Related
I've generated a network figure using vedo library and I'm trying to add this as an inset to a figure generated in matplotlib
import networkx as nx
import matplotlib.pyplot as plt
from vedo import *
from matplotlib.offsetbox import OffsetImage, AnnotationBbox
G = nx.gnm_random_graph(n=10, m=15, seed=1)
nxpos = nx.spring_layout(G, dim=3, seed=1)
nxpts = [nxpos[pt] for pt in sorted(nxpos)]
nx_lines = [(nxpts[i], nxpts[j]) for i, j in G.edges()]
pts = Points(nxpts, r=12)
edg = Lines(nx_lines).lw(2)
# node values
values = [[1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[30, 80, 10, 79, 70, 60, 75, 78, 65, 10],
[1, .30, .10, .79, .70, .60, .75, .78, .65, .90]]
time = [0.0, 0.1, 0.2] # in seconds
vplt = Plotter(N=1)
pts1 = pts.cmap('Blues', values[0])
vplt.show(
pts1, edg,
axes=False,
bg='white',
at=0,
interactive=False,
zoom=1.5
).screenshot("network.png")
ax = plt.subplot(111)
ax.plot(
[1, 2, 3], [1, 2, 3],
'go-',
label='line 1',
linewidth=2
)
arr_img = vplt.screenshot(returnNumpy=True, scale=1)
im = OffsetImage(arr_img, zoom=0.25)
ab = AnnotationBbox(im, (1, 0), xycoords='axes fraction', box_alignment=(1.1, -0.1), frameon=False)
ax.add_artist(ab)
plt.show()
ax.figure.savefig(
"output.svg",
transparent=True,
dpi=600,
bbox_inches="tight"
)
There resolution of the image in the inset is too low. Suggestions on how to add the inset without loss of resolution will be really helpful.
EDIT:
The answer posted below works for adding a 2D network, but I am still looking for ways that will be useful for adding a 3D network in the inset.
I am not familiar with vedo but the general procedure would be to create an inset_axis and plot the image with imshow. However, your code is using networkx which has matplotlib bindings and you can directly do this without vedo
EDIT: code edited for 3d plotting
import networkx as nx
import matplotlib.pyplot as plt
G = nx.gnm_random_graph(n=10, m=15, seed=1)
nxpos = nx.spring_layout(G, dim=3, seed=1)
nxpts = [nxpos[pt] for pt in sorted(nxpos)]
nx_lines = [(nxpts[i], nxpts[j]) for i, j in G.edges()]
# node values
values = [[1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[30, 80, 10, 79, 70, 60, 75, 78, 65, 10],
[1, .30, .10, .79, .70, .60, .75, .78, .65, .90]]
time = [0.0, 0.1, 0.2] # in seconds
fig, ax = plt.subplots()
ax.plot(
[1, 2, 3], [1, 2, 3],
'go-',
label='line 1',
linewidth=2
)
from mpl_toolkits.mplot3d import (Axes3D)
from matplotlib.transforms import Bbox
rect = [.6, 0, .5, .5]
bbox = Bbox.from_bounds(*rect)
inax = fig.add_axes(bbox, projection = '3d')
# inax = add_inset_axes(,
# ax_target = ax,
# fig = fig, projection = '3d')
# inax.axis('off')
# set angle
angle = 25
inax.view_init(10, angle)
# hide axes, make transparent
# inax.set_facecolor('none')
# inax.grid('off')
import numpy as np
# plot 3d
seen = set()
for i, j in G.edges():
x = np.stack((nxpos[i], nxpos[j]))
inax.plot(*x.T, color = 'k')
if i not in seen:
inax.scatter(*x[0], color = 'skyblue')
seen.add(i)
if j not in seen:
inax.scatter(*x[1], color = "skyblue")
seen.add(j)
fig.show()
I want to make an auto calibration system using PyTorch.
I try to deal with a homogeneous transform matrix as weights of neural networks.
I write a code referring to PyTorch tutorials, but my custom parameters are not updated after backward method is called.
When I print a 'grad' attribute of each parameter, it is a None.
My code is below. Is there anything wrong?
Please give any advise to me. Thank you.
