def plot_loss(l1, l2, ac1, ac2, name, fold):
"""
Plots train and validation losses and accuracies
:param l1: list of train losses
:param l2: list of validation losses
:param ac1: list of train accuracies
:param ac2: list of validation accuracies
:param name: train file name
:param fold: extension to identify plot images
"""
plt.clf()
viz = Visdom()
#os.system("python -m visdom.server &")
plt.plot(l1, 'k', label='train', linewidth=.5)
plt.plot(l2, 'r', label='val', linewidth=.5)
plt.plot(ac1, 'y', label='Train accuracy', linewidth=.5)
plt.plot(ac2, 'g', label='Val accuracy', linewidth=.5)
plt.title(name)
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.gca().spines['top'].set_visible(False)
plt.gca().spines['right'].set_visible(False)
plt.legend(loc=0)
viz.matplot(plt)
plt.savefig('plots/loss_fold_' + str(fold) + '.png')
def show_pe_chart(max_pe=150, pe_key='3'):
plt.style.use('seaborn-whitegrid')
plt.ylabel(u'市盈率', fontproperties='SimHei')
plt.xlabel(u'时间轴', fontproperties='SimHei')
plt.title(titles[pe_key], fontproperties='SimHei')
plt.ylim(-1, max_pe + 10)
plt.legend(loc=0, prop=font)
plt.grid(True)
viz = Visdom(env='main')
viz.matplot(plt)
def vis_show_image(image: np.array):
from matplotlib import pyplot as plt
import matplotlib
import cv2
from visdom import Visdom
matplotlib.use('Agg')
rgb_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
plt.imshow(rgb_image)
vis = Visdom()
vis.matplot(rgb_image)
def show_salary_chart(ylabel=''):
global ylabel_dict
plt.style.use('seaborn-whitegrid')
plt.xlabel(u'时间轴', fontproperties='SimHei')
plt.xticks(rotation=-90)
plt.title(ylabel, fontproperties='SimHei')
# plt.xlim(2000, 2020)
# plt.ylim(-1, max_pe+10)
plt.legend(loc=0, prop=font)
plt.grid(True)
viz = Visdom(env='main')
viz.matplot(plt)
class VisdomWebServer(object):
def __init__(self):
DEFAULT_PORT = 8097
DEFAULT_HOSTNAME = "http://localhost"
self.vis = Visdom(port=DEFAULT_PORT, server=DEFAULT_HOSTNAME)
def update(self, metrics):
if not self.vis.check_connection():
'No connection could be formed quickly'
return
# Learning curve
try:
fig, ax = plt.subplots()
plt.plot(metrics['train_loss'],
label='Training loss',
color='#32526e')
plt.plot(metrics['val_loss'],
label='Validation loss',
color='#ff6b57')
plt.legend()
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
plt.grid(zorder=0, color='lightgray', linestyle='--')
self.vis.matplot(plt, win='lrcurve')
plt.close()
plt.clf()
fig, ax = plt.subplots()
plt.plot(metrics['learning_rate'], color='#32526e')
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
plt.grid(zorder=0, color='lightgray', linestyle='--')
self.vis.matplot(plt, win='lr_rate')
plt.close()
plt.clf()
#plt.figure()
#plt.plot(metrics['zernike_train_loss'], label='Zernike train loss', color='blue')
#plt.plot(metrics['zernike_val_loss'], label='Zernike val loss', color='red')
#plt.legend()
#plt.grid()
#self.vis.matplot(plt, win='lrcurve_z')
#plt.close()
#plt.clf()
except BaseException as err:
print('Skipped matplotlib example')
print('Error message: ', err)
class VisdomLinePlotter(object):
def __init__(self, env_name='main'):
self.vis = Visdom()
self.env = env_name
self.plots = {}
def plot(self, var_name, split_name, title_name, x, y, xlabel='Epochs'):
if var_name not in self.plots:
self.plots[var_name] = self.vis.line(X=np.array([x, x]),
Y=np.array([y, y]),
env=self.env,
opts=dict(legend=[split_name],
title=title_name,
xlabel=xlabel,
ylabel=var_name))
else:
self.vis.line(X=np.array([x]),
Y=np.array([y]),
env=self.env,
win=self.plots[var_name],
name=split_name,
update='append')
def plot_matplotlib(self, plot_name, plt):
self.plots[plot_name] = self.vis.matplot(plt, env=self.env)
def plot_text(self, text, title='Text'):
self.vis.text(text, env=self.env, opts=dict(title=title))
def show_detect_result(imgs_arr=[], all_boxes=[]):
trans_func = transforms.ToPILImage()
imgs_one_line = int(len(imgs_arr) / 2 + (len(imgs_arr) % 2))
for idx, img in enumerate(imgs_arr):
# img_plt = trans_func(img).convert('RGB')
axes = plt.subplot(2, imgs_one_line, (idx + 1))
i = 0
objs = all_boxes[idx]
print('len(objs): ', len(objs))
for obj_class in objs:
# print('item[0]: ', int(item[0].item()))
for item in obj_class:
print('item.shape: ', item.shape)
name = labelmap[int(item[0])]
score = item[1]
xmin = int(item[2])
ymin = int(item[3])
xmax = int(item[4])
ymax = int(item[5] )
# print('name: ', name)
# print('(xmin, ymin, xmax, ymax): ', xmin, ymin, xmax, ymax)
i += 1
i %= len(color_list)
rect = patches.Rectangle((xmin, ymin), (xmax - xmin), (ymax - ymin),
linewidth=2, edgecolor=color_list[i], fill=False)
if args.score == True:
axes.text(rect.xy[0], rect.xy[1], str(score),
va='center', ha='center', color=color_list[i],
bbox=dict(facecolor='w'))
if args.bbox == True:
axes.add_patch(rect)
if args.label == True:
axes.text(rect.xy[0], rect.xy[1], name,
va='center', ha='center', color='k',
bbox=dict(facecolor='w'))
plt.imshow(img)
plt.axis('off')
plt.ioff()
viz = Visdom(env='ssd_obj_detect')
viz.matplot(plt)
class Display_board:
def __init__(self, port=8097, viz=None, env_name=None):
if viz is None:
self.viz = Visdom(port=port, env=env_name)
else:
self.viz = viz
def add_Line_windows(self, name, X=0, Y=0):
w = self.viz.line(X=np.array([X]),
Y=np.array([Y]),
opts=dict(title=name))
return w
def update_line(self, w, X, Y):
self.viz.line(X=np.array([X]), Y=np.array([Y]), win=w, update="append")
def show_image(self, image):
plt.imshow(image)
self.viz.matplot(plt)
class VisdomPlotter:
def __init__(self):
self.viz = Visdom(env='BachPropagation')
self.plots = {}
def plot_line(self,
plot_name,
line_label=None,
title=None,
y_label=None,
x=None,
y=None,
color=None):
if plot_name not in self.plots:
opts = {
'title': title,
'xlabel': 'Epochs',
'ylabel': y_label,
'linecolor': color
}
if line_label is not None:
opts['legend'] = [line_label]
self.plots[plot_name] = self.viz.line(X=x, Y=y, opts=opts)
else:
self.viz.line(X=x,
Y=y,
win=self.plots[plot_name],
name=line_label,
update='append',
opts={'linecolor': color})
def display_matplot_figure(self, figure, plot_name):
if plot_name not in self.plots:
self.plots[plot_name] = self.viz.matplot(figure)
else:
self.viz.matplot(figure, win=self.plots[plot_name])
def add_song(self, path):
self.viz.audio(audiofile=path, tensor=None)
def train_test_model(self, model_info):
model_name = model_info["model_name"]
print("[Current model] : ", model_name, "\nModel setting : ",
model_info)
model, model_info = model_handler.create_model_object(model_info)
model = self.model_cuda_selecter(model)
# If you want to use pre-trained model, and train the last layer, set the "train_last_layer" as True
if model_info["train_last_layer"]:
model = layer_manipulation.train_last_layer(model)
# create transform object.
train_loader = self.__create_dataloader(self.train_folder,
self.train_transforms,
model_info["input_size"],
model_info["batch_size"],
model_info["mean"],
model_info["std"], True)
test_loader = self.__create_dataloader(self.test_folder,
self.test_transforms,
model_info["input_size"],
model_info["batch_size"],
model_info["mean"],
model_info["std"], False)
loss_func = loss_func_handler.create_loss_func(
loss_func_name=model_info["loss_func"], **model_info["lf_setting"])
optimizer = optimizer_handler.create_optimizer(
model,
learning_rate=model_info["lr"],
optimizer_name=model_info["optimizer"],
**model_info["optimizer_setting"])
print(model_info["learning_scheduler_setting"])
scheduler = scheduler_handler.create_train_scheduler(
optimizer, model_info["learning_scheduler"],
**model_info["learning_scheduler_setting"])
epoch_train_acc_list, epoch_train_loss_list = [], []
epoch_test_acc_list, epoch_test_loss_list = [], []
# create a several folders into the log folder
model_main_path = os.path.join(self.log_folder, model_name)
model_save_path = os.path.join(self.log_folder, model_name,
"model_epoch")
best_model_save_path = os.path.join(self.log_folder, model_name,
"best_model")
if os.path.exists(model_main_path) == False:
os.mkdir(model_main_path)
if os.path.exists(model_save_path) == False:
os.mkdir(model_save_path)
if os.path.exists(best_model_save_path) == False:
os.mkdir(best_model_save_path)
txt_path = os.path.join(model_main_path, model_name + "_train_log.txt")
model_txt_path = os.path.join(model_main_path,
model_name + "_model_architecture.txt")
train_environment_path = os.path.join(
model_main_path, model_name + "_train_environment.txt")
excel_path = os.path.join(model_main_path,
model_name + "_train_result.xlsx")
result_excel_log = {
"train_acc": epoch_train_acc_list,
"train_loss": epoch_train_loss_list,
"test_acc": epoch_test_acc_list,
"test_loss": epoch_test_loss_list
}
logger.save_txt(model_txt_path, model.__str__())
logger.save_txt(train_environment_path, train_loader.__str__())
logger.save_txt(train_environment_path, test_loader.__str__())
logger.save_txt(train_environment_path, str(model_info))
best_acc = 0.0
best_loss = 100000.0
logger.save_txt(txt_path,
"EPOCH, train_acc, train_loss, test_acc, test_loss")
# Visdom logging session
if self.use_visdom:
model_visdom = Visdom(env=model_name)
self.visdom.append(model_visdom)
