Related
I have a LSTM model I am using to predict the unemployment rate from federal reserve filings. It uses glove vectors and vocab2index embedding and the training went as planned. However, upon attempting to feed a word embedding into the model for prediction testing it keeps throwing various errors.
Here is the model:
def load_glove_vectors(glove_file= glove_embedding_vectors_text_file):
"""Load the glove word vectors"""
word_vectors = {}
with open(glove_file) as f:
for line in f:
split = line.split()
word_vectors[split[0]] = np.array([float(x) for x in split[1:]])
return word_vectors
def get_emb_matrix(pretrained, word_counts, emb_size = 300):
""" Creates embedding matrix from word vectors"""
vocab_size = len(word_counts) + 2
vocab_to_idx = {}
vocab = ["", "UNK"]
W = np.zeros((vocab_size, emb_size), dtype="float32")
W[0] = np.zeros(emb_size, dtype='float32') # adding a vector for padding
W[1] = np.random.uniform(-0.25, 0.25, emb_size) # adding a vector for unknown words
vocab_to_idx["UNK"] = 1
i = 2
for word in word_counts:
if word in word_vecs:
W[i] = word_vecs[word]
else:
W[i] = np.random.uniform(-0.25,0.25, emb_size)
vocab_to_idx[word] = i
vocab.append(word)
i += 1
return W, np.array(vocab), vocab_to_idx
word_vecs = load_glove_vectors()
pretrained_weights, vocab, vocab2index = get_emb_matrix(word_vecs, counts)
Unfortunately when I feed this array
[array([ 3, 10, 6287, 6, 113, 271, 3, 6639, 104, 5105, 7525,
104, 7526, 9, 23, 9, 10, 11, 24, 7527, 7528, 104,
11, 24, 7529, 7530, 104, 11, 24, 7531, 7530, 104, 11,
24, 7532, 7530, 104, 11, 24, 7533, 7534, 24, 7535, 7536,
104, 7537, 104, 7538, 7539, 7540, 6643, 7541, 7354, 7542, 7543,
7544, 9, 23, 9, 10, 11, 24, 25, 8, 10, 11,
24, 3, 10, 663, 168, 9, 10, 290, 291, 3, 4909,
198, 10, 1478, 169, 15, 4621, 3, 3244, 3, 59, 1967,
113, 59, 520, 198, 25, 5105, 7545, 7546, 7547, 7546, 7548,
7549, 7550, 1874, 10, 7551, 9, 10, 11, 24, 7552, 6287,
7553, 7554, 7555, 24, 7556, 24, 7557, 7558, 7559, 6, 7560,
323, 169, 10, 7561, 1432, 6, 3134, 3, 7562, 6, 7563,
1862, 7144, 741, 3, 3961, 7564, 7565, 520, 7566, 4833, 7567,
7568, 4901, 7569, 7570, 4901, 7571, 1874, 7572, 12, 13, 7573,
10, 7574, 7575, 59, 7576, 59, 638, 1620, 7577, 271, 6488,
59, 7578, 7579, 7580, 7581, 271, 7582, 7583, 24, 669, 5932,
7584, 9, 113, 271, 3764, 3, 5930, 3, 59, 4901, 7585,
793, 7586, 7587, 6, 1482, 520, 7588, 520, 7589, 3246, 7590,
13, 7591])
into torch.LongTensor() I keep getting the following error:
TypeError: can't convert np.ndarray of type numpy.object_. The only supported types are: float64, float32, float16, complex64, complex128, int64, int32, int16, int8, uint8, and bool.
Any ideas on how to remedy? I am fairly new to AI in general, and I am an economist by trade so I am almost certain I have made a boneheaded error.
I am performing an image segmentation with a u-net model.
My mask has classes from 0-50.
I also have a text file dictionary with codes representing each class.
For example -
{1: '1234', 2:'5678', 3:'1245'} etc.
How do I combine when the 2 first string characters are the same so for example above key 1 and 3 are the same because they both start with "12".
