Related
I have a 3-D numpy array representing a model domain of 39 layers, 279 rows, 153 columns. The values in the array are either 0 or 1 and signify if the cell in the domain is inactive or active, respectively. I am trying to create a 2-D array of shape 279 rows and 153 columns where the array values equal the layer number for the uppermost active layer in the grid. Essentially, at each row, col location I want to loop through the layers to find the first one that is a 1 and not a 0 and then put that layer number in the 2-D array at that row, col location. For example:
If a four layer (layers 0-3) array looks like this:
array([[[ 0., 1., 0., 0.],
[ 1., 0., 0., 0.],
[ 1., 0., 0., 0.]],
[[ 0., 1., 1., 0.],
[ 1., 1., 0., 0.],
[ 1., 1., 0., 0.]],
[[ 0., 0., 1., 1.],
[ 0., 1., 1., 0.],
[ 0., 1., 1., 0.]],
[[ 0., 0., 1., 1.],
[ 0., 1., 1., 1.],
[ 0., 1., 1., 1.]]])
The 2-D array should look like this:
array([[[ 0., 0., 1., 2.],
[ 0., 1., 2., 3.],
[ 0., 1., 2., 3.]],
If the row-col location is not active (not equal to 1) in any layer , the value in the resulting array should be 0 (like at 1,1), the same as if it were active in layer 0.
I have tried modifying a couple of solutions where the z-axis values are summed, or averaged, but can't seem to figure out how to get exactly what I am looking for.
You could try numpy.argmax:
import numpy as np
a = np.array([[[ 0., 1., 0., 0.],
[ 1., 0., 0., 0.],
[ 1., 0., 0., 0.]],
[[ 0., 1., 1., 0.],
[ 1., 1., 0., 0.],
[ 1., 1., 0., 0.]],
[[ 0., 0., 1., 1.],
[ 0., 1., 1., 0.],
[ 0., 1., 1., 0.]],
[[ 0., 0., 1., 1.],
[ 0., 1., 1., 1.],
[ 0., 1., 1., 1.]]])
print(np.argmax(a,0))
array([[0, 0, 1, 2],
[0, 1, 2, 3],
[0, 1, 2, 3]])
This works because argmax returns the first max value when searching over the defined axis (in this case the 0th axis).
So I am trying to figure out how to train my matrix in a way that I will get a BandRNN.
BandRnn is a diagonalRNN model with a different number of connections per neuron.
For example:
C is the number of connections per neuron.
I found out that there is a way to turn off some of the gradients in a for loop, in a way that prevents them from being trained as follows:
for p in model.input.parameters():
p.requires_grad = False
But I can't find a proper way to do so, in a way that will make my matrix become a BandRNN.
Hopefully, someone will be able to help me with this issue.
As far as I know you can only activate/deactivate requires_grad on a tensor, and not on distinct components of that tensor. Instead what you could do is zero out the values outside the band.
First create a mask for the band, you could use torch.ones with torch.diagflat:
>>> torch.diagflat(torch.ones(5), offset=1)
By setting the right dimension for torch.ones as well as the right offset you can generate offset diagonal matrices with consistent shapes.
>>> N = 10; i = -1
>>> torch.diagflat(torch.ones(N-abs(i)), offset=i)
tensor([[0., 0., 0., 0., 0.],
[1., 0., 0., 0., 0.],
[0., 1., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 1., 0.]])
>>> N = 10; i = 0
>>> torch.diagflat(torch.ones(N-abs(i)), offset=i)
tensor([[1., 0., 0., 0., 0.],
[0., 1., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 1., 0.],
[0., 0., 0., 0., 1.]])
>>> N = 10; i = 1
>>> torch.diagflat(torch.ones(N-abs(i)), offset=i)
tensor([[0., 1., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 1., 0.],
[0., 0., 0., 0., 1.],
[0., 0., 0., 0., 0.]])
You get the point, summing these matrices element-wise allows use to get a mask:
>>> N = 10; b = 3
>>> mask = sum(torch.diagflat(torch.ones(N-abs(i)), i) for i in range(-b//2,b//2+1))
>>> mask
tensor([[1., 1., 0., 0., 0.],
[1., 1., 1., 0., 0.],
[1., 1., 1., 1., 0.],
[0., 1., 1., 1., 1.],
[0., 0., 1., 1., 1.]])
Then you can zero out the values outside the band on your nn.Linear:
>>> m = nn.Linear(N, N)
>>> m.weight.data = m.weight * mask
>>> m.weight
Parameter containing:
tensor([[-0.3321, -0.3377, -0.0000, -0.0000, -0.0000],
[-0.4197, 0.1729, 0.2101, 0.0000, 0.0000],
[ 0.3467, 0.2857, -0.3919, -0.0659, 0.0000],
[ 0.0000, -0.4060, 0.0908, 0.0729, -0.1318],
[ 0.0000, -0.0000, -0.4449, -0.0029, -0.1498]], requires_grad=True)
Note, you might need to perform this on each forward pass as the parameters outside the band might get updated to non-zero values during the training. Of course, you can initialize mask once and keep it in memory.
It would be more convenient to wrap everything into a custom nn.Module.
So, what I have is a data file in the form of
1 , 1 , 2
2 , 5 , 8
3 , 9 , 10
...
...
In my case, every single triplet is in the form of: value , x-position , y-position.
What i want to achieve is to insert this data in a 2d-matrix, which I already created using the np.zeros function. However, I am stuck and can't figure out how to write a function which puts the given values to the right x and y position in the matrix :/
My current Matrix (named matrix) looks like:
array([[0,0,0,...,0]
[0,0,0,...,0]
[... ]
[0,0,0,...,0]])
and if i would use matrix[1,1]=2 (first line of data) i would get:
array([[0,0,0,...,0]
[0,2,0,...,0]
[... ]
[0,0,0,...,0]])
My goal is to insert all lines of data in this way.