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.params = nn.Parameter(torch.rand(6))
self.rx, self.ry, self.rz = self.params[0], self.params[1], self.params[2]
self.tx, self.ty, self.tz = self.params[3], self.params[4], self.params[5]
def forward(self, x):
tr_mat = torch.tensor([[1, 0, 0, self.params[3]],
[0, 1, 0, self.params[4]],
[0, 0, 1, self.params[5]],
[0, 0, 0, 1]], requires_grad=True)
rz_mat = torch.tensor([[torch.cos(self.params[2]), -torch.sin(self.params[2]), 0, 0],
[torch.sin(self.params[2]), torch.cos(self.params[2]), 0, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]], requires_grad=True)
ry_mat = torch.tensor([[torch.cos(self.params[1]), 0, torch.sin(self.params[1]), 0],
[0, 1, 0, 0],
[-torch.sin(self.params[1]), 0, torch.cos(self.params[1]), 0],
[0, 0, 0, 1]], requires_grad=True)
rx_mat = torch.tensor([[1, 0, 0, 0],
[0, torch.cos(self.params[0]), -torch.sin(self.params[0]), 0],
[0, torch.sin(self.params[0]), torch.cos(self.params[0]), 0],
[0, 0, 0, 1]], requires_grad=True)
tf1 = torch.matmul(tr_mat, rz_mat)
tf2 = torch.matmul(tf1, ry_mat)
tf3 = torch.matmul(tf2, rx_mat)
tr_local = torch.tensor([[1, 0, 0, x[0]],
[0, 1, 0, x[1]],
[0, 0, 1, x[2]],
[0, 0, 0, 1]])
tf_output = torch.matmul(tf3, tr_local)
output = tf_output[:3, 3]
return output
def get_loss(self, output):
pass
model = Net()
input_ex = np.array([[-0.01, 0.05, 0.92],
[-0.06, 0.03, 0.94]])
output_ex = np.array([[-0.3, 0.4, 0.09],
[-0.5, 0.2, 0.07]])
print(list(model.parameters()))
optimizer = optim.Adam(model.parameters(), 0.001)
criterion = nn.MSELoss()
for input_np, label_np in zip(input_ex, output_ex):
input_tensor = torch.from_numpy(input_np).float()
label_tensor = torch.from_numpy(label_np).float()
output = model(input_tensor)
optimizer.zero_grad()
loss = criterion(output, label_tensor)
loss.backward()
optimizer.step()
print(list(model.parameters()))
What happens
Your problem is related to PyTorch's implicit conversion of torch.tensor to float. Let's say you have this:
tr_mat = torch.tensor(
[
[1, 0, 0, self.params[3]],
[0, 1, 0, self.params[4]],
[0, 0, 1, self.params[5]],
[0, 0, 0, 1],
],
requires_grad=True,
)
torch.tensor can only be constructed from list which has Python like values, it cannot have torch.tensor inside it. What happens under the hood (let's say) is each element of self.params which can be converted to float is (in this case all of them can, e.g. self.params[3], self.params[4], self.params[5]).
When tensor's value is casted to float it's value is copied into Python counterpart hence it is not part of computational graph anymore, it's a new pure Python variable (which cannot be backpropagated obviously).
Solution
What you can do is choose elements of your self.params and insert them into eye matrices so the gradient flows. You can see a rewrite of your forward method taking this into account:
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.params = nn.Parameter(torch.randn(6))
def forward(self, x):
sinus = torch.cos(self.params)
cosinus = torch.cos(self.params)
tr_mat = torch.eye(4)
tr_mat[:-1, -1] = self.params[3:]
rz_mat = torch.eye(4)
rz_mat[0, 0] = cosinus[2]
rz_mat[0, 1] = -sinus[2]
rz_mat[1, 0] = sinus[2]
rz_mat[1, 1] = cosinus[2]
ry_mat = torch.eye(4)
ry_mat[0, 0] = cosinus[1]
ry_mat[0, 2] = sinus[1]
ry_mat[2, 0] = -sinus[1]
ry_mat[2, 2] = cosinus[1]
rx_mat = torch.eye(4)
rx_mat[1, 1] = cosinus[0]
rx_mat[1, 2] = -sinus[0]
rx_mat[2, 1] = sinus[0]
rx_mat[2, 2] = cosinus[0]
tf1 = torch.matmul(tr_mat, rz_mat)
tf2 = torch.matmul(tf1, ry_mat)
tf3 = torch.matmul(tf2, rx_mat)
tr_local = torch.tensor(
[[1, 0, 0, x[0]], [0, 1, 0, x[1]], [0, 0, 1, x[2]], [0, 0, 0, 1]],
)
tf_output = torch.matmul(tf3, tr_local)
output = tf_output[:3, 3]
return output
(you may want to double check this rewrite but the idea holds).