How can I do this for all classes?
firstTwoCharDict = {}
for key, value in dictionary.items():
if key == 0:
value == value
firstTwoCharDict[key] = value
else:
value = value[:2]
firstTwoCharDict[key] = value
newDict = {}
for key, value in firstTwoCharDict.items():
if value not in newDict:
newDict[value] = [key]
else:
newDict[value].append(key)
This provides this
{'62': [1, 39],
'90': [2, 5, 9, 20, 32, 42, 47, 72, 88, 91, 95],
'97': [3, 49, 55],
'98': [4, 24, 34, 40, 53, 76, 81, 90, 96],
'31': [6, 17, 30, 48, 83],
'69': [7, 13, 15, 16, 27, 44, 51, 54, 56, 75],
'79': [8, 50],
'71': [10, 19, 22, 35, 61, 63, 65],
'99': [11, 12, 21, 46, 52, 69, 78, 84, 89],
'48': [14, 36, 74],
'60': [18],
'64': [23, 38, 66, 97]
```
Now i have an 2d array with integers, how do I replace them with they keys if the array values are equal to the values in the dict?
I have a small matrix A with dimensions MxNxO
I have a large matrix B with dimensions KxMxNxP, with P>O
I have a vector ind of indices of dimension Ox1
I want to do:
B[1,:,:,ind] = A
But, the lefthand of my equation
B[1,:,:,ind].shape
is of dimension Ox1xMxN and therefore I can not broadcast A (MxNxO) into it.
Why does accessing B in this way change the dimensions of the left side?
How can I easily achieve my goal?
Thanks
There's a feature, if not a bug, that when slices are mixed in the middle of advanced indexing, the sliced dimensions are put at the end.
Thus for example:
In [204]: B = np.zeros((2,3,4,5),int)
In [205]: ind=[0,1,2,3,4]
In [206]: B[1,:,:,ind].shape
Out[206]: (5, 3, 4)
The 3,4 dimensions have been placed after the ind, 5.
We can get around that by indexing first with 1, and then the rest:
In [207]: B[1][:,:,ind].shape
Out[207]: (3, 4, 5)
In [208]: B[1][:,:,ind] = np.arange(3*4*5).reshape(3,4,5)
In [209]: B[1]
Out[209]:
array([[[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]],
[[20, 21, 22, 23, 24],
[25, 26, 27, 28, 29],
[30, 31, 32, 33, 34],
[35, 36, 37, 38, 39]],
[[40, 41, 42, 43, 44],
[45, 46, 47, 48, 49],
[50, 51, 52, 53, 54],
[55, 56, 57, 58, 59]]])
This only works when that first index is a scalar. If it too were a list (or array), we'd get an intermediate copy, and couldn't set the value like this.
https://docs.scipy.org/doc/numpy-1.15.0/reference/arrays.indexing.html#combining-advanced-and-basic-indexing
It's come up in other SO questions, though not recently.
weird result when using both slice indexing and boolean indexing on a 3d array
The following script computes R-squared value between two numpy arrays(x and y).
The R-squared value is very low due to outliers in the data. How can I extract the indices of those outliers?
import numpy as np, matplotlib.pyplot as plt, scipy.stats as stats
x = np.random.random_integers(1,50,50)
y = np.random.random_integers(1,50,50)
r2 = stats.linregress(x, y) [3]**2
print r2
plt.scatter(x, y)
plt.show()
An outlier is defined as: value-mean > 2*standard deviation.
You can do this with the line
[i for i in range(len(x)) if (abs(x[i] - np.mean(x)) > 2*np.std(x))]
What is does:
A list is constructed from the indices of x, where the element at that index satisfies the condition described above.
A quick test:
x = np.random.random_integers(1,50,50)
this gives me the array:
array([16, 6, 13, 18, 21, 37, 31, 8, 1, 48, 4, 40, 9, 14, 6, 45, 20,
15, 14, 32, 30, 8, 19, 8, 34, 22, 49, 5, 22, 23, 39, 29, 37, 24,
45, 47, 21, 5, 4, 27, 48, 2, 22, 8, 12, 8, 49, 12, 15, 18])
Now I add some outliers manually as there are none initially:
x[4] = 200
x[15] = 178
lets test:
[i for i in range(len(x)) if (abs(x[i] - np.mean(x)) > 2*np.std(x))]
result:
[4, 15]
Is this what you was looking for?
EDIT:
I added the abs() function in the line above, because when you are working with negative numbers this might end bad. The abs() function takes the absolute value.
I think Sander's approach is the correct one, but if you must see R2 without those outliers before making a decision here is a way to do it.