model_visdom.text(model.__str__(),
opts=dict(title="model architecture"))
train_environment = model_visdom.text(
train_loader.__str__(), opts=dict(title="Training setting"))
model_visdom.text(test_loader.__str__(),
win=train_environment,
opts=dict(title="Training setting"),
append=True)
model_visdom.text(str(model_info),
win=train_environment,
opts=dict(title="Training setting"),
append=True)
model_acc_graph = lineplotstream(
model_visdom, "{} Train Test Accuracy".format(model_name))
model_loss_graph = lineplotstream(
model_visdom, "{} Train Test Loss".format(model_name))
weight_change = logger.PolygonHistogram3D("Weight", "Epoch",
"Number of Count",
"Weight Histogram")
weight_change_plot = model_visdom.matplot(weight_change.plot())
# Train, test model
for epoch in range(1, model_info["max_epoch"] + 1):
torch.set_grad_enabled(True)
model, train_acc, train_loss = self.train_model(
epoch, train_loader, model, loss_func, optimizer)
scheduler.step()
torch.set_grad_enabled(False)
test_acc, test_loss = self.test_model(epoch, test_loader, model,
loss_func)
epoch_train_acc_list.append(train_acc)
epoch_train_loss_list.append(train_loss)
epoch_test_acc_list.append(test_acc)
epoch_test_loss_list.append(test_loss)
text_log = [
str(x)
for x in [epoch, train_acc, train_loss, test_acc, test_loss]
]
text_log = "\t".join(text_log)
logger.save_txt(txt_path, text_log)
#print(result_excel_log)
logger.save_excel(
excel_path,
[("{}_result".format(model_name), result_excel_log)])
if self.use_visdom:
model_acc_graph.update([train_acc], [epoch], legend="Train")
model_acc_graph.update([test_acc], [epoch], legend="Test")
model_loss_graph.update([train_loss], [epoch], legend="Train")
model_loss_graph.update([test_acc], [epoch], legend="Test")
weights = logger.trainable_parameters(model)
print(weights)
weights = [
weight.detach().numpy().reshape(-1) for weight in weights
]
weights = np.concatenate(weights)
print(weights)
weight_change.update(weights)
model_visdom.matplot(weight_change.plot(),
win=weight_change_plot)
model_handler.save_checkpoint(
model,
os.path.join(model_save_path, "epoch_{}.pth".format(epoch)))
if best_acc < test_acc:
model_handler.save_checkpoint(
model,
os.path.join(best_model_save_path,
"best_acc_model.pth".format(epoch)))
best_acc = test_acc
if best_loss > test_loss:
model_handler.save_checkpoint(
model,
os.path.join(best_model_save_path,
"best_loss_model.pth".format(epoch)))
best_loss = test_loss
return model, epoch_train_acc_list, epoch_test_loss_list, epoch_test_acc_list, epoch_test_loss_list
class VisdomPlotter(object):
def __init__(self, env_name='main', port=80, hostname='localhost'):
'''
Params:
* env_name : str
* port : int
* hostname : str
'''
self.server_is_running = server_is_running(port, hostname)
self.viz = Visdom(port=port) if self.server_is_running else None
self.env = env_name
self.plots = {}
def line_plot(self, var_name, split_name, title_name, x, y, x_label='Epochs'):
'''
Params:
* var_name : variable name (e.g. loss, acc)
* split_name : split name (e.g. train, val)
* title_name : titles of the graph (e.g. Classification Accuracy)
* x : x axis value (e.g. epoch number)
* y : y axis value (e.g. epoch loss)
'''
if not self.server_is_running:
return
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=x_label,
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 heatmap_plot(self, var_name, split_name, title_name, data, x_label='Epochs'):
'''
Params:
* var_name : str
* split_name : str
* title_name : str
* data
* x_label (optional) : str
'''
if not self.server_is_running and data is not None:
return
labels = list(map(lambda x: x[0], data[0]))
y = np.array([list(map(lambda x: x[1], line)) for line in data]).T
ax = sns.heatmap(y, yticklabels=labels, cbar=False)
plt.tight_layout()
plt.title(title_name)
plt.xlabel(x_label)
plt.ylabel('layers')
self.viz.matplot(plt, win=var_name, env=self.env)
def matplot_plot(self, var_name, data):
'''
Params:
* var_name : str
* data
'''
if not self.server_is_running and data is not None:
return
self.viz.matplot(data, win=var_name, env=self.env)
def image_plot(self, var_name, image):
'''
Params:
* var_name : str
* image
'''
if not self.server_is_running or image is None:
return
self.viz.image(image, win=var_name, env=self.env)
class visual_block(nn.Module):
def __init__(self, max_row=10, max_column=10, w=60, h=60, margin_right=2, margin_top=2, dpi=30):
super(visual_block, self).__init__()
self.axes_pool = {}
self.max_column = max_column
self.max_row = max_row
self.dpi = dpi
self.layer_count = 0
self.layer_pool = {}
self.image_figure_pool = {}
self.figure_width = self.max_column * (w + margin_right) * 1.0 / dpi
self.figure_height = self.max_row * (h + margin_top) * 1.0 / dpi
self.image_w = w * 1.0 / (self.figure_width * dpi)
self.image_h = h * 1.0 / (self.figure_height * dpi)
self.width_bias = margin_right / (self.figure_width * dpi)
self.height_bias = margin_top / (self.figure_height * dpi)
self.viz = Visdom()
def update_graph(self, figure, data, params):
data = data.cpu()
# 解析参数
options = {'mode': 'feature_map', 'layer': 'conv1', 'channel_num': 'all',
'cmap': None, 'image': ''}
for i in params:
if i in options:
options[i] = params[i]
if options['channel_num'] == 'all':
channel_num = len(data)
else:
channel_num = options['channel_num']
index = options['image'] + '_' + options['layer']
if index not in self.axes_pool:
if options['layer'] not in self.layer_pool:
self.layer_count += 1
self.layer_pool[options['layer']] = self.layer_count
cli = current_layer_index = self.layer_pool[options['layer']]
self.axes_pool[index] = []
for i in range(len(data)):
width = 1.0 / self.max_column - self.width_bias
height = 1.0 / self.max_row - self.height_bias
left = (cli - 1) * 1.0 / self.max_column
bottom = (self.max_row - i - 1) * 1.0 / self.max_row
react = [left, bottom, width, height]
axi = figure.add_axes(react)
axi.axis('off')
self.axes_pool[index].append(axi)
axes = self.axes_pool[index]
if options['mode'] == 'source_image':
data = self.denormalize(data)
img = data.detach().numpy()
img = np.transpose(img, [1, 2, 0])
axes[0].imshow(img)
else:
data = data.detach().numpy()
if channel_num > self.max_row:
channel_num = self.max_row
for channel in range(channel_num):
axes[channel].imshow(data[channel], cmap=options['cmap'])
return figure
def denormalize(self, x_hat):
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
mean = torch.tensor(mean).unsqueeze(1).unsqueeze(1)
std = torch.tensor(std).unsqueeze(1).unsqueeze(1)
x = x_hat * std + mean
return x
def forward(self, x):
"""
可视化
Options:
mode: 'feature_map','kernel_weights','source_image'
指定输入的tensor是属于特征图、卷积核，或者是原图
layer: 'conv1',''
当前层名称,必填。用于索引 ax 子图
channel: 'all',1,2,...
指定要显示的特征图或者卷积核参数数量
"""
params = x[1]
x = x[0]
show_data = x
for image in range(len(show_data)):
if 'image%d' % image not in self.image_figure_pool:
fig = plt.figure(figsize=(self.figure_width, self.figure_height), dpi=self.dpi)
fig.tight_layout()
self.image_figure_pool['image%d' % image] = fig
params['image'] = 'image%d' % image
figure = self.image_figure_pool['image%d' % image]
self.update_graph(figure, show_data[image], params)
if 'end' in params and params['end']:
self.viz.matplot(plot=self.image_figure_pool['image%d' % image], win='image%d' % image,
env='image%d' % image)
return x
elif event['key'] == 'Backspace':
curr_txt = curr_txt[:-1]
elif event['key'] == 'Delete':
curr_txt = txt
elif len(event['key']) == 1:
curr_txt += event['key']
viz.text(curr_txt, win=callback_text_window)
viz.register_event_handler(type_callback, callback_text_window)
# matplotlib demo:
try:
import matplotlib.pyplot as plt
plt.plot([1, 23, 2, 4])
plt.ylabel('some numbers')
viz.matplot(plt)
except BaseException as err:
print('Skipped matplotlib example')
print('Error message: ', err)
# 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
video_url = 'http://media.w3.org/2010/05/sintel/trailer.ogv'
# linux
image, label = data
if (cuda_predicate == True):
image = image.cuda()
label = label.cuda()
output = model(image)
optimizer.zero_grad()
loss = criterion(output, label)
loss.backward()
optimizer.step()
print("Batch Num:", i)
if i % WINDOW == WINDOW - 1:
torch.save(model, './current_model')
print(loss.item())
LOSSES.append(loss.item())
plt.plot(LOSSES)
viz.matplot(plt, win="loss")
#===============================================================================
# Testing Phase
#===============================================================================
correct = 0
with torch.no_grad():
for image, label in test_loader:
if (cuda_predicate == True):
image = image.cuda()
label = label.cuda()
output = model(image)
pred = output.max(1, keepdim=True)[1]
correct += pred.eq(label.view_as(pred)).sum().item()
print("Accuracy: ", correct)
class VisdomLogger():
def __init__(self,**kwargs):
if Visdom is None:
self.viz = None # do nothing
return
self.connected = True
try:
self.viz = Visdom(raise_exceptions=True,**kwargs)
except Exception as e:
print("Could not reach visdom server...")
self.connected = False
pass
self.windows = dict()
r = np.random.RandomState(1)
self.colors = r.randint(0,255, size=(255,3))
self.colors[0] = np.array([1., 1., 1.])