Also notice tr_local can be done "your way" as we don't need any values to keep gradient.
requires_grad
You can see requires_grad wasn't used anywhere in the code. It's because what requires gradient is not the whole eye matrix (we will not optimize 0 and 1), but parameters which are inserted into it. Usually you don't need requires_grad at all in your neural network code because:
input tensors are not optimized (usually, those could be when you are doing adversarial attacks or such)
nn.Parameter requires gradient by default (unless frozen)
layers and other neural network specific stuff requires gradient by default (unless frozen)
values which don't need gradient (input tensors) going through layers which do require it (or parameters or w/e) can be backpropagated
I have 10,000 images in RGB in an ndarray the size of (10000, 32, 32, 3).
I'd like to efficiently compress the images (take the means of colors) to 2x2, 4x4 etc. using numpy. The only idea I've got so far is to manually split the images, compress, and put together the pieces within the loops. Is there a more elegant solution?
You could do something like this, using scipy.ndimage.zoom:
import numpy as np
import scipy.ndimage as si
def resample(img, dims):
orig = img.shape[1]
new_imgs = []
for dim in dims:
factor = dim / orig
new_img = si.zoom(img, zoom=[1, factor, factor, 1])
new_imgs.append(new_img)
return new_imgs
For example, with random data:
>>> img = np.random.random((100, 32, 32, 3))
>>> resample(img, dims = [2, 4, 8, 16, 32])
>>> [img.shape for img in new_imgs]
[(100, 2, 2, 3),
(100, 4, 4, 3),
(100, 8, 8, 3),
(100, 16, 16, 3),
(100, 32, 32, 3)]
Note from the comment (below) that you might need to adjust the mode parameter in the zoom function.
You can use SciKit image's view_as_blocks and np.mean():
import numpy as np
import skimage
images = np.random.rand(10000, 32, 32, 3)
images_rescaled = skimage.util.view_as_blocks(images, (1, 4, 4, 1)).mean(axis=(-2, -3)).squeeze()
images_rescaled.shape
# (10000, 8, 8, 3)
I have my label tensor of shape (1,1,128,128,128) in which the values might range from 0,24. I want to convert this to one hot encoded tensor, using the nn.fucntional.one_hot function
n = 24
one_hot = torch.nn.functional.one_hot(indices, n)
but this expects a tensor of indices, honestly, I am not sure how to get those. The only tensor I have is the label tensor of the shape described above and it contains values ranging from 1-24, not the indices
How can I get a tensor of indices from my tensor? Thanks in advance.
If the error you are getting is this one:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: one_hot is only applicable to index tensor.
Maybe you just need to convert to int64:
import torch
# random Tensor with the shape you said
indices = torch.Tensor(1, 1, 128, 128, 128).random_(1, 24)
# indices.shape => torch.Size([1, 1, 128, 128, 128])
# indices.dtype => torch.float32
n = 24
one_hot = torch.nn.functional.one_hot(indices.to(torch.int64), n)
# one_hot.shape => torch.Size([1, 1, 128, 128, 128, 24])
# one_hot.dtype => torch.int64
You can use indices.long() too.
The torch.as_tensor function can also be helpful if your labels are stored in a list or numpy array:
import torch
import random
n_classes = 5
n_samples = 10
# Create list n_samples random labels (can also be numpy array)
labels = [random.randrange(n_classes) for _ in range(n_samples)]
# Convert to torch Tensor
labels_tensor = torch.as_tensor(labels)
# Create one-hot encodings of labels
one_hot = torch.nn.functional.one_hot(labels_tensor, num_classes=n_classes)
print(one_hot)
The output one_hot has shape (n_samples, n_classes) and should look something like:
tensor([[0, 0, 0, 1, 0],
[0, 1, 0, 0, 0],
[0, 1, 0, 0, 0],
[0, 0, 0, 1, 0],
[0, 0, 0, 1, 0],
[0, 0, 0, 1, 0],
[1, 0, 0, 0, 0],
[1, 0, 0, 0, 0],
[0, 0, 0, 1, 0],
[1, 0, 0, 0, 0]])
Usually, this issue can be solved by adding long().
for example,
import torch
import torch.nn.functional as F
labels=torch.Tensor([[0, 2, 1]])
n_classes=3
encoded=F.one_hot(labels, n_classes)
It gives an error as:
RuntimeError: one_hot is only applicable to index tensor.
