def dot(x,y):
return sum(x*y)
def matmult_high_level(X,Y):
return np.array([[np.dot(x,y) for y in Y.T] for x in X])
def matmult_loops(X,Y,Z):
m, d = X.shape
n = Y.shape[1]
for i in xrange(m):
for j in xrange(n):
total = X[i,0] * Y[0,j]
for k in xrange(1,d):
total += X[i,k] * Y[k,j]
Z[i,j] = total
return Z
n, d = 250, 1000
m = 250
X = np.random.randn(n,d)
Y = np.random.randn(d,m)
Z = np.zeros((m,n))
from timer import compare_perf
compare_perf(matmult_high_level, [X,Y],cpython=True)
compare_perf(matmult_loops, [X, Y, Z], cpython=False)
def clamp(i, offset, maxval):
j = max(0, i + offset)
return min(j, maxval)
def reflect(pos, offset, bound):
idx = pos + offset
return min(2 * (bound - 1) - idx, max(idx, -idx))
def conv(x, weights, mode=clamp):
sx = x.shape
sw = weights.shape
result = np.zeros_like(x)
for i in xrange(sx[0]):
for j in xrange(sx[1]):
for ii in xrange(sw[0]):
for jj in xrange(sw[1]):
idx = mode(i, ii - sw[0] / 2,
sw[0]), mode(j, jj - sw[0] / 2, sw[0])
result[i, j] += x[idx] * weights[ii, jj]
return result
xsize = (300, 300)
x = np.random.randn(*xsize)
wsize = (5, 5)
w = np.random.randn(*wsize)
compare_perf(conv, [x, w])
for ii in xrange(max(i-window_radius, 0), min(i+window_radius+1, height)):
if ii != i or jj != j:
d = image[i, j, :] - image[ii, jj, :]
s = np.sum(d**2)
gval = 1.0 - np.sqrt(s) / np.sqrt(3)
attack_strength = gval * state[ii, jj, 1]
if attack_strength > defense_strength:
defense_strength = attack_strength
winning_colony = state[ii, jj, 0]
changes += 1
state_next[i, j, 0] = winning_colony
state_next[i, j, 1] = defense_strength
return changes
N = 50
dtype = np.double
image = np.zeros((N, N, 3), dtype=dtype)
state = np.zeros((N, N, 2), dtype=dtype)
state_next = np.empty_like(state)
# colony 1 is strength 1 at position 0,0
# colony 0 is strength 0 at all other positions
state[0, 0, 0] = 1
state[0, 0, 1] = 1
window_radius = 10
from timer import compare_perf
compare_perf(growcut_python, [image, state, state_next, window_radius])
from parakeet import jit
import timer
def covariance(x,y):
return ((x-x.mean()) * (y-y.mean())).mean()
def fit_simple_regression(x,y):
slope = covariance(x,y) / covariance(x,x)
offset = y.mean() - slope * x.mean()
return slope, offset
import numpy as np
N = 10**7
x = np.random.randn(N).astype('float64')
slope = 903.29
offset = 102.1
y = slope * x + offset
jit(fit_simple_regression)(x,y)
timer.compare_perf(fit_simple_regression, (x,y))
def rosen_der_np(x):
der = np.empty_like(x)
der[1:-1] = (+ 200 * (x[1:-1] - x[:-2] ** 2)
- 400 * (x[2:] - x[1:-1] ** 2) * x[1:-1]
- 2 * (1 - x[1:-1]))
der[0] = -400 * x[0] * (x[1] - x[0] ** 2) - 2 * (1 - x[0])
der[-1] = 200 * (x[-1] - x[-2] ** 2)
return der
def rosen_der_loops(x):
n = x.shape[0]
der = np.empty_like(x)
for i in range(1, n - 1):
der[i] = (+ 200 * (x[i] - x[i - 1] ** 2)
- 400 * (x[i + 1]
- x[i] ** 2) * x[i]
- 2 * (1 - x[i]))
der[0] = -400 * x[0] * (x[1] - x[0] ** 2) - 2 * (1 - x[0])
der[-1] = 200 * (x[-1] - x[-2] ** 2)
return der
if __name__ == '__main__':
N = 10**5
x = np.arange(N) / float(N)
jit(rosen_der_np)(x)
from timer import compare_perf
# numba still crashes on negative indexing
compare_perf(rosen_der_np, [x.copy()], numba=False)
compare_perf(rosen_der_loops, [x.copy()], numba=False)
import numpy as np
from timer import compare_perf
# Simple convolution of 3x3 patches from a given array x
# by a 3x3 array of filter weights
def conv_3x3_trim(x, weights):
return np.array([[(x[i-1:i+2, j-1:j+2]*weights).sum()
for j in xrange(1, x.shape[1] -2)]
for i in xrange(1, x.shape[0] -2)])
x = np.random.randn(500,500)
w = np.random.randn(3,3)
compare_perf(conv_3x3_trim, [x,w])
def clamp(i, offset, maxval):
j = max(0, i + offset)
return min(j, maxval)
def reflect(pos, offset, bound):
idx = pos+offset
return min(2*(bound-1)-idx,max(idx,-idx))
def conv(x, weights, mode=clamp):
sx = x.shape
sw = weights.shape
result = np.zeros_like(x)
for i in xrange(sx[0]):
for j in xrange(sx[1]):
for ii in xrange(sw[0]):
for jj in xrange(sw[1]):
idx = mode(i,ii-sw[0]/2,sw[0]), mode(j,jj-sw[0]/2,sw[0])
result[i,j] += x[idx] * weights[ii,jj]
return result
xsize = (300,300)
x = np.random.randn(*xsize)
wsize = (5,5)
w = np.random.randn(*wsize)
compare_perf(conv, [x,w])
import numpy as np def dists(X,Y): return np.array([[np.sum( (x-y) ** 2) for x in X] for y in Y]) d = 100 X = np.random.randn(1000,d) Y = np.random.randn(200,d) from timer import compare_perf compare_perf(dists, [X,Y])
return distance_matrix
def arc_distance_numpy_broadcast(a, b):
"""
Calculates the pairwise arc distance between all points in vector a and b.