self.colors[1] = np.array([0. , 0.18431373, 0.65490196]) # ikb blue
def update(self, data):
if self.connected:
self.plot_epochs(data)
def bar(self,X, name="barplot"):
if self.connected:
X[np.isnan(X)] = 0
win = name.replace(" ","_")
opts = dict(
title=name,
xlabel='t',
ylabel="P(t)",
width=600,
height=200,
marginleft=20,
marginright=20,
marginbottom=20,
margintop=30
)
self.viz.bar(X,win=win,opts=opts)
def plot(self, X, name="plot",**kwargs):
if self.connected:
X[np.isnan(X)] = 0
win = "pl_"+name.replace(" ","_")
opts = dict(
title=name,
xlabel='t',
ylabel="P(t)",
width=600,
height=200,
marginleft=20,
marginright=20,
marginbottom=20,
margintop=30,
**kwargs
)
self.viz.line(X ,win=win, opts=opts)
def confusion_matrix(self, cm, title="Confusion Matrix", norm=None):
if self.connected:
plt.clf()
if norm is not None:
cm /= np.expand_dims(cm.sum(norm),axis=norm)
cm[np.isnan(cm)] = 0
cm[np.isinf(cm)] = 0
vmin = 0
vmax = 1
else:
vmin = None
vmax = None
name=title
plt.rcParams['figure.figsize'] = (9, 9)
#sn.set(font_scale=1.4) # for label size
ax = sn.heatmap(cm, annot=True, annot_kws={"size": 11}, vmin=vmin, vmax=vmax) # font size
ax.set(xlabel='ground truth', ylabel='predicted', title=title)
plt.tight_layout()
opts = dict(
resizeable=True
)
self.viz.matplot(plt, win=name, opts=opts)
def plot_class_p(self,X):
if self.connected:
plt.clf()
x = X.detach().cpu().numpy()
plt.plot(x[0, :])
name="confusion matrix"
plt.rcParams['figure.figsize'] = (6, 6)
#sn.set(font_scale=1.4) # for label size
ax = sn.heatmap(cm, annot=True, annot_kws={"size": 11}) # font size
ax.set(xlabel='ground truth', ylabel='predicted', title="Confusion Matrix")
plt.tight_layout()
opts = dict(
resizeable=True
)
self.viz.matplot(plt, win=name, opts=opts)
def plot_boxplot(self, labels, t_stops, tmin=None, tmax=None):
if self.connected:
grouped = [t_stops[labels == i] for i in np.unique(labels)]
#legend = ["class {}".format(i) for i in np.unique(labels)]
plt.clf()
name = "boxplot"
plt.rcParams['figure.figsize'] = (9, 9)
# sn.set(font_scale=1.4) # for label size
ax = sn.boxplot(data=grouped, orient="h")
ax.set_xlabel("t_stop")
ax.set_ylabel("class")
ax.set_xlim(tmin, tmax)
#ax = sn.heatmap(cm, annot=True, annot_kws={"size": 11}, vmin=vmin, vmax=vmax) # font size
#ax.set(xlabel='ground truth', ylabel='predicted', title=title)
plt.tight_layout()
opts = dict(
resizeable=True
)
self.viz.matplot(plt, win=name, opts=opts)
pass
def plot_epochs(self, data):
"""
Plots mean of epochs
:param data:
:return:
"""
if self.connected:
data_mean_per_epoch = data.groupby(["mode", "epoch"]).mean()
cols = data_mean_per_epoch.columns
modes = data_mean_per_epoch.index.levels[0]
for name in cols:
if name in self.windows.keys():
win = self.windows[name]
update = 'new'
else:
win = name # first log -> new window
update = None
opts = dict(
title=name,
showlegend=True,
xlabel='epochs',
ylabel=name)
for mode in modes:
epochs = data_mean_per_epoch[name].loc[mode].index
values = data_mean_per_epoch[name].loc[mode]
win = self.viz.line(
X=epochs,
Y=values,
name=mode,
win=win,
opts=opts,
update=update
)
update='insert'
self.windows[name] = win
class ppo:
def __init__(self,
env_fn,
seed=0,
steps_per_epoch=4000,
epochs=50,
gamma=0.99,
clip_ratio=0.2,
pi_lr=3e-4,
vf_lr=1e-3,
train_pi_iters=80,
train_v_iters=80,
lam=0.97,
max_ep_len=1000,
target_kl=0.01,
save_freq=10,
exp_name='',
test_agent=False,
load=False):
# seed
self.seed = seed
self.seed += 10000 * proc_id()
tf.set_random_seed(self.seed)
np.random.seed(self.seed)
random.seed(self.seed)
# Hyper-parameter
self.env = env_fn()
self.obs_dim = self.env.observation_space.shape
self.act_dim = self.env.action_space.shape
self.gamma, self.lam = gamma, lam
self.steps_per_epoch, self.epochs = steps_per_epoch, epochs
self.pi_lr, self.vf_lr = pi_lr, vf_lr
self.clip_ratio, self.target_kl = clip_ratio, target_kl
self.train_pi_iters, self.train_v_iters = train_pi_iters, train_v_iters
self.exp_name, self.save_freq, self.max_ep_len = exp_name, save_freq, max_ep_len
if test_agent:
from visdom import Visdom
self.viz = Visdom()
assert self.viz.check_connection()
self.win = self.viz.matplot(plt)
# Experience buffer
self.buf = PPOBuffer(self.obs_dim, self.act_dim, self.steps_per_epoch,
self.gamma, self.lam)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.Session(config=config, graph=self.graph)
self.__make_model()
self.sess.run(tf.global_variables_initializer())
# Sync params across processes
self.sess.run(sync_all_params())
Count_Variables()
print('Trainable_variables:')
for v in tf.compat.v1.trainable_variables():
print('{}\t {}'.format(v.name, str(v.shape)))
var_list = tf.global_variables()
self.saver = tf.compat.v1.train.Saver(var_list=var_list,
max_to_keep=1)
# summary
self.writer = tf.compat.v1.summary.FileWriter("logs/" + exp_name)
if load:
self.load()
self.sess.run(sync_all_params())
self.ep_ret_ph = tf.placeholder(tf.float32,
shape=(),
name="ep_ret_ph")
self.ep_Entropy_ph = tf.placeholder(tf.float32,
shape=(),
name="Entropy")
self.clipfrac_ph = tf.placeholder(tf.float32,
shape=(),
name="clipfrac")
self.ep_len_ph = tf.placeholder(tf.float32,
shape=(),
name="ep_len_ph")
self.test_summary = tf.compat.v1.summary.merge([
tf.compat.v1.summary.scalar('EP_ret',
self.ep_ret_ph,
family='test'),
tf.compat.v1.summary.scalar('EP_len',
self.ep_len_ph,
family='test')
])
self.entropy_summary = tf.compat.v1.summary.merge([
tf.compat.v1.summary.scalar('Entropy',
self.ep_Entropy_ph,
family='test'),
tf.compat.v1.summary.scalar('clipfrac',
self.clipfrac_ph,
family='test')
])
def __make_model(self):
self.o1_ph = tf.placeholder(tf.float32, [None, 128, 128, 3],
name='o1_ph')
self.o2_ph = tf.placeholder(tf.float32, [None, 128, 128, 3],
name='o2_ph')
self.o_low_dim_ph = tf.placeholder(tf.float32, [None, 5],
name='o_low_dim_ph')
self.f_s_ph = tf.placeholder(tf.float32, [None, 14], name='f_s_ph')
a_ph, adv_ph, ret_ph, logp_old_ph = ppo_core.placeholders(
self.act_dim, None, None, None)
pi, logp, logp_pi, self.v = ppo_core.mlp_actor_critic(
self.o1_ph,
self.o2_ph,
self.o_low_dim_ph,
self.f_s_ph,
a_ph,
action_space=self.env.action_space)
self.all_phs = [
self.o1_ph, self.o2_ph, self.o_low_dim_ph, self.f_s_ph, a_ph,
adv_ph, ret_ph, logp_old_ph
]
self.get_action_ops = [pi, self.v, logp_pi]
# PPO objectives
ratio = tf.exp(logp - logp_old_ph) # pi(a|s) / pi_old(a|s)
min_adv = tf.where(adv_ph > 0, (1 + self.clip_ratio) * adv_ph,
(1 - self.clip_ratio) * adv_ph)
self.pi_loss = -tf.reduce_mean(tf.minimum(ratio * adv_ph, min_adv))
self.v_loss = tf.reduce_mean((ret_ph - self.v)**2)
# Info (useful to watch during learning)
self.approx_kl = tf.reduce_mean(
logp_old_ph -
logp) # a sample estimate for KL-divergence, easy to compute
self.approx_ent = tf.reduce_mean(
-logp) # a sample estimate for entropy, also easy to compute
clipped = tf.logical_or(ratio > (1 + self.clip_ratio), ratio <
(1 - self.clip_ratio))
self.clipfrac = tf.reduce_mean(tf.cast(clipped, tf.float32))
# Optimizers
self.train_pi = MpiAdamOptimizer(learning_rate=self.pi_lr).minimize(
self.pi_loss)
self.train_v = MpiAdamOptimizer(learning_rate=self.vf_lr).minimize(
self.v_loss)
def update(self):
inputs = {k: v for k, v in zip(self.all_phs, self.buf.get())}
# Training
for i in range(self.train_pi_iters):
_, kl = self.sess.run([self.train_pi, self.approx_kl],
feed_dict=inputs)
kl = mpi_avg(kl)
if kl > 1.5 * self.target_kl:
print(
'process %d: Early stopping at step %d due to reaching max kl.'
% (proc_id(), i))
break
for _ in range(self.train_v_iters):
self.sess.run(self.train_v, feed_dict=inputs)
def rollout(self):
o, r, d, ep_ret, ep_len = self.env.reset(), 0, False, 0, 0
# Main loop: collect experience in env and update/log each epoch
for epoch in tqdm(range(self.epochs)):
for t in range(self.steps_per_epoch):
f_s = self.env.full_state()
a, v_t, logp_t = self.sess.run(self.get_action_ops,
feed_dict={
self.o1_ph:
o[0][np.newaxis, ],
self.o2_ph:
o[1][np.newaxis, ],
self.o_low_dim_ph:
o[2][np.newaxis, ],
self.f_s_ph:
f_s[np.newaxis, ]
})
o2, r, d, _ = self.env.step(a[0])
ep_ret += r
ep_len += 1
# save and log
self.buf.store(o, f_s, a, r, v_t, logp_t)
# Update obs (critical!)
o = o2
terminal = d or (ep_len == self.max_ep_len)
if terminal or (t == self.steps_per_epoch - 1):
if not (terminal):
print(
'process %d: trajectory cut off by epoch at %d steps.'
% (proc_id(), ep_len))
# if trajectory didn't reach terminal state, bootstrap value target
last_val = 0 if d else self.sess.run(
self.v,
feed_dict={
self.o1_ph: o[0][np.newaxis, ],
self.o2_ph: o[1][np.newaxis, ],
self.o_low_dim_ph: o[2][np.newaxis, ],
self.f_s_ph: f_s[np.newaxis, ]
})
self.buf.finish_path(last_val)
if proc_id() == 0:
ep_s = self.sess.run(self.test_summary, {
self.ep_ret_ph: ep_ret,
self.ep_len_ph: ep_len
})
self.writer.add_summary(ep_s, global_step=epoch)
o, ep_ret, ep_len = self.env.reset(), 0, 0
# Save model
if (epoch % self.save_freq == 0) or (epoch == self.epochs - 1):
if proc_id() == 0:
self.Save()
# Perform PPO update!
self.update()
def Save(self):
path = "model/" + self.exp_name + "/model.ckpt"
print("process %d: Save model to the: '{}'".format(path) % proc_id())
self.saver.save(self.sess, save_path=path)
def load(self):
path = "model/" + self.exp_name + "/model.ckpt"
print("\nLoad model From the: '{}'\n".format(path))
self.saver.restore(self.sess, save_path=path)
def _choose_action(self, o):
a, v_t, logp_t = self.sess.run(self.get_action_ops,
feed_dict={
self.o1_ph: o[0][np.newaxis, ],
self.o2_ph: o[1][np.newaxis, ],
self.o_low_dim_ph:
o[2][np.newaxis, ]
})
return a[0], v_t
def test_agent(self, n=1):
ep_r = []
ep_l = []
plt.figure(figsize=(9, 5))
for j in range(n):
print('------------------Epoch: %d-------------------' % j)
o, r, d, ep_ret, ep_len, level_plot = self.env.reset(
), 0, False, 0, 0, []
v_plot = []
while not (d or (ep_len == self.max_ep_len)):
# Take deterministic actions at test time
a, v_t = self._choose_action(o)
print('V: {}'.format(v_t[0]))
v_plot.append(v_t)
# _ = input()
o, r, d, info = self.env.step(a, _step=ep_len)
print("step: {} a: {} r: {}".format(ep_len,
np.around(a, decimals=2),
r))
ep_ret += r
ep_len += 1
level_plot.append(self.env._rank_before)
self.win = R_plot(level_plot, v_plot, self.viz, self.win)
print('EP_ret: %d, \t EP_Len: %d' % (ep_ret, ep_len))
if info['is_success']:
print('success!')
ep_r.append(ep_ret), ep_l.append(ep_len)
plt.close()
type=int,
default=201,
help='case id to visualize.')
