def gpu_calculation(train_data):
train_data = train_data - train_data.mean(axis=1, keepdims=True)
cov_mat = np.cov(train_data)
np.save('gpu_cov_mat.npy', cov_mat)
cov_mat = gpuarray.to_gpu(cov_mat)
va_eig, ve_eig = linalg.eig(cov_mat, 'V', 'N')
val_eig = np.asarray(va_eig)
vec_eig = np.asarray(ve_eig)
np.save('gpu_eig_val.npy', val_eig)
np.save('gpu_eig_vec.npy', vec_eig)
# cov_mat=np.load('gpu_cov_mat.npy')
# val_eig=np.load('gpu_eig_val.npy')
# vec_eig=np.load('gpu_eig_vec.npy')
idx = val_eig.argsort()[::-1]
aval_eig = val_eig[idx]
avec_eig = vec_eig[:, idx].T
pca_vec = []
total = sum(val_eig)
eig_sum = 0
for i in range(0, len(aval_eig)):
eig_sum = eig_sum + aval_eig[i]
if (eig_sum / total) < 0.9:
pca_vec.append(avec_eig[i])
else:
break
# train_data_pca=np.multiply(np.transpose(vec_eig),train_data)
# test_data_pca = np.multiply(np.transpose(vec_eig), test_data)
# Display first eigenface
pca_vec = np.asarray(pca_vec)
pca_vec = np.real(pca_vec)
dis_eig = np.real(avec_eig[0])
dis_eig[dis_eig < 0] = 0
def sorted_eig(X, ascending=True, mode='cpu'):
if mode == 'cpu':
e_vals, e_vecs = np.linalg.eig(X)
idx = np.argsort(e_vals)
if not ascending:
idx = idx[::-1]
e_vecs = e_vecs[:, idx]
e_vals = e_vals[idx]
return e_vals, e_vecs
elif mode == 'gpu':
import skcuda.linalg as LA
import pycuda.gpuarray as gpuarray
e_vecs_gpu, e_vals_gpu = LA.eig(X, 'N', 'V', lib='cusolver')
e_vals = e_vals_gpu.get()
idx = np.argsort(e_vals)
V_gpu = gpuarray.empty((X.shape[0], X.shape[1]), np.float32)
d = X.shape[0]
for i in range(d):
V_gpu[i] = e_vecs_gpu[idx[i]]
V_gpu = LA.transpose(V_gpu)
return e_vals, V_gpu
linalg.init()
# loading the sparse matrix
supmat = np.load(
'/scratch/gpfs/sbdas/Helioseismology/qdpy_output/output_files/w135_antia/super_matrix.npy'
)
supmat_gpu = gpuarray.to_gpu(supmat)
# number of timeit iterations
niter = 10
# timing the numpy eigenvalue solver
time_numpy = timeit(lambda: np.linalg.eigh(supmat), number=niter)
# timing the skcuda eigenvalue solver
time_skcuda = timeit(lambda: linalg.eig(supmat_gpu), number=niter)
time_numpy /= niter
time_skcuda /= niter
print(f'numpy: {time_numpy}')
print(f'skcuda: {time_skcuda}')
# getting eigenvalues and eigenvectors for both cases to compare
eval_np, evec_np = np.linalg.eigh(supmat)
eval_sk_gpu, evec_sk_gpu = linalg.eig(supmat_gpu)
evec_sk, eval_sk = eval_sk_gpu.get(), evec_sk_gpu.get()
# reorderign the skcuda matrix
evec_sk = np.transpose(evec_sk)
def test_eig_complex128(self):
a = np.array(np.random.rand(9, 9) + 1j*np.random.rand(9,9), np.complex128, order='F')
a_gpu = gpuarray.to_gpu(a)
w_gpu = linalg.eig(a_gpu, 'N', 'N')
assert np.allclose(np.trace(a), sum(w_gpu.get()), atol=atol_float64)
def test_eig_float32(self):
a = np.asarray(np.random.rand(9, 9), np.float32, order='F')
a_gpu = gpuarray.to_gpu(a)
w_gpu = linalg.eig(a_gpu, 'N', 'N')
assert np.allclose(np.trace(a), sum(w_gpu.get()), atol=1e-4)
def getEigenValues(inData,
maxNumClusters,
catCols=np.array([], dtype=int),
save_to_file="eigenValues.csv",
additional_comments=""):
### Input:
### # inData: (numpy.ndarray) Array of shape (numSamples, numFeatures)
### # maxNumClusters: (int) Maximum number of clusters we want
### # catCols: (numpy.ndarray, dtype=int) 1-D array containing indices of categorical columns in inData
### # save_to_file: (str) Filepath to save the eigenvalues. Default: ./eigenValues.csv
### # additional_comments: (str) Any additional comments to be added in the header of output file. (e.g. "Data that I collected last Monday")
###
### Output:
### # eigVals: (numpy.ndarray) Array of shape (maxNumClusters+1, numFeatures+1). eigVals[:,i] (i>0) contains the first maxNumClusters+1 eigenvalues of the laplacian with i'th feature removed (i = 1, 2, ... numFeatures). eigVals[:,0] contains the first maxNumClusters+1 eigenvalues of the laplacian with no feature removed.
