),
)
# Assure that the stacked area plot isn't giant
viz.update_window_opts(
win=win,
opts=dict(
width=300,
height=300,
),
)
# boxplot
X = np.random.rand(100, 2)
X[:, 1] += 2
viz.boxplot(X=X, opts=dict(legend=['Men', 'Women']))
# stemplot
Y = np.linspace(0, 2 * math.pi, 70)
X = np.column_stack((np.sin(Y), np.cos(Y)))
viz.stem(X=X, Y=Y, opts=dict(legend=['Sine', 'Cosine']))
# quiver plot
X = np.arange(0, 2.1, .2)
Y = np.arange(0, 2.1, .2)
X = np.broadcast_to(np.expand_dims(X, axis=1), (len(X), len(X)))
Y = np.broadcast_to(np.expand_dims(Y, axis=0), (len(Y), len(Y)))
U = np.multiply(np.cos(X), Y)
V = np.multiply(np.sin(X), Y)
viz.quiver(
X=U,
class Visualizer(threading.Thread):
def __init__(self, env='Visualizer'):
super(Visualizer, self).__init__()
self.name = env
self.vis = Visdom(env=env)
self.taskList = queue.Queue()
self.fStop = False
self.funcList = dict(result=self.dispCurrentResults,
tTrain=self.dispTrainTime,
tIter=self.dispIterTime,
text=self.infoText)
self.plotData = dict(loss=dict(
x=[],
y=[],
))
self.lossData = {'X': [], 'loss': []}
self.trainTime = {'X': [], 'Time': []}
self.iterTime = {'iter': [], 'time': []}
self.epochIterTime = [[]]
# self.boxEpoch = 0
# self.boxIter = 0
# |visuals|: dictionary of images to display or save
# vis.task(msg=dict(
# func='result',
# data=dict(
# epoch=,
# data=,
# errors=
# )
# ))
def dispCurrentResults(self, msg):
epoch = msg['epoch']
data = msg['data']
errors = msg['error']
for item in data.keys():
# print(data[item].shape)
self.vis.image(data[item], win=item, opts=dict(title=item))
self.lossData['X'].append(epoch)
self.lossData['loss'].append(errors['loss'])
self.vis.line(X=np.array(self.lossData['X']),
Y=np.array(self.lossData['loss']),
opts={
'title': 'Loss',
'xlabel': 'Epoch',
'ylabel': 'Loss'
},
win='loss')
# vis.task(msg=dict(
# func='tTrain',
# data=dict(
# epoch=,
# time=
# )
# ))
def dispTrainTime(self, msg):
epoch = msg['epoch']
time = msg['time']
self.trainTime['X'].append(epoch)
self.trainTime['Time'].append(time)
self.vis.line(X=np.array(self.trainTime['X']),
Y=np.array(self.trainTime['Time']),
opts={
'title': 'Training Time',
'xlabel': 'Epoch',
'ylabel': 'Time'
},
win='trainTime')
# vis.task(msg=dict(
# func='tIter',
# data=dict(
# iter=,
# time=
# )
# ))
# def addIterTime(self, msg):
def dispIterTime(self, msg):
epoch = msg['epoch']
time = msg['time']
if len(self.epochIterTime) <= epoch:
self.epochIterTime.append([time])
else:
self.epochIterTime[epoch].append(time)
if epoch != 0:
self.epochIterTime[0][0] = self.epochIterTime[0][1]
self.vis.boxplot(X=np.array(
self.epochIterTime[:epoch]).transpose(),
opts={
'title': 'Iter Time per Epoch',
'xlabel': 'Epoch',
'ylabel': 'Time'
},
win='iterTime')
def task(self, msg):
self.taskList.put(msg)
def run(self):
while not self.fStop:
if not self.taskList.empty():
msg = self.taskList.get()
self.funcList[msg['func']](msg['data'])
def stop(self):
self.fStop = False
# vis.task(msg=dict(
# func='text',
# data=dict(
# text=[],
# title=, # 可选
# win= # 可选
# )
# ))
def infoText(self, msg):
text = msg['text']
try:
title = msg['title']
except:
title = "info"
try:
win = msg['win']
except:
win = title
text_ = ""
if win is None:
win = title
if isinstance(text, str):
text_ = text
elif isinstance(text, list):
text_ = ""
for s in text:
text_ += s + "<br>"
self.vis.text(text_, win=win, opts=dict(title=title, font='Calibri'))
xlabel='Time',
ylabel='Volume',
ytype='log',
title='Stacked area plot',
marginleft=30,
marginright=30,
marginbottom=80,
margintop=30,
),
)
# boxplot
X = np.random.rand(100, 2)
X[:, 1] += 2
viz.boxplot(
X=X,
opts=dict(legend=['Men', 'Women'])
)
# stemplot
Y = np.linspace(0, 2 * math.pi, 70)
X = np.column_stack((np.sin(Y), np.cos(Y)))
viz.stem(
X=X,
Y=Y,
opts=dict(legend=['Sine', 'Cosine'])
)
# pie chart
X = np.asarray([19, 26, 55])
viz.pie(
X=X,
if isinstance(text, str):
text_ = text
elif isinstance(text, list):
text_ = ""
for s in text:
text_ += s + "<br>"
self.vis.text(text_, win=win, opts=dict(title=title, font='Calibri'))
if __name__ == "__main__":
vis = Visdom(env="boxPlotTest")
# uneven = np.array([[1,2,3,4,5,6,7,8,9,10],[1,2,3,4,5]])
# vis.boxplot(
# uneven.transpose(),
# win="unEven"
# )
even = np.array([[1, 1], [2, 2], [3, 3], [4, 4], [5], [6], [7], [8]])
vis.boxplot(
even,
win="even1",
)
# vis.boxplot(
# even.transpose(),
# win="even2"
# )
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)]
})