args = parser.parse_args()
# --------- Connect to visdom server -------------------------------------------------
viz = Visdom(server='http://127.0.0.1', port=8097, env='test')
assert viz.check_connection()
# img_data = load_volume(args.case_num).get_fdata()
img_data = \
Kits2019DataLoader3D.load_patient(os.path.join('/home/data_share/npy_data/', str(args.case_num).zfill(5)))[0][1]
img_shape = img_data.shape
print(img_data.dtype)
assert len(np.unique(img_data)) == 3
print(img_shape)
plt.imshow(img_data[:, :, img_shape[-1] // 2], cmap='gray')
viz.matplot(plt, opts={'title': 'case_original', 'showlegend': True})
mirrored = np.flip(img_data.copy(), axis=(0, 1, 2))
# print(mirrored.shape)
# plt.imshow(mirrored[img_shape[-1] // 2, :, :], cmap='gray')
# viz.matplot(
# plt,
# opts={
# 'title': 'case_mirrored',
# 'showlegend': True
# }
# )
patches = image.extract_patches(img_data, patch_size, strides)
print(patches.shape)
m_patches = image.extract_patches(mirrored, patch_size, strides)
print(m_patches.shape)
# patches = patches.reshape((-1, patch_size[0], patch_size[1], patch_size[2]))
def main():
global args
args = parser.parse_args()
# Get options
opt_list = []
for filename in args.path_opts:
with open(filename, 'r') as f:
opt = yaml.load(f)
opt_list.append(opt)
with open(args.vis_path_opt, 'r') as handle:
vis_options = yaml.load(handle)
print('## args')
pprint(vars(args))
# Get the relevant curves
algo = args.algo
env_name = args.env_name
num_trials = args.num_trials
trial_offset = args.trial_offset
x_curves = []
mean_curves = []
bottom_curves = []
top_curves = []
for opt in opt_list:
# Get the curves for all the trials for each yaml file (mean and variance)
x_curve, mean_curve, top_curve, bottom_curve = get_eval_vals(
opt, vis_options, args.eval_key, algo, env_name, num_trials,
trial_offset, args.bin_size, args.smooth, args.mode)
x_curves.append(x_curve)
mean_curves.append(mean_curve)
bottom_curves.append(bottom_curve)
top_curves.append(top_curve)
assert (len(x_curves) == len(args.legend_names))
assert (len(mean_curves) == len(args.legend_names))
assert (len(bottom_curves) == len(args.legend_names))
assert (len(top_curves) == len(args.legend_names))
# Plot the curves
if args.no_vis:
raise Exception("No, we need visdom to display")
vis = Visdom(port=args.port)
plt.figure()
for leg_name, x, mean, bot, top in zip(args.legend_names, x_curves,
mean_curves, bottom_curves,
top_curves):
# Subsample (because why not?)
x = x[::100]
mean = mean[::100]
p = plt.plot(x, mean, label=leg_name)
if args.mode == 'all':
for curve in top:
curve = curve[::100]
plt.plot(x, curve)
else:
bot = bot[::100]
top = top[::100]
plt.fill_between(x, bot, top, alpha=0.2)
plt.legend()
plt.xlabel('Number of Timesteps')
plt.ylabel(args.eval_key)
plt.title(args.eval_key)
plt.grid(True)
vis.matplot(plt)
plt.close()
policy_losses = np.zeros(args.train_episodes)
value_losses = np.zeros(args.train_episodes)
lengths = np.zeros(args.train_episodes)
rewards_mean = np.zeros(args.train_episodes // args.episodes_per_eval + 1)
rewards_std = np.zeros(args.train_episodes // args.episodes_per_eval + 1)
rewards_mean[0], rewards_std[0] = a2c.eval(args.test_episodes)
print('episode', 0, 'reward average', rewards_mean[0], 'reward std', rewards_std[0])
plt.xlabel('episodes')
plt.ylabel('average reward')
errbar = plt.errorbar(np.arange(1), rewards_mean[:1], rewards_std[:1], capsize=3)
viz = Visdom()
policy_loss_plot = None
value_loss_plot = None
length_plot = None
reward_plot = viz.matplot(plt, env=args.task_name)
for i in range(args.train_episodes):
policy_losses[i], value_losses[i], lengths[i] = a2c.train(args.gamma, args.r_scale)
if (i + 1) % args.episodes_per_plot == 0:
if policy_loss_plot is None:
opts = dict(xlabel='episodes', ylabel='policy loss')
policy_loss_plot = viz.line(X=np.arange(1, i + 2), Y=policy_losses[:i + 1],
env=args.task_name, opts=opts)
else:
viz.line(X=np.arange(i - args.episodes_per_plot + 1, i + 2),
Y=policy_losses[i - args.episodes_per_plot:i + 1],
env=args.task_name, win=policy_loss_plot, update='append')
if value_loss_plot is None:
opts = dict(xlabel='episodes', ylabel='value loss')
value_loss_plot = viz.line(X=np.arange(1, i + 2), Y=value_losses[:i + 1],
class Visdom_Plot(object):
def __init__(self, title, port=8097, env_name='main'):
#当‘env_name'不是main时，创建一个新环境
self.viz = Visdom(port=port, env=env_name)
self.loss_win = {}
self.acc_win = {}
self.text_win = {}
self.plt_img_win = {}
self.images_win = {}
self.gray_win = {}
self.title = title
def _new_win(self,
type='loss_win',
win_name='default_loss_win',
id='train_loss',
H_img=100):
'''
type: loss_win, acc_win, text_win, plt_img_win
name: default is the default win in class. you can specify a window's name
id: the line's name
'''
assert type in [
'loss_win', 'acc_win', 'text_win', 'plt_img_win', 'gray_win'
], "win type must a string inside ['loss_win', 'acc_win', 'text_win', 'plt_img_win'] "
if type == 'loss_win':
self.loss_win[win_name] = self.viz.line(X=np.array([0]),
Y=np.array([0]),
name=id,
opts=dict(
xlabel='Epoch.batch',
ylabel='Loss',
title=win_name,
marginleft=60,
marginbottom=60,
margintop=80,
width=800,
height=600,
))
elif type == 'acc_win':
self.acc_win[win_name] = self.viz.line(X=np.array([0]),
Y=np.array([0]),
name=id,
opts=dict(
xlabel='Epoch.batch',
ylabel='Top1 accuracy',
title=win_name,
showlegend=True,
markercolor=np.array(
[[255, 0, 0]]),
marginleft=60,
marginbottom=60,
margintop=60,
width=800,
height=600,
))
elif type == 'plt_img_win' or type == 'gray_win':
getattr(self,
type)[win_name] = self.viz.images(np.random.randn(
1, 3, 100, 100),
opts=dict(
height=H_img * 5,
width=H_img * 5,
))
elif type == 'text_win':
self.text_win[win_name] = self.viz.text('Text Window')
def append_loss(self,
loss,
epoch_batches,
win_name='default_loss_win',
id='train_loss'):
if win_name not in self.loss_win:
self._new_win(type='loss_win', win_name=win_name, id=id)
self.viz.line(X=np.array([epoch_batches]),
Y=np.array([loss]),
win=self.loss_win[win_name],
name=id,
opts=dict(showlegend=True),
update='append')
def append_acc(self,
train_acc,
epoch_batches,
win_name='default_acc_win',
id='train_acc'):
if win_name not in self.acc_win:
self._new_win(type='acc_win', win_name=win_name, id=id)
self.viz.line(X=np.array([epoch_batches]),
Y=np.array([train_acc]),
win=self.acc_win[win_name],
name=id,
opts=dict(showlegend=True),
update='append')
def lr_scatter(self, epoch, lr, win_name='default_acc_win'):
self.viz.scatter(
X=np.array([[epoch, 20]]),
name='lr=' + str(lr),
win=self.acc_win[win_name],
opts=dict(showlegend=True),
update='append',
)
def img_plot(self, images, lm=None, mode='update', caption=''):
'''
Input:
images : tensors, N x 3 x H x W, so transfer to N x H x W x 3 is needed
lm : N x K x 2, is not None, then landmarks will be scattered.
'''
win_exist = len(self.plt_img_win)
N, C, H, W = images.size()
if N > win_exist:
for i in range(win_exist, N, 1):
self._new_win(type='plt_img_win',
win_name='image' + str(i),
H_img=H)
if lm is not None:
N, K, m = lm.size()
assert N == images.size(
)[0] and m == 2, "landmarks have illegal size"
lm = lm.cpu()
images = images.cpu()
plt.figure(figsize=(H * 0.06, W * 0.06))
for n, image in enumerate(images[:]):
# print(image.size())
image = image.transpose(0, 1).transpose(1, 2)
# print(image.size())
plt.imshow(image.detach().numpy(
)) # convert to H x W x 3. plt的输入是HxWx3，而viz.images())的输入是3xHxW
if lm is not None:
color = np.linspace(0, 1, num=K)
plt.scatter(x=lm[n, :, 0].detach().numpy(),
y=lm[n, :, 1].detach().numpy(),
c=color,
marker='x',
s=200)
self.viz.matplot(plt,
win=self.plt_img_win['image' + str(n)],
opts=dict(caption='image' + str(n)))
plt.clf()
def images(self, images, win_name='default_images_win'):
'''
Input:
images:N x 3 x H x W, tensors
'''
images = images.cpu()
if win_name not in self.images_win:
self.images_win[win_name] = self.viz.images(
images.detach().numpy())
else:
self.viz.images(images.detach().numpy(),
win=self.images_win[win_name])
def gray_images(self, images, win_name='default_gray_win'):
'''
Input:
images : K x H x W, tensors
'''
images = images.cpu()
win_exist = len(self.gray_win)
K, H, W = images.size()
if K > win_exist:
for i in range(win_exist, K, 1):
self._new_win(type='gray_win',
win_name='gray' + str(i),
H_img=H // 2)
plt.figure(figsize=(H / 2 * 0.06, W / 2 * 0.06))
for n, image in enumerate(images):
plt.imshow(image.detach().numpy(
)) # convert to H x W x 3. plt的输入是HxWx3，而viz.images())的输入是3xHxW
self.viz.matplot(plt, win=self.gray_win['gray' + str(n)])
plt.clf()
def append_text(self, text, win_name='default_text_win', append=True):
if win_name not in self.text_win:
self._new_win(type='text_win', win_name=win_name)
self.viz.text(text, win=self.text_win[win_name], append=append)
class GenericVisdomPlotter(GenericPlotter):
'''
Visdom based generic plotter implementation
'''
def __init__(self, config):
'''
Initializer
'''
super(GenericVisdomPlotter, self).__init__(config)
# prepare the environment name
self.env = 'Cami_' + config
# default host and port
hostname = 'deep' # "deep"
baseurl = "/visdom"
myport = 80
# replace host and port by the one provided in the config file
# if 'hostname' in config['visualization']:
# hostname = config['visualization']['hostname']
# if 'port' in config['visualization']:
# port = int(config['visualization']['port'])
# initialize the object for visualization
self.viz = Visdom(server=hostname,
base_url=baseurl,
port=myport,
use_incoming_socket=False,
use_polling=True)
# the dictionary of plots and figures
self.figures = dict()
self.plots = dict()
# initialize the current epoch in 0
self.current_epoch = 0
def plot(self, plot_name, split_name, x, y, x_label='Epochs'):
'''
Plot a line plot
'''
# if the plot is not in the dictionary, initialize one
if (plot_name not in self.plots):
self.plots[plot_name] = self.viz.line(X=np.array([x, x]),
Y=np.array([y, y]),
env=self.env,
opts=dict(
legend=[str(split_name)],
title=plot_name,
xlabel=x_label,
ylabel=plot_name))
# if the plot is already there, update
else:
self.viz.line(X=np.array([x]),
Y=np.array([y]),
env=self.env,
update='append',
win=self.plots[plot_name],
name=str(split_name))
def plot_matplotlib(self, plt, plot_name="matplotlib_progreso"):
self.viz.matplot(plt, env=self.env, win=plot_name)
def plot_multiple_statistics(self, plot_name, x, y_values):
'''
Plot multiple statistics within the same plot
'''
# get the split names
split_names = y_values.keys()
# iterate for each of them
for split in split_names:
# plot the values
self.plot(plot_name, split, x, y_values[split])
def plot_text(self, info, plot_name="some_text"):
'''
Plot some text
'''
self.viz.text(info, env=self.env, win=plot_name)
def plot_scalar(self, plot_name, x, y, legend):
'''
Plot a line plot
'''
self.plot(plot_name, legend, x, y)
def display_image(self, image_key, image, caption=''):
'''
Display given image in the plot
'''
# if the image is already in the plot, remove it to replace it for the new one
if image_key in self.figures:
self.viz.close(win=self.figures[image_key], env=self.env)
del self.figures[image_key]
# plot the image
self.figures[image_key] = self.viz.images(image,
env=self.env,
opts=dict(title=caption))
def close_client(self):
self.viz.use_socket = False
class Visualizer:
def __init__(self, cfg: DictConfig):
self.cfg = cfg.visdom
visdom_command = 'screen -S visdom_{} -d -m bash -c "python -m visdom.server -port {}"'.format(
self.cfg.port, self.cfg.port
)
os.mkdir("visdom")
os.system(visdom_command)
time.sleep(2)
# self.env = self.cfg.default_env_name # TODO: What is this
self.vis = Visdom(
port=self.cfg.port,
log_to_filename=os.path.join("visdom", self.cfg.log_to_filename),
