCropPhoto2PhotoData
from squid.model import SuperviseModel
import torch
import torch.nn as nn
from squid.loss import VGGLoss
from squid.net import AOD_Deep1_Net
target_net = AOD_Deep1_Net()
target_net = nn.DataParallel(target_net).cuda()
model = SuperviseModel({
'net':
def __init__(self, env_name='main', imsize=None):
self.vis = Visdom()
self.env = env_name
self.images = {}
self.imsize = imsize
class Callback(object):
"""
Used to log/visualize the evaluation metrics during training. The values are stored at the end of each epoch.
"""
def __init__(self, metrics):
"""
Args:
metrics : a list of callbacks. Possible values:
"CoherenceMetric"
"PerplexityMetric"
"DiffMetric"
"ConvergenceMetric"
"""
# list of metrics to be plot
self.metrics = metrics
def set_model(self, model):
"""
Save the model instance and initialize any required variables which would be updated throughout training
"""
self.model = model
self.previous = None
# check for any metric which need model state from previous epoch
if any(isinstance(metric, (DiffMetric, ConvergenceMetric)) for metric in self.metrics):
self.previous = copy.deepcopy(model)
# store diff diagonals of previous epochs
self.diff_mat = Queue()
if any(metric.logger == "visdom" for metric in self.metrics):
if not VISDOM_INSTALLED:
raise ImportError("Please install Visdom for visualization")
self.viz = Visdom()
# store initial plot windows of every metric (same window will be updated with increasing epochs)
self.windows = []
if any(metric.logger == "shell" for metric in self.metrics):
# set logger for current topic model
self.log_type = logging.getLogger('gensim.models.ldamodel')
def on_epoch_end(self, epoch, topics=None):
"""
Log or visualize current epoch's metric value
Args:
epoch : current epoch no.
topics : topic distribution from current epoch (required for coherence of unsupported topic models)
"""
# stores current epoch's metric values
current_metrics = {}
# plot all metrics in current epoch
for i, metric in enumerate(self.metrics):
label = str(metric)
value = metric.get_value(topics=topics, model=self.model, other_model=self.previous)
current_metrics[label] = value
if metric.logger == "visdom":
if epoch == 0:
if value.ndim > 0:
diff_mat = np.array([value])
viz_metric = self.viz.heatmap(
X=diff_mat.T, env=metric.viz_env, opts=dict(xlabel='Epochs', ylabel=label, title=label)
)
# store current epoch's diff diagonal
self.diff_mat.put(diff_mat)
# saving initial plot window
self.windows.append(copy.deepcopy(viz_metric))
else:
viz_metric = self.viz.line(
Y=np.array([value]), X=np.array([epoch]), env=metric.viz_env,
opts=dict(xlabel='Epochs', ylabel=label, title=label)
)
# saving initial plot window
self.windows.append(copy.deepcopy(viz_metric))
else:
if value.ndim > 0:
# concatenate with previous epoch's diff diagonals
diff_mat = np.concatenate((self.diff_mat.get(), np.array([value])))
self.viz.heatmap(
X=diff_mat.T, env=metric.viz_env, win=self.windows[i],
opts=dict(xlabel='Epochs', ylabel=label, title=label)
)
self.diff_mat.put(diff_mat)
else:
self.viz.updateTrace(
Y=np.array([value]), X=np.array([epoch]), env=metric.viz_env, win=self.windows[i]
)
if metric.logger == "shell":
statement = "".join(("Epoch ", str(epoch), ": ", label, " estimate: ", str(value)))
self.log_type.info(statement)
# check for any metric which need model state from previous epoch
if isinstance(metric, (DiffMetric, ConvergenceMetric)):
self.previous = copy.deepcopy(self.model)
return current_metrics
from io import BytesIO
from PIL import Image
from argparse import ArgumentParser
from visdom import Visdom
visdom_endpoint = variables.get("ENDPOINT_VISDOM") if variables.get("ENDPOINT_VISDOM") else results[0].__str__()
print("VISDOM_ENDPOINT: ", visdom_endpoint)
assert visdom_endpoint is not None
visdom_endpoint = visdom_endpoint.replace("http://", "")
(VISDOM_HOST, VISDOM_PORT) = visdom_endpoint.split(":")
print("Connecting to %s:%s" % (VISDOM_HOST, VISDOM_PORT))
vis = Visdom(server="http://"+VISDOM_HOST, port=int(VISDOM_PORT))
assert vis.check_connection()
# text plot
textwindow = vis.text('Hello World!')
#updatetextwindow = vis.text('Hello World! More text should be here')
#vis.text('And here it is', win=updatetextwindow, append=True)
# show ActiveEon logo
url_image = 'http://s3.eu-west-2.amazonaws.com/activeeon-public/images/logo.jpg'
http = urllib3.PoolManager()
r = http.request('GET', url_image)
image = np.asarray(Image.open(BytesIO(r.data))).astype(np.uint8)
vis_image = image.transpose((2, 0, 1)).astype(np.float64)
vis.image(vis_image,opts=dict(title='ActiveEon', caption='ActiveEon'))