Setup data and introduce outlier:
In [1]:
import numpy as np, scipy.stats as stats
np.random.seed(123)
x = np.random.random_integers(1,50,50)
y = np.random.random_integers(1,50,50)
y[5] = 100
Calculate R2 taking out one y value at a time (along with matching x value):
m = np.eye(y.shape[0])
r2 = np.apply_along_axis(lambda a: stats.linregress(np.delete(x, a.argmax()), np.delete(y, a.argmax()))[3]**2, 0, m)
Get index of the biggest outlier:
r2.argmax()
Out[1]:
5
Get R2 when this outlier is taken out:
In [2]:
r2[r2.argmax()]
Out[2]:
0.85892084723588935
Get the value of the outlier:
In [3]:
y[r2.argmax()]
Out[3]:
100
To get top n outliers:
In [4]:
n = 5
sorted_index = r2.argsort()[::-1]
sorted_index[:n]
Out [4]:
array([ 5, 27, 34, 0, 17], dtype=int64)
i am newbie and just want to implement Hierarchical Agglomerative clustering for RGB images. For this I extract all values of RGB from an image. And I process image.Next I find its distance and then develop the linkage. Now from linkage I want to extract my original data (i.e RGB values) on specified indices with indices id. Here is code I have done so far.
image = Image.open('image.jpg')
image = image.convert('RGB')
im = np.array(image).reshape((-1,3))
rgb = list(im.getdata())
X = pdist(im)
Y = linkage(X)
I = inconsistent(Y)
based on the 4th column of consistency. I opt minimum value of the cutoff in order to get maximum clusters.
cutoff = 0.7
cluster_assignments = fclusterdata(Y, cutoff)
# Print the indices of the data points in each cluster.
num_clusters = cluster_assignments.max()
print "%d clusters" % num_clusters
indices = cluster_indices(cluster_assignments)
ind = np.array(enumerate(rgb))
for k, ind in enumerate(indices):
print "cluster", k + 1, "is", ind
dendrogram(Y)
I got results like this
cluster 6 is [ 6 11]
cluster 7 is [ 9 12]
cluster 8 is [15]
Means cluster 6 contains the indices of 6 and 11 leafs. Now at this point I stuck in how to map these indices to get original data(i.e rgb values). indices of each rgb values to each pixel in the image. And then I have to generate codebook to implement Agglomeration Clustering. I have no idea how to approach this task. Read a lot of stuff but nothing clued.
Here is my solution:
import numpy as np
from scipy.cluster import hierarchy
im = np.array([[54,101,9],[ 67,89,27],[ 67,85,25],[ 55,106,1],[ 52,108,0],
[ 55,78,24],[ 19,57,8],[ 19,46,0],[ 95,110,15],[112,159,57],
[ 67,118,26],[ 76,127,35],[ 74,128,30],[ 25,62,0],[100,120,9],
[127,145,61],[ 48,112,25],[198,25,21],[203,11,10],[127,171,60],
[124,173,45],[120,133,19],[109,137,18],[ 60,85,0],[ 37,0,0],
[187,47,20],[127,170,52],[ 30,56,0]])
groups = hierarchy.fclusterdata(im, 0.7)
idx_sorted = np.argsort(groups)
group_sorted = groups[idx_sorted]
im_sorted = im[idx_sorted]
split_idx = np.where(np.diff(group_sorted) != 0)[0] + 1
np.split(im_sorted, split_idx)
output:
[array([[203, 11, 10],
[198, 25, 21]]),
array([[187, 47, 20]]),
array([[127, 171, 60],
[127, 170, 52]]),
array([[124, 173, 45]]),
array([[112, 159, 57]]),
array([[127, 145, 61]]),
array([[25, 62, 0],
[30, 56, 0]]),
array([[19, 57, 8]]),
array([[19, 46, 0]]),
array([[109, 137, 18],
[120, 133, 19]]),
array([[100, 120, 9],
[ 95, 110, 15]]),
array([[67, 89, 27],
[67, 85, 25]]),
array([[55, 78, 24]]),
array([[ 52, 108, 0],
[ 55, 106, 1]]),
array([[ 54, 101, 9]]),
array([[60, 85, 0]]),
array([[ 74, 128, 30],
[ 76, 127, 35]]),
array([[ 67, 118, 26]]),
array([[ 48, 112, 25]]),
array([[37, 0, 0]])]