offline=self.cfg.offline,
)
(self.x_min, self.x_max), (self.y_min, self.y_max) = (
(self.cfg.x_min, self.cfg.x_max),
(self.cfg.y_min, self.cfg.y_max),
)
self.counter = 0
self.plots = {}
def img_result(
self,
img_list: List[np.ndarray],
nrow: int,
caption: str = "view",
title: str = "title",
win: int = 1,
env: str = None,
):
self.vis.images(
img_list,
nrow=nrow,
win=win,
opts={"caption": caption, "title": title},
env=env,
)
def plot_img_255(
self, img_key: str, img: np.ndarray, env: str = None, opts: Dict = None,
):
"""Visdom plot a single image (channels-first CxHxW)"""
self.vis.image(img, win=img_key, opts=opts, env=env)
def plot_matplotlib(self, fig, caption="view", title="title", win=1, env=None):
self.vis.matplot(
fig,
win=win,
opts={"caption": caption, "title": title, "resizable": True},
env=env,
)
def plot_plotly(self, fig, caption="view", title="title", win=1, env=None):
self.vis.plotlyplot(fig, win=win, env=env)
def plot(
self,
plot_key: str,
split_name: str,
x: int,
y: int,
env: str = None,
opts: Dict = None,
):
"""Visdom plot line data"""
if plot_key not in self.plots:
self.plots[plot_key] = self.vis.line(
X=np.array([x, x]),
Y=np.array([y, y]),
env=env,
opts={**opts, "legend": [split_name]},
)
else:
self.vis.line(
X=np.array([x]),
Y=np.array([y]),
env=env,
win=self.plots[plot_key],
name=split_name,
update="append",
)
def align_execute_mean_two():
temp = []
f = open(two_task_time_csv, 'r')
csv_r = csv.reader(f)
for p in csv_r:
temp.append(p)
f.close()
length = len(temp)
print('length of rough pearson correlation list: %d' % length)
# Build the dictionary for storing the complemented correlation
str_i = '-'
inference_exe_join = []
for i in inference_exe:
# inference_exe_join.append(str_i.join(i))
inference_exe_join.append(i[2] + '_' + i[3] + '_' + i[1])
dict_two = {}
# em_exe = poly_exe + inference_exe_join
em_exe = poly_exe + inference_exe_join + example_exe
for p in em_exe:
dict_two[p] = []
# search each kind of benchmark and complement all the combinations with it
for i in range(length):
obj = temp[i]
print('*** %s *** & *** %s ***' % (obj[0], obj[1]))
group = temp[i]
if group[0] != group[1]:
x1, x2 = dict_two[group[0]], dict_two[group[1]]
x1.append(group)
x2.append(
[group[1], group[0], group[4], group[5], group[2], group[3]])
dict_two[group[0]], dict_two[group[1]] = x1, x2
else:
x1 = dict_two[group[0]]
x1.append(group)
dict_two[group[0]] = x1
# write the dictionary to csv file
f = open(align_two_task_time_csv, 'a+')
csv_w = csv.writer(f)
for p in dict_two.keys():
list_cor = dict_two[p]
for i in list_cor:
# print i
csv_w.writerow(i)
f.close()
# write the dictionary to csv file
f = open(align_two_task_time_gap_csv, 'a+')
csv_w = csv.writer(f)
x = []
y1 = []
y2 = []
for p in dict_two.keys():
list_cor = dict_two[p]
for i in list_cor:
gap_1 = float(i[3]) - float(i[2])
gap_2 = float(i[5]) - float(i[4])
csv_w.writerow([i[0], i[1], gap_1, gap_2])
x.append(i[0] + '&' + i[1])
y1.append(gap_1)
y2.append(gap_2)
f.close()
# plot the diagram
viz = Visdom()
assert viz.check_connection()
# ----------------------------------------------------------------------------------------------
# plot all the points via matplotlib.pyplot
# ----------------------------------------------------------------------------------------------
try:
import matplotlib.pyplot as plt
plt.switch_backend('agg')
# plt.plot(y1[6500:7000])
plt.plot(y2)
plt.ylabel('numerical numbers')
win = viz.matplot(plt)
except BaseException as err:
print('Skipped matplotlib example')
print('Error message: ', err)
assert viz.win_exists(win), 'Created window marked as not existing'
class Dashboard(object):
def __init__(self, opt, vis_opt, logpath, vis=True, port=8097, mode='realtime'):
# Get relevant options
exp_name = opt['logs']['exp_name']
if 'env' in opt:
self.game = opt['env']['env-name']
self.alg_name = opt['alg']['algo']
self.num_steps = opt['optim']['num_frames']
trial = opt['trial']
self.unique_name = exp_name + ' ' + self.game + ' ' + self.alg_name + ' trial %d' % trial
else:
self.unique_name = exp_name + ' supervised'
# Get strings for different monitors
self.episode_monitor_str = vis_opt['episode_monitor_str'] if 'episode_monitor_str' in vis_opt else None
self.step_monitor_str = vis_opt['step_monitor_str'] if 'step_monitor_str' in vis_opt else None
self.alg_monitor_str = vis_opt['alg_monitor_str'] if 'alg_monitor_str' in vis_opt else None
self.ll_alg_monitor_str = vis_opt['ll_alg_monitor_str'] if 'll_alg_monitor_str' in vis_opt else None
# TODO - do this later if we care
self.logpath = logpath
# Get keys we want to plot
self.plot_keys = vis_opt['plot_keys']
# Timing stuff
# Keep track of last time we updated info for a specific episode
# This lets us
self.time_since_update = {}
# Set up Visdom server and windows
if vis:
self.viz = Visdom(port=port)
self.wins = {}
# Save opt for special cases
self.opt = opt
# Save the current windows to file
def dump_plots(self):
# TODO unsure of usage, but this seems good to add
return
# Do some smoothing on our curves
def smooth_curve(self, x, y):
# Halfwidth of our smoothing convolution
halfwidth = min(31, int(np.ceil(len(x) / 30)))
k = halfwidth
xsmoo = x[k:-k]
ysmoo = np.convolve(y, np.ones(2 * k + 1), mode='valid') / \
np.convolve(np.ones_like(y), np.ones(2 * k + 1), mode='valid')
downsample = max(int(np.floor(len(xsmoo) / 1e3)), 1)
return xsmoo[::downsample], ysmoo[::downsample]
# Does binning of x and y over the given interval
def fix_point(self, x, y, interval):
np.insert(x, 0, 0)
np.insert(y, 0, 0)
fx, fy = [], []
pointer = 0
ninterval = int(max(x) / interval + 1)
for i in range(ninterval):
tmpx = interval * i
while pointer + 1 < len(x) and tmpx > x[pointer + 1]:
pointer += 1
if pointer + 1 < len(x):
alpha = (y[pointer + 1] - y[pointer]) / \
(x[pointer + 1] - x[pointer])
tmpy = y[pointer] + alpha * (tmpx - x[pointer])
fx.append(tmpx)
fy.append(tmpy)
return fx, fy
# TODO - Make a function that outputs our variance plots over multiple trials
def plot_trial_var_values(self):
return
# Given keys and a list of lists, return a dictionary which for each key contains a list
def format_data(self, keys, data):
# Make struct, and init each key with empty list
ret_struct = {}
for key in keys:
ret_struct[key] = []
# Go through our data, and populate the lists
for data_row in data:
if len(keys) != len(data_row):
pdb.set_trace()
# TODO - there is some weird error where there is a missing newline between the header row and first data row
# Figure out how to deal with this elegantly, or just restart the job
#assert(len(keys) == len(data_row))
for key, val in zip(keys, data_row):
ret_struct[key].append(val)
# Check length makes sense
for key in keys:
assert(len(ret_struct[key]) == len(data))
return ret_struct
# Preload all the file data
def preload_data(self):
# Episode monitor (one for each thread)
if self.episode_monitor_str:
episode_monitor_data = []
infiles = glob.glob(os.path.join(self.logpath, '*.' + self.episode_monitor_str))
for inf in infiles:
# Read all the lines at once quickly
with open(inf, 'r') as f:
rawlines = f.readlines()
rawlines = [l.rstrip('\n') for l in rawlines]
# Get the header with key names
episode_monitor_keys = rawlines[1].split(',')
time_index = episode_monitor_keys.index('episode_dt')
# Get values line by line
for line in rawlines[2:]:
line_vals = line.split(',')
t_time = float(line_vals[time_index])
tmp = [t_time, line_vals]
episode_monitor_data.append(tmp)
# Resort episode monitor by timestamp
episode_monitor_data = sorted(episode_monitor_data, key=lambda x: x[0])
episode_monitor_data = [x[1] for x in episode_monitor_data]
# Extract the lists for each key
self.episode_monitor_data = self.format_data(episode_monitor_keys, episode_monitor_data)
# Get the x axis (episode_len, but using cumulitive sum)
self.episode_monitor_x = np.cumsum([int(x) for x in self.episode_monitor_data['episode_len']])
self.episode_monitor_x = np.concatenate((np.array([0]), self.episode_monitor_x[:-1]))
# Load the last episode's step monitor into memory (note, only one, even for multithreaded)
if self.step_monitor_str:
step_monitor_data = []
# Read all the lines at once quickly
with open(os.path.join(self.logpath, self.step_monitor_str), 'r') as f:
rawlines = f.readlines()
rawlines = [l.rstrip('\n') for l in rawlines]
# Get the header with key names
step_monitor_keys = rawlines[1].split(',')
# Get values line by line
for line in rawlines[2:]:
line_vals = line.split(',')
step_monitor_data.append(line_vals)
# Extract lists for each key
self.step_monitor_data = self.format_data(step_monitor_keys, step_monitor_data)
# Get the x axis (just the step count)
self.step_monitor_x = [i for i in range(len(step_monitor_data))]
# Load the algorithm monitor data
if self.alg_monitor_str:
alg_monitor_data = []
# Read all the lines at once quickly
with open(os.path.join(self.logpath, self.alg_monitor_str), 'r') as f:
rawlines = f.readlines()
rawlines = [l.rstrip('\n') for l in rawlines]
# Get the header with key names
alg_monitor_keys = rawlines[1].split(',')
# Get values line by line
for line in rawlines[2:]:
line_vals = line.split(',')
alg_monitor_data.append(line_vals)
# Extract lists for each key
self.alg_monitor_data = self.format_data(alg_monitor_keys, alg_monitor_data)
# Get the x axis (just the update count)
self.alg_monitor_x = [i for i in range(len(alg_monitor_data))]
# Load the raw plotting data from source
def load_data(self, data_src, data_type, log_name):
# Load the raw values
if data_src == 'episode_monitor':
if not self.episode_monitor_str:
raise Exception("Episode monitor data not loaded")
# Get the data for this log name from the right data source (should be already loaded)
monitor = self.episode_monitor_data
raw_data_x = self.episode_monitor_x
elif data_src == 'step_monitor':
if not self.step_monitor_str:
raise Exception("Step monitor data not loaded")
monitor = self.step_monitor_data
raw_data_x = self.step_monitor_x
elif data_src == 'alg_monitor':
if not self.alg_monitor_str:
raise Exception("Alg monitor data not loaded")
monitor = self.alg_monitor_data
raw_data_x = self.alg_monitor_x
else:
raise NotImplementedError
# If multiscalar, load from all the compnent keys
if data_type in ['multiscalar', 'special']:
raw_data_y = {}
for key in log_name:
#pdb.set_trace()
if log_name[key] or type(log_name[key]) is str:
assert(key in monitor)
if key in monitor:
raw_data_y[key] = monitor[key]
# Else, just get from monitor
else:
raw_data_y = monitor[log_name]
return raw_data_x, raw_data_y
# Plot a trace of an agents x y movement
def display_movement(self, xypos, plot_struct, thetas=None, yaws=None):
# Do matplot for xy movement
xlabel_name = 'X'
ylabel_name = 'Y'
plt.figure()
x = [dat[0] for dat in xypos]
y = [dat[1] for dat in xypos]
plt.scatter(x, y, marker='*')
plt.xlabel(xlabel_name)
plt.ylabel(ylabel_name)
plt.title(self.unique_name)
# If applicable, draw theta direction
if thetas is not None:
last_theta = None
for t, theta in enumerate(thetas):
if len(thetas) > 100 and t % 10 > 0:
continue
if last_theta is None or np.linalg.norm(last_theta-theta) > 1e-6:
if np.linalg.norm(theta) > 1e-3:
x_offset = x[t]
y_offset = y[t]
plt.arrow(x_offset, y_offset, theta[0], theta[1], head_width=0.05, head_length=0.1, fc='k', ec='k')
last_theta = theta
if yaws is not None:
last_yaw = None
for t, yaw in enumerate(yaws):
if len(yaws) > 100 and t % 10 > 0:
continue
if last_yaw is None or np.linalg.norm(last_yaw-yaw) > 1e-6:
if np.linalg.norm(yaw) > 1e-3:
x_offset = x[t]
y_offset = y[t]
plt.arrow(x_offset, y_offset, yaw[0], yaw[1], head_width=0.05, head_length=0.1, fc='g', ec='g')