"""
theta_1 = a[:, 0][:, None]
theta_2 = b[:, 0][None, :]
phi_1 = a[:, 1][:, None]
phi_2 = b[:, 1][None, :]
temp = (np.sin((theta_2 - theta_1) / 2)**2
+
np.cos(theta_1) * np.cos(theta_2)
* np.sin((phi_2 - phi_1) / 2)**2)
distance_matrix = 2 * (np.arctan2(np.sqrt(temp), np.sqrt(1 - temp)))
return distance_matrix
from timer import compare_perf
n = 1000
import numpy as np
a = np.random.rand(n, 2)
b = np.random.rand(n, 2)
compare_perf(arc_distance_python_nested_for_loops, [a,b])
compare_perf(arc_distance_numpy_broadcast, [a,b])
compare_perf(arc_distance_numpy_tile, [a,b])
# Simple convolution of 3x3 patches from a given array x
# by a 3x3 array of filter weights
def conv_3x3_trim(x, weights):
return np.array([[(x[i-1:i+2, j-1:j+2]*weights).sum()
for j in xrange(1, x.shape[1] -2)]
for i in xrange(1, x.shape[0] -2)])
x = np.random.randn(500,500)
w = np.random.randn(3,3)
# compare_perf(conv_3x3_trim, [x,w])
w = np.random.randn(5,5)
# Simple convolution of 5x5 patches from a given array x
# by a 5x5 array of filter weights
def conv_3x3_trim_loops(image, weights):
result = np.zeros_like(image)
for i in xrange(3,x.shape[0]-3):
for j in xrange(3,x.shape[1]-3):
for ii in xrange(5):
for jj in xrange(5):
result[i,j] += image[i-ii+2, j-jj+2] * weights[ii, jj]
return result
compare_perf(conv_3x3_trim_loops, [x,w])
min_dist = dist(x, C[0, :])
min_idx = 0
for cidx in xrange(1,k):
centroid = C[cidx,:]
curr_dist = np.sum(x-centroid)**2
if curr_dist < min_dist:
min_dist = curr_dist
min_idx = cidx
A[i] = min_idx
# recompute the clusters by averaging data points
# assigned to them
for cidx in xrange(k):
# reset centroids
for dim_idx in xrange(ndims):
C[cidx, dim_idx] = 0
# add each data point only to its assigned centroid
cluster_count = 0
for i in xrange(n):
if A[i] == cidx:
C[cidx, :] += X[i, :]
cluster_count += 1
C[cidx, :] /= cluster_count
return C
n, d = 10**3, 100
X = np.random.randn(n,d)
k = 5
from timer import compare_perf
compare_perf(kmeans_loops, [X, k, 10])
import numpy as np
def dists(X,Y):
result = np.zeros( (X.shape[0], Y.shape[0]), dtype=X.dtype)
for i in xrange(X.shape[0]):
for j in xrange(Y.shape[0]):
result[i,j] = np.sum( (X[i,:] - Y[j,:]) ** 2)
return result
d = 100
X = np.random.randn(1000,d)
Y = np.random.randn(200,d)
import timer
timer.compare_perf(dists, [X,Y])
def clamp(i, offset, maxval):
j = max(0, i + offset)
return min(j, maxval)
def reflect(pos, offset, bound):
idx = pos+offset
return min(2*(bound-1)-idx,max(idx,-idx))
def conv(x, weights, mode=clamp):
sx = x.shape
sw = weights.shape
result = np.zeros_like(x)
for i in xrange(sx[0]):
for j in xrange(sx[1]):
for ii in xrange(sw[0]):
for jj in xrange(sw[1]):
idx = mode(i,ii-sw[0]/2,sw[0]), mode(j,jj-sw[0]/2,sw[0])
result[i,j] += x[idx] * weights[ii,jj]
return result
xsize = (300,300)
x = np.random.randn(*xsize)
wsize = (5,5)
w = np.random.randn(*wsize)
# Nmba can't run this benchmark yet
compare_perf(conv, [x,w], numba=False)
for xidx in xrange(ndims):
curr_dist += (x[xidx] - centroid[xidx])**2
if curr_dist < min_dist:
min_dist = curr_dist
min_idx = cidx
A[i] = min_idx
# recompute the clusters by averaging data points
# assigned to them
for cidx in xrange(k):
# reset centroids
for dim_idx in xrange(ndims):
C[cidx, dim_idx] = 0
# add each data point only to its assigned centroid
cluster_count = 0
for i in xrange(n):
if A[i] == cidx:
C[cidx, :] += X[i, :]
cluster_count += 1
C[cidx, :] /= cluster_count
return C
n, d = 10**4, 50
X = np.random.randn(n,d)
k = 25
from timer import compare_perf
compare_perf(kmeans_comprehensions, [X, k, 5],cpython=False)
compare_perf(kmeans_loops, [X, k, 5], cpython=True)