You can make use of the np.genfromtxt function [numpy-doc] where you set as delimiter=… parameter, the comma (','). So given you made a file data.txt, you can load that file into a numpy array with:
>>> import numpy as np
>>> np.genfromtxt('data.txt', delimiter=',')
array([[ 1., 1., 2.],
[ 2., 5., 8.],
[ 3., 9., 10.]])
Or if you are only interested in the x/y values, you can use the usecols=… parameter:
>>> np.genfromtxt('data.txt', delimiter=',', usecols=(1,2))
array([[ 1., 2.],
[ 5., 8.],
[ 9., 10.]])
You can load the data using genfromtxt():
import numpy as np
tmp = np.genfromtxt('data.txt', delimiter=',', dtype=int)
and then generate an empty data matrix a from the first two columns of tmp
a = np.zeros(np.max(tmp[:, :2], axis=0) + 1)
and populate it with values from tmp
a[tmp[:, 0], tmp[:, 1]] = tmp[:, 2]
a
# array([[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
# [ 0., 2., 0., 0., 0., 0., 0., 0., 0., 0.],
# [ 0., 0., 0., 0., 0., 8., 0., 0., 0., 0.],
# [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 10.]])
I am trying to calculate KL divergence using the entropy function of scipy.
My p is:
array([[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1.],
[ 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])
and q is:
array([[ 0.05242718, 0.04436347, 0.04130855, 0.04878344, 0.04310538,
0.02856853, 0.03303122, 0.02517992, 0.08525434, 0.03450324,
0.14580068, 0.1286993 , 0.28897473],
[ 0.65421444, 0.11592199, 0.0642645 , 0.02989768, 0.01385762,
0.01756484, 0.01024294, 0.00891479, 0.01140301, 0.00718939,
0.00938009, 0.01070139, 0.04644726],
[ 0.65984136, 0.13251236, 0.06345234, 0.02891162, 0.02429709,
0.02025307, 0.01073064, 0.01170066, 0.00678652, 0.00703361,
0.00560414, 0.00651137, 0.02236522],
[ 0.32315928, 0.23900077, 0.05460232, 0.03953635, 0.02901102,
0.01294443, 0.02372061, 0.02092882, 0.01188251, 0.01377188,
0.02976672, 0.05854314, 0.14313218],
[ 0.7717858 , 0.09692616, 0.03415596, 0.01713088, 0.01108141,
0.0128005 , 0.00847301, 0.01049734, 0.0052889 , 0.00514799,
0.00442508, 0.00485477, 0.01743218]], dtype=float32)
When I do:
entropy(p[0],q[0])
I am getting the following error:
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-201-563ea7d4decf> in <module>()
4 print('p0:',p[0])
5 print('q0:',q[0])
----> 6 entropy(p[0],q[0])
/Users/freelancer/anaconda/envs/py35/lib/python3.5/site-packages/matplotlib/mlab.py in entropy(y, bins)
1570 y = np.zeros((len(x)+2,), x.dtype)
1571 y[1:-1] = x
-> 1572 dif = np.diff(y)
1573 up = (dif == 1).nonzero()[0]
1574 dn = (dif == -1).nonzero()[0]
/Users/freelancer/anaconda/envs/py35/lib/python3.5/site-packages/numpy/lib/function_base.py in histogram(a, bins, range, normed, weights, density)
781 if (np.diff(bins) < 0).any():
782 raise ValueError(
--> 783 'bins must increase monotonically.')
784
785 # Initialize empty histogram
ValueError: bins must increase monotonically.
Why is it?
This works with the example arrays:
import scipy as sp
sp.stats.entropy(p[0], q[0])
Looking at the stack trace in the error massage, it becomes apparent that you did not call scipy's entropy function but matplotlib's entropy, which works differently.
Here is the relevant part:
/Users/freelancer/anaconda/envs/py35/lib/python3.5/site-packages/matplotlib/mlab.pyin entropy(y, bins)
Is there a higher (than two) dimensional equivalent of diag?
L = [...] # some arbitrary list.
A = ndarray.diag(L)
will create a diagonal 2-d matrix shape=(len(L), len(L)) with elements of L on the diagonal.
I'd like to do the equivalent of:
length = len(L)
A = np.zeros((length, length, length))
for i in range(length):
A[i][i][i] = L[i]
Is there a slick way to do this?
Thanks!
You can use diag_indices to get the indices to be set. For example,
x = np.zeros((3,3,3))
L = np.arange(6,9)
x[np.diag_indices(3,ndim=3)] = L
gives
array([[[ 6., 0., 0.],
[ 0., 0., 0.],
[ 0., 0., 0.]],
[[ 0., 0., 0.],
[ 0., 7., 0.],
[ 0., 0., 0.]],
[[ 0., 0., 0.],
[ 0., 0., 0.],
[ 0., 0., 8.]]])
Under the hood diag_indices is just the code Jaime posted, so which to use depends on whether you want it spelled out in a numpy function, or DIY.
You can use fancy indexing:
In [2]: a = np.zeros((3,3,3))
In [3]: idx = np.arange(3)
In [4]: a[[idx]*3] = 1
In [5]: a
Out[5]:
array([[[ 1., 0., 0.],
[ 0., 0., 0.],
[ 0., 0., 0.]],
[[ 0., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 0.]],
[[ 0., 0., 0.],
[ 0., 0., 0.],
[ 0., 0., 1.]]])
For a more general approach, you could set the diagonal of an arbitrarily sized array doing something like:
def set_diag(arr, values):
idx = np.arange(np.min(arr.shape))
arr[[idx]*arr.ndim] = values