last_yaw = yaw
# Update window
# TODO - is there a draw over option maybe?
if plot_struct['window_once']:
win = self.viz.matplot(plt)
elif keyname in self.wins:
self.wins[keyname] = self.viz.matplot(plt, win=self.wins[keyname])
else:
self.wins[keyname] = self.viz.matplot(plt)
plt.close()
# Do simple x, y plot
def simple_plot(self, x, y, keyname, plot_struct):
# Check we've reached bin_size
bin_size = plot_struct['bin_size']
if len(x) < bin_size:
return
# Time subset (if applicable)
if 'time_start' in plot_struct:
start = plot_struct['time_start']
end = plot_struct['time_end']
x = x[start:end+1]
y = y[start:end+1]
# Do smoothing or any other work for x and y
if plot_struct['smooth'] == 1:
x, y = self.smooth_curve(x, y)
elif plot_struct['smooth'] == 2:
y = medfilt(y, kernel_size=9)
if bin_size > 1:
x, y = self.fix_point(x, y, bin_size)
# Do matplot
if plot_struct['data_src'] == 'episode_monitor':
xlabel_name = 'Number of Timesteps'
elif plot_struct['data_src'] == 'step_monitor':
xlabel_name = 'Number of Steps'
elif plot_struct['data_src'] == 'alg_monitor':
xlabel_name = 'Number of Updates'
else:
raise NotImplementedError
plt.figure()
plt.plot(x, y)
plt.xlabel(xlabel_name)
plt.ylabel(keyname)
plt.title(self.unique_name)
plt.grid(True)
# Update window
if keyname in self.wins:
self.wins[keyname] = self.viz.matplot(plt, win=self.wins[keyname])
else:
self.wins[keyname] = self.viz.matplot(plt)
plt.close()
# Plot multiple y values
def multi_plot(self, x, y_dict, keyname, plot_struct):
# Check we've reached bin_size
bin_size = plot_struct['bin_size']
if len(x) < bin_size:
return
# Convert values to float
for k in y_dict:
y_dict[k] = [float(val.replace('\x00','')) for val in y_dict[k]]
# Do smoothing or any other work for x and y
orig_x = x
if plot_struct['smooth'] == 1:
for key in y_dict:
y_orig = y_dict[key]
x = np.copy(orig_x)
x, y = self.smooth_curve(x, y_orig)
y_dict[key] = y
elif plot_struct['smooth'] == 2:
for key in y_dict:
y_dict[key] = medfilt(y_dict[key], kernel_size=9)
orig_x = x
for key in y_dict:
x = np.copy(orig_x)
y_orig = y_dict[key]
x, y = self.fix_point(x, y_orig, bin_size)
y_dict[key] = y
# Do matplot
if plot_struct['data_src'] == 'episode_monitor':
xlabel_name = 'Number of Timesteps'
elif plot_struct['data_src'] == 'step_monitor':
xlabel_name = 'Number of Steps'
elif plot_struct['data_src'] == 'alg_monitor':
xlabel_name = 'Number of Updates'
else:
raise NotImplementedError
fig = plt.figure()
# Put each y value on
legend_handles = []
legend_labels = []
for key in y_dict: # Kind of hack. First few values often plot garbage
plt.plot(x, y_dict[key], label=key)
plt.legend()
plt.xlabel(xlabel_name)
plt.ylabel(keyname)
plt.title(self.unique_name)
plt.grid(True)
# Update window
if keyname in self.wins:
self.wins[keyname] = self.viz.matplot(plt, win=self.wins[keyname])
else:
self.wins[keyname] = self.viz.matplot(plt)
plt.close()
# Do simple value display
def display_simple_value(self, value_str, keyname, plot_struct):
display_text = self.unique_name + '\n' + keyname + ': ' + value_str
if keyname in self.wins:
self.wins[keyname] = self.viz.text(display_text, win=self.wins[keyname])
else:
self.wins[keyname] = self.viz.text(display_text)
# Update visdom plot
def update_display(self, raw_data_x, raw_data_y, keyname, plot_struct):
# First split on data type
# If scalar, pretty simple plotting
data_type = plot_struct['data_type']
# If scalar, call simple plot
if data_type == 'scalar':
raw_data_y = [float(y) for y in raw_data_y]
self.simple_plot(raw_data_x, raw_data_y, keyname, plot_struct)
# If multiscalar, do more complex plot with legend
elif data_type == 'multiscalar':
self.multi_plot(raw_data_x, raw_data_y, keyname, plot_struct)
# If array, either do elementwise or get norm
elif data_type == 'array':
# Clean up and convert arrays
conv_array = lambda s: np.fromstring(s.replace('[', '').replace(']', '').replace('"', ''), dtype=float, sep=' ')
raw_data_y = [conv_array(y) for y in raw_data_y]
display_type = plot_struct['display_type']
# If norm, take np norm and do simple plot
if display_type == 'norm':
raw_data_y = [np.linalg.norm(y) for y in raw_data_y]
self.simple_plot(raw_data_x, raw_data_y, keyname, plot_struct)
# If elementwise, plot for each element
elif display_type == 'elementwise':
for i in range(len(raw_data_y[0])):
new_key = keyname + '[%d]' % i
data_y = [y[i] for y in raw_data_y]
self.simple_plot(raw_data_x, data_y, new_key, plot_struct)
# Same as above, but only show subset of indices
elif display_type == 'elementwise_subset':
# Display for [start_ind, end_ind)
start_ind = plot_struct['start_ind']
end_ind = plot_struct['end_ind']
for i in range(start_ind, end_ind):
new_key = keyname + '[%d]' % i
data_y = [y[i] for y in raw_data_y]
self.simple_plot(raw_data_x, data_y, new_key, plot_struct)
else:
raise NotImplementedError
# If single value, display as text
elif data_type == 'single_value':
dat = raw_data_y[0]
self.display_simple_value(dat, keyname, plot_struct)
elif data_type == 'special':
if keyname == 'theta_xy_plot':
# Get xy and get theta
conv_array = lambda s: np.fromstring(s.replace('[', '').replace(']', '').replace('"', ''), dtype=float, sep=' ')
state_raw = [conv_array(y) for y in raw_data_y['state']]
obs_raw = [conv_array(y) for y in raw_data_y['obs']]
xypos = [y[:2] for y in state_raw]
quats = [y[3:7] for y in state_raw]
yaws = []
for q in quats:
_, _, yaw = geom_utils.quaternion_to_euler_angle(q)
yaws.append(yaw)
if 'theta_sz' in self.opt['env']:
theta_sz = self.opt['env']['theta_sz']
if self.opt['env']['add_timestep']:
start_ind = -(theta_sz+1)
end_ind = -2
else:
start_ind = -theta_sz
end_ind = -1
theta_space_mode = self.opt['env']['theta_space_mode']
yaws_2d = None
if theta_space_mode in ['arbitrary', 'arbitrary_stop']:
thetas = [np.array([y[start_ind], y[end_ind]]) for y in obs_raw]
assert(abs(np.linalg.norm(thetas[0]) - 1) < 1e-6)
elif theta_space_mode in ['simple_four']:
thetas = []
yaws_2d = []
for obs, yaw in zip(obs_raw, yaws):
if obs[start_ind] == 1:
theta = np.array([1, 0])
elif obs[start_ind+1] == 1:
theta = np.array([-1, 0])
elif obs[start_ind+2] == 1:
theta = np.array([0, 1])
elif obs[start_ind+3] == 1:
theta = np.array([0, -1])
else:
raise Exception("Something is wrong")
# Turn with yaw
#pdb.set_trace()
theta = geom_utils.convert_vector_to_egocentric(-yaw, theta)
thetas.append(theta)
yaws_2d.append(np.array([math.cos(yaw), math.sin(yaw)]))
elif theta_space_mode in ['simple_eight']:
thetas = []
yaws_2d = []
for obs, yaw in zip(obs_raw, yaws):
if obs[start_ind] == 1:
theta = np.array([1, 0])
elif obs[start_ind+1] == 1:
theta = np.array([-1, 0])
elif obs[start_ind+2] == 1:
theta = np.array([0, 1])
elif obs[start_ind+3] == 1:
theta = np.array([0, -1])
elif obs[start_ind+4] == 1:
theta = np.array([math.sqrt(0.5), math.sqrt(0.5)])
elif obs[start_ind+5] == 1:
theta = np.array([-math.sqrt(0.5), math.sqrt(0.5)])
elif obs[start_ind+6] == 1:
theta = np.array([-math.sqrt(0.5), -math.sqrt(0.5)])
elif obs[start_ind+7] == 1:
theta = np.array([math.sqrt(0.5), -math.sqrt(0.5)])
else:
raise Exception("Something is wrong")
# Turn with yaw
#pdb.set_trace()
theta = geom_utils.convert_vector_to_egocentric(-yaw, theta)
thetas.append(theta)
yaws_2d = None
elif theta_space_mode in ['k_theta']:
yaws_2d = []
for yaw in yaws:
yaws_2d.append(np.array([math.cos(yaw), math.sin(yaw)]))
thetas = None
else:
thetas = None
else:
thetas = None
yaws_2d = None
#pdb.set_trace()
self.display_movement(xypos, plot_struct, thetas, yaws_2d)
else:
raise NotImplementedError
# Plot all values in visdom
def visdom_plot(self):
# Preload the monitor data
self.preload_data()
# For each plot key
for key in self.plot_keys:
# Get values
plot_struct = self.plot_keys[key]
data_src = plot_struct['data_src']
data_type = plot_struct['data_type']
log_name = plot_struct['log_name']
# Check if we want to wait until a certain delay
if 'update_delay' in plot_struct:
update_delay = plot_struct['update_delay']
# Skip this if it hasn't been long enough since last update
if key in self.time_since_update and time.time() - self.time_since_update[key] < update_delay:
continue
else:
# Update time delay value
self.time_since_update[key] = time.time()
# Get the data from the correct source
data_x, data_y = self.load_data(data_src, data_type, log_name)