To solve this issue, use long().
import torch
import torch.nn.functional as F
labels=torch.Tensor([[0, 2, 1]]).long()
n_classes=3
encoded=F.one_hot(labels, n_classes)
Now it would be executed without errors.
I'm running LBP algorithm to classify images by their texture features. Classifying method is LinearSVC in sklearn.svm package.
Getting histogram and fitting by SVM is done, but sometimes length of histogram varies depending on image.
Example is below:
from skimage import feature
from scipy.stats import itemfreq
from sklearn.svm import LinearSVC
import numpy as np
import cv2
import cvutils
import csv
import os
def __get_hist(image, radius):
NumPoint = radius*8
lbp = feature.local_binary_pattern(image, NumPoint, radius, method="uniform")
x = itemfreq(lbp.ravel())
hist = x[:,1]/sum(x[:,1])
return hist
def get_trainHist_list(train_txt):
train_dic = {}
with open(train_txt, 'r') as csvfile:
reader = csv.reader(csvfile, delimiter = ' ')
for row in reader:
train_dic[row[0]] = int(row[1])
hist_list=[]
key_list=[]
label_list=[]
for key, label in train_dic.items():
img = cv2.imread("D:/Python36/images/texture/%s" %key, cv2.IMREAD_GRAYSCALE)
key_list.append(key)
label_list.append(label)
hist_list.append(__get_hist(img,3))
bundle = [np.array(key_list), np.array(label_list), np.array(hist_list)]
return bundle
train_txt = 'D:/Python36/images/class_train.txt'
train_hist = get_trainHist_list(train_txt)
model = LinearSVC(C=100.0, random_state=42)
model.fit(train_hist[2], train_hist[1])
for i in train_hist[2]:
print(len(i))
test_img = cv2.imread("D:/Python36/images/texture_test/flat-3.png", cv2.IMREAD_GRAYSCALE)
hist= np.array(__get_hist(test_img, 3))
print(len(hist))
prediction = model.predict([hist])
print(prediction)
result
26
26
26
26
26
26
25
Traceback (most recent call last):
File "D:\Python36\texture.py", line 44, in <module>
prediction = model.predict([hist])
File "D:\Python36\lib\site-packages\sklearn\linear_model\base.py", line 324, in predict
scores = self.decision_function(X)
File "D:\Python36\lib\site-packages\sklearn\linear_model\base.py", line 305, in decision_function
% (X.shape[1], n_features))
ValueError: X has 25 features per sample; expecting 26
As you can see, length of histogram for training images is all 26, but test_img's is 25. For this reason, predict in SVM doesn't work.
I guess test_img has empty parts in the histogram, and that empty parts could have skipped. (I'm not sure)
Someone have idea to fix it?
There are 59 different uniform LBPs for a neighbourhood of 8 points. This should be the dimension of your feature vectors, but it is not because you used itemfreq to compute the histograms (as a side note, itemfreq is deprecated). The length of the histograms obtained throug itemfreq is the number of different uniform LBPs in the image. If some uniform LBPs are not present in the image the number of bins of the resulting histogram will be lower than 59. This issue can be easily fixed by utilizing bincount as demonstrated in the toy example below:
import numpy as np
from skimage import feature
from scipy.stats import itemfreq
lbp = np.array([[0, 0, 0, 0],
[1, 1, 1, 1],
[8, 8, 9, 9]])
hi = itemfreq(lbp.ravel())[:, 1] # wrong approach
hb = np.bincount(lbp.ravel(), minlength=59) # proposed method
The output looks like this:
In [815]: hi
Out[815]: array([4, 4, 2, 2], dtype=int64)
In [816]: hb
Out[816]:
array([4, 4, 0, 0, 0, 0, 0, 0, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0], dtype=int64)