# Display the data
self.update_display(data_x, data_y, key, plot_struct)
class GPU_logger():
def __init__(self,
host_names,
time_interval,
time_range,
show_summary=True,
show_monitor=True,
show_cpu_mem=False,
port=8098,
env_name='main',
verbose=False):
self.viz = Visdom(port=port)
self.env = env_name
self.hosts = host_names
self.sleep = time_interval
self.range = time_range
self.show_summary = show_summary
self.show_monitor = show_monitor
self.show_cpu_mem = show_cpu_mem
self.verbose = verbose
self.win = {}
self.num_gpus = 4
self.max_length = self.range // self.sleep
self.tick_ds = self.max_length // 6
self.time_indices = deque()
self.memory_queue = np.zeros([len(host_names), self.num_gpus, 0])
self.usages_queue = np.zeros([len(host_names), self.num_gpus, 0])
self.table_W = 780
self.table_H = 470
self.restore(PWD)
def reset(self):
self.viz.close(win=None, env=self.env)
def save(self, output_dir):
if not isdir(output_dir):
os.makedirs(output_dir)
with open(join(output_dir, 'cached_data'), 'wb') as f:
pickle.dump(
{
'time': self.time_indices,
'memory': self.memory_queue,
'usages': self.usages_queue
}, f, pickle.HIGHEST_PROTOCOL)
def restore(self, output_dir):
if isfile(join(output_dir, 'cached_data')):
with open(join(output_dir, 'cached_data'), 'rb') as f:
cache = pickle.load(f)
self.time_indices = cache['time']
self.memory_queue = cache['memory']
self.usages_queue = cache['usages']
def record(self):
if self.show_summary:
gpu_tabel_opts = {
'title': 'README',
'resize': False,
'width': self.table_W,
'height': self.table_H
}
col_W = self.table_W / len(self.hosts)
self.win['summary'] = self.viz.text(README,
env=self.env,
opts=gpu_tabel_opts)
if self.show_cpu_mem:
cpu_tabel_opts = {
'title': 'CPU Memory',
'resize': True,
'width': self.table_W,
'height': 150
}
col_W_c = self.table_W / len(self.hosts)
self.win['cpu_mem'] = self.viz.text('',
env=self.env,
opts=cpu_tabel_opts)
while True:
if len(self.time_indices) >= self.max_length:
self.time_indices.popleft()
self.time_indices.append(time.strftime("%H:%M"))
if isfile(join(HOME, '.tcshrc')): # hotfix to handle my zsh
os.rename(join(HOME, '.tcshrc'), join(HOME, '.tcshrc.bak'))
#gpu_memory = ssh_host(self.hosts, mode='memory')
#gpu_usage = ssh_host(self.hosts, mode='gpu')
gpu_memory, gpu_usage, cpu_memory = ssh_host(self.hosts,
mode='all')
if self.verbose:
print('gpu memory:', gpu_memory)
print('gpu usage:', gpu_usage)
print('cpu memory:', cpu_memory)
if isfile(join(HOME, '.tcshrc.bak')):
os.rename(join(HOME, '.tcshrc.bak'), join(HOME, '.tcshrc'))
if self.memory_queue.shape[-1] < self.max_length:
self.memory_queue = np.append(
self.memory_queue,
np.reshape(gpu_memory,
[len(self.hosts), self.num_gpus, 1]), -1)
self.usages_queue = np.append(
self.usages_queue,
np.reshape(gpu_usage, [len(self.hosts), self.num_gpus, 1]),
-1)
else:
self.memory_queue = np.append(
np.delete(self.memory_queue, 0, -1),
np.reshape(gpu_memory,
[len(self.hosts), self.num_gpus, 1]), -1)
self.usages_queue = np.append(
np.delete(self.usages_queue, 0, -1),
np.reshape(gpu_usage, [len(self.hosts), self.num_gpus, 1]),
-1)
gpu_status_table, cpu_status_table = [], []
gpu_status_table.append(
get_tablerow('', ['GPU:01', 'GPU:02', 'GPU:03', 'GPU:04']))
cpu_status_table.append(
get_tablerow('', ['Total:', 'Used:', 'Available:']))
for k, host in enumerate(self.hosts):
fig = plt.figure(figsize=(20, 2.7))
fig.suptitle(host, size=24, fontweight="bold", y=1.)
gpu_status_row = []
for idx in range(4):
memory_, usage_ = self.memory_queue[
k, idx, :], self.usages_queue[k, idx, :]
nonzero_idx = np.where(memory_ > 1)[0]
m_memory_ = np.mean(memory_[nonzero_idx[0]:]) if len(
nonzero_idx) > 0 else np.mean(memory_)
m_usage_ = np.mean(usage_[nonzero_idx]) if len(
nonzero_idx) > 0 else np.mean(usage_)
if m_memory_ == -1: # I've set empty gpu to -1
gpu_status_row.append(' ')
continue
if np.any(
memory_[-5:] > 1
): #if last 5 checkpoint is empty -> not used now
status = get_gpu_health(m_memory_, m_usage_)
else:
status = status_kw[0]
box_prop = dict(facecolor=color_map[status],
edgecolor='none',
pad=2,
alpha=0.6)
gpu_status_row.append(status)
plt.subplot(1, 4, idx + 1)
plt.xticks(
np.arange(0, len(self.time_indices), self.tick_ds))
plt.fill_between(list(self.time_indices),
memory_,
color='r',
label="Memory",
alpha=0.4)
plt.fill_between(list(self.time_indices),
usage_,
color='g',
label="Usages",
alpha=0.4)
plt.ylim(-1, 101)
plt.title('GPU-{:02d}'.format(idx), size=15, bbox=box_prop)
plt.xlabel('Average Memory: {:.1f}% Average Usage: {:.1f}%'\
.format(m_memory_, m_usage_), size=14)
plt.tight_layout()
plt.legend()
gpu_status_table.append(
get_tablerow(host.replace('emon', '-'),
gpu_status_row)) #shorten hostname
cpu_status_table.append(
get_tablerow(host.replace('emon', '-'), [
str(cpu_memory[k][0]) + 'G',
str(cpu_memory[k][1]) + 'G',
str(cpu_memory[k][-1]) + 'G'
]))
if self.show_monitor:
opts = {
'title': host,
'resizable': True,
'height': 190,
'width': 1400
}
if host not in self.win.keys():
self.win[host] = self.viz.matplot(plt,
env=self.env,
opts=opts)
else:
self.viz.matplot(plt,
win=self.win[host],
env=self.env,
opts=opts)
plt.close()
TIMESTAMP = f"Updated at {time.strftime('%Y/%m/%d-%H:%M:%S')}<br>"
if 'summary' in self.win.keys():
status_table_g = [list(t) for t in zip(*gpu_status_table)]
table_ = HTML.table(status_table_g,
border='3',
cellpadding=5,
col_width=[col_W] * (len(self.hosts) + 1),
col_align=['center'] *
(len(self.hosts) + 1))
self.viz.text(README + table_ + TIMESTAMP + ACKNOWLEDGE,
env=self.env,
win=self.win['summary'],
opts=gpu_tabel_opts)
if 'cpu_mem' in self.win.keys():
status_table_c = [list(t) for t in zip(*cpu_status_table)]
table_c = HTML.table(
status_table_c,
border='2',
cellpadding=2,
col_width=[col_W_c] * (len(self.hosts) + 1),
col_align=['center'] * (len(self.hosts) + 1))
self.viz.text(table_c + TIMESTAMP,
env=self.env,
win=self.win['cpu_mem'],
opts=cpu_tabel_opts)
time.sleep(self.sleep)
class sac_ES:
def __init__(self,
env_fn,
seed=0,
steps_per_epoch=5000,
epochs=100,
gamma=0.7,
polyak=0.995,
lr=3e-4,
alpha=0.2,
batch_size=1024,
start_steps=1,
max_ep_len=1000,
memory_size=int(1e+4),
max_kl=0.01,
experiment_name='none',
use_adv=False,
use_half_dete=False,
use_sample_dist=False,
Test_agent=False):
tf.set_random_seed(seed)
np.random.seed(seed)
self._success_obs_item = len(os.listdir('result/success_obs/'))
self.env = env_fn()
self.ac_per_state = 30
self.memory_size = memory_size
self.batch_size = batch_size
self.gamma = gamma
self.alpha = alpha
self.lr = lr
self.polyak = polyak
self.start_steps = start_steps
self.max_ep_len = max_ep_len
self.epochs = epochs
self.steps_per_epoch = steps_per_epoch
self.max_kl = max_kl
self.use_adv = use_adv
self.entropy_too_low = False
self.win = None
if Test_agent:
from visdom import Visdom
self.viz = Visdom()
assert self.viz.check_connection()
self.win = self.viz.matplot(plt)
self.act_dim = self.env.action_space.shape[0]
self.replay_buffer = ReplayBuffer(act_dim=self.act_dim,
size=self.memory_size)
self.use_half_dete = use_half_dete
self.use_sample_dist = use_sample_dist
# Build computer graph
self.graph = tf.Graph()
with tf.device('/gpu:0'), self.graph.as_default():
self.sess = tf.Session(config=config, graph=self.graph)
self.__make_model()
# Summary
self.experiment_name = experiment_name
self.writer = tf.compat.v1.summary.FileWriter("logs/" +
self.experiment_name)
self.ep_ret_ph = tf.placeholder(tf.float32,
shape=(),
name="ep_ret_ph")
self.ep_len_ph = tf.placeholder(tf.float32,
shape=(),
name="ep_len_ph")
self.test_summary = tf.compat.v1.summary.merge([
tf.compat.v1.summary.scalar('EP_ret',
self.ep_ret_ph,
family='test'),
tf.compat.v1.summary.scalar('EP_len',
self.ep_len_ph,
family='test')
])
self.sess.run(tf.global_variables_initializer())
self.sess.run(self.V_target_init)
_ = os.system("clear")
Count_Variables()
print('Trainable_variables:')
for v in tf.compat.v1.trainable_variables():
print('{}\t {}'.format(v.name, str(v.shape)))
def __make_model(self):
self.x1_ph = tf.placeholder(tf.float32, [None, 128, 128, 3],
name='o1_ph')
self.x2_ph = tf.placeholder(tf.float32, [None, 128, 128, 3],
name='o2_ph')
self.a_ph, self.r_ph, self.d_ph, self.q_ij_ph = core.placeholders(
self.act_dim, None, None, self.ac_per_state)
with tf.variable_scope('main'):
self.dist, self.sample_dist, self.mean, self.pi, self.q1, self.q2, self.q1_pi, self.q2_pi, self.v, self.std = core.actor_critic(
self.x1_ph,
self.a_ph,
action_space=self.env.action_space,
hidden_size=(64, 32))
vf_mlp = lambda x_: tf.squeeze(
core.mlp(core._cnn(x_), [64, 32, 1], tf.nn.relu, None), axis=1)
with tf.variable_scope('target'):
self.v_target = vf_mlp(self.x2_ph)
# Policy Evaluate
with tf.variable_scope('Policy_Evaluate'):
# Min Double-Q:
min_q_pi = tf.minimum(self.q1_pi, self.q2_pi)
# Targets for Q and V regression
q_target = tf.stop_gradient(self.r_ph + self.gamma *
(1 - self.d_ph) * self.v_target)
logp_pi = tf.reduce_sum(self.dist.log_prob(self.pi),
axis=-1,
keep_dims=True) # sum the each dim_of_A
v_target = tf.stop_gradient(min_q_pi - self.alpha * logp_pi)
q1_loss = 0.5 * tf.reduce_mean((q_target - self.q1)**2)
q2_loss = 0.5 * tf.reduce_mean((q_target - self.q2)**2)
v_loss = 0.5 * tf.reduce_mean((v_target - self.v)**2)
value_loss = q1_loss + q2_loss + v_loss
# set_target_update
self.V_target_init = tf.group([
tf.assign(v_targ, v_main)
for v_main, v_targ in zip(get_vars('main/v'), get_vars('target'))
])
# Policy Improvement
with tf.variable_scope('Policy_Improvement'):
# (samples_per_state, batch, self.act_dim) --> (-1, self.act_dim)
self.a_flatten_s_dist = self._samp_a(self.sample_dist)
self.a_flatten_dist = self._samp_a(self.dist)
if self.use_adv:
q = self.q1 - self.v_target
else:
q = self.q1
q_ij = tf.stop_gradient(
tf.nn.softmax(tf.reshape(q, [self.ac_per_state, -1]), axis=0))
logp = tf.reduce_sum(self.dist.log_prob(self.a_ph),
axis=-1,
keep_dims=True) # sum the each dim_of_A
logp_ij = tf.reshape(logp, [self.ac_per_state, -1])
self.entropy = tf.reduce_mean(self.dist.entropy())
likelihood_term = q_ij * logp_ij
# Value train op
value_optimizer = tf.compat.v1.train.AdamOptimizer(
learning_rate=self.lr)
value_params = get_vars('main/q') + get_vars('main/v')
self.train_value_op = value_optimizer.minimize(value_loss,
var_list=value_params)
# value_learn -> update_target_value
self.soft_update = tf.group([
tf.assign(v_targ,
self.polyak * v_targ + (1 - self.polyak) * v_main)
for v_main, v_targ in zip(get_vars('main/v'), get_vars('target'))
])
# Maximum likelihood
pi_loss = -tf.reduce_sum(likelihood_term)
pi_optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=self.lr)
self.train_pi_op = pi_optimizer.minimize(pi_loss,
var_list=get_vars('main/pi'))
var_list = tf.global_variables()
self.saver = tf.compat.v1.train.Saver(var_list=var_list, max_to_keep=1)
assert tf.get_collection(tf.GraphKeys.UPDATE_OPS) == [
] # 确保没有 batch norml
# summary
self.a_summary = tf.compat.v1.summary.merge(
[tf.compat.v1.summary.scalar('pi_loss', pi_loss, family='actor')])
self.c_summary = tf.compat.v1.summary.merge([
tf.compat.v1.summary.scalar('value_loss',
value_loss,
family='critic'),
tf.compat.v1.summary.scalar('q1_loss', q1_loss, family='critic'),
tf.compat.v1.summary.scalar('q2_loss', q2_loss, family='critic'),
tf.compat.v1.summary.scalar('v_loss', v_loss, family='critic'),
tf.compat.v1.summary.scalar('pi_entropy',
self.entropy,
family='actor')
])
def _samp_a(self, dist):
# (samples_per_state, batch, self.act_dim) --> (-1, self.act_dim)
a = tf.reshape(dist.sample(self.ac_per_state), [-1, self.act_dim]),
return tf.clip_by_value(a, self.env.action_space.low[0],
self.env.action_space.high[0])
def Save(self):
path = "model/" + self.experiment_name + "/model.ckpt"
# print("Save model to the: '{}'".format(path))
self.saver.save(self.sess, save_path=path)
def load(self):
path = "model/" + self.experiment_name + "/model.ckpt"
print("\nLoad model From the: '{}'\n".format(path))
self.saver.restore(self.sess, save_path=path)
def _choose_action(self, o, deterministic=False):
act_op = self.mean if deterministic else self.pi
return self.sess.run(act_op, feed_dict={self.x1_ph:
o[np.newaxis, ]})[0]
def learn(self):
batch = self.replay_buffer.sample_batch(self.batch_size)
a_flatten_opt = self.a_flatten_s_dist #if self.entropy_too_low else self.a_flatten_dist
opt_step_1 = [self.train_value_op, a_flatten_opt, self.c_summary]
feed_step_1 = {
self.x1_ph: batch['obs1'],
self.x2_ph: batch['obs2'],
self.a_ph: batch['acts'],
self.r_ph: batch['rews'],
self.d_ph: batch['done']
}
_, a_Flatten, c_s = self.sess.run(opt_step_1, feed_step_1)
self.writer.add_summary(c_s)
feed_step_2 = {
self.a_ph: np.squeeze(a_Flatten),
self.x1_ph: np.tile(batch['obs1'], [self.ac_per_state, 1, 1, 1]),
self.x2_ph: np.tile(batch['obs2'], [self.ac_per_state, 1, 1, 1])
}
opt_step_2 = [self.entropy, self.train_pi_op, self.a_summary]
# update policy by max_likelihood
entropy, _, a_s = self.sess.run(opt_step_2, feed_step_2)
self.entropy_too_low = True if entropy < -0.5 else False
self.writer.add_summary(a_s)
# update target_net and old pi
self.sess.run(self.soft_update)
def test_agent(self, deterministic=False, n=1):
ep_r = []
ep_l = []
plt.figure(figsize=(9, 5))
for j in range(n):
print('------------------Epoch: %d-------------------' % j)
o, r, d, ep_ret, ep_len, level_plot = self.env.reset(
), 0, False, 0, 0, []
while not (d or (ep_len == self.max_ep_len)):
# Take deterministic actions at test time
a = self._choose_action(o, deterministic=deterministic)
o, r, d, info = self.env.step(a, _step=ep_len)
print("step: {} a: {} r: {}".format(ep_len,
np.around(a, decimals=2),
r))
ep_ret += r
ep_len += 1
level_plot.append(self.env._before)
self.win = R_plot(level_plot, self.viz, self.win)
print('EP_ret: %d, \t EP_Len: %d' % (ep_ret, ep_len))
if info['is_success']:
print('success!')
ep_r.append(ep_ret), ep_l.append(ep_len)
plt.close()
def rollout(self):
total_steps = self.steps_per_epoch * self.epochs
o, r, d, ep_ret, ep_len = self.env.reset(), 0, False, 0, 0
deterministic = False
log_t = []
log_episode_ret = []
for t in tqdm(range(total_steps)):
if t > self.start_steps:
if self.use_half_dete:
deterministic = True if np.random.randn() > 0 else False
a = self._choose_action(o, deterministic)
else:
a = self.env.action_space.sample()
o2, r, d, info = self.env.step(a)
ep_ret += r
ep_len += 1
if info['is_success']:
im = Image.fromarray(o2)
im.save('result/success_obs/' + str(self._success_obs_item) +
'.png')
self._success_obs_item += 1
d = False if ep_len == self.max_ep_len else d
self.replay_buffer.store(o, a, r, o2, d)
o = o2
if d or (ep_len == self.max_ep_len):
if (t + 1) >= 500:
[self.learn() for _ in range(ep_len)] # todo::
ep_s = self.sess.run(self.test_summary, {
self.ep_ret_ph: ep_ret,
self.ep_len_ph: ep_len
})
self.writer.add_summary(ep_s, global_step=t)
log_episode_ret.append(ep_ret)
log_t.append(t)
print("episode return: ", ep_ret)
o, r, d, ep_ret, ep_len = self.env.reset(), 0, False, 0, 0
if (t + 1) % 100 == 0:
self.Save()
np.save('result/' + self.experiment_name + '_reward.npy',
(log_t, log_episode_ret),
allow_pickle=True)
if use_vis:
viz = Visdom()
assert viz.check_connection()
initial_c = torch.rand(batch_size, 3, img_size, img_size)
initial_s = torch.rand(batch_size, 1, img_size, img_size)
temp = np.linspace(1, 5, 5)
plt.plot(temp)
result_c = viz.images(initial_c,
nrow=batch_size // 2,
opts=dict(caption="cover vs. stego"))
result_s = viz.images(initial_s,
nrow=batch_size // 2,
opts=dict(caption="secret vs. reveal"))
lc_ssim = viz.matplot(plt)
ls_ssim = viz.matplot(plt)
l_dis = viz.matplot(plt)
l_net = viz.matplot(plt)
# loss hyper
# L(c, c0) = α (1-SSIM(c, c0)) +(1-α)(1 MSSIM(c,c0))+ β MSE(c,c0)
# L(s, s0) = α (1-SSIM(s, s0)) +(1-α)(1 MSSIM(s,s0))+ β MSE(s,s0)
# L(c, c0, s, s0) = L(c, c0) + λ L(s, s0)
alpha = 0.7
beta = 0.5
gamma = 0.85
lambda0 = 0.002
# optimizer hyper for adam
def visom_show(plt=None):
viz = Visdom(env='main')
viz.matplot(plt)
elif event['key'] == 'Backspace':
curr_txt = curr_txt[:-1]
elif event['key'] == 'Delete':
curr_txt = txt
elif len(event['key']) == 1:
curr_txt += event['key']
viz.text(curr_txt, win=callback_text_window)
viz.register_event_handler(type_callback, callback_text_window)
# matplotlib demo:
try:
import matplotlib.pyplot as plt
plt.plot([1, 23, 2, 4])
plt.ylabel('some numbers')
viz.matplot(plt)
except BaseException as err:
print('Skipped matplotlib example')
print('Error message: ', err)
# 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
video_url = 'http://media.w3.org/2010/05/sintel/trailer.ogv'
# linux