###
### Description:
### # Gets eigenValues of the laplacian matrix for input data inData, with catCols. This is done iteratively numFeatures+1 ( = inData.shape[1]+1) times.
###
### Note:
### # This function is expected to take a long time to execute, depending on the size on inData. Suggest calling this function on smaller data first, and proceed on actual data, if all is good.
start_time = time.time()
if np.argwhere(np.isnan(inData)).size != 0:
raise ValueError("Input array inData contains NaN entries.")
return 0
inData = np.asanyarray(inData, dtype=np.float32, order='C')
numSamples = inData.shape[0]
numFeatures = inData.shape[1]
eigVals = np.zeros((maxNumClusters + 1, numFeatures + 1),
dtype=np.float32,
order='C')
print("\n#### Working on base data\n")
similarity_matrix_gpu = ck.clinicalKernel(inData,
catCols=catCols,
return_to_CPU=False)
laplacian_matrix = laplacian.laplacian_normalised(similarity_matrix_gpu,
numRows=numSamples,
simMatrix_on_CPU=False,
return_to_CPU=True)
### Change the order of laplacian_matrix from 'C' to 'F', because cuSolver and skcuda like 'F'
### Send laplacian_matrix back to GPU as pycuda.gpuarray.GPUArray, with order 'F'
laplacian_matrix = np.asanyarray(laplacian_matrix,
order='F',
dtype=np.float32)
laplacian_matrix_gpu = gpuarray.to_gpu(laplacian_matrix)
linalg.init()
print("\n# Computing eigenvalues ...")
eigVals_current = linalg.eig(laplacian_matrix_gpu,
jobvl='N',
jobvr='N',
lib='cusolver')
eigVals_current = eigVals_current.get(
) #Convert GPUArray to numpy array and sort in increasing order
eigVals_current.sort()
eigVals_current = eigVals_current[:(maxNumClusters + 1)]
eigVals[:, 0] = eigVals_current
### Main for loop that iterates over features
for i in range(numFeatures):
time_elapsed = time.strftime("%H:%M:%S",
time.gmtime(time.time() - start_time))
print("\n#### Feature dropped: " + repr(i + 1) + " out of " +
repr(numFeatures) + " (time elapsed: " + time_elapsed + ")")
catCols_reduced = np.delete(
catCols,
np.nonzero(catCols == i)) #If a catCols is currently being dropped
mask = catCols_reduced > i
catCols_reduced[mask] = catCols_reduced[
mask] - 1 # When i'th feature is dropped, indices > i decrease by 1
# print("catCols: "+repr(catCols_reduced))
inData_reduced = np.delete(inData, i, axis=1)
similarity_matrix_gpu = ck.clinicalKernel(inData_reduced,
catCols=catCols_reduced,
return_to_CPU=False)
laplacian_matrix = laplacian.laplacian_normalised(
similarity_matrix_gpu,
numRows=numSamples,
simMatrix_on_CPU=False,
return_to_CPU=True)
### Change the order of laplacian_matrix from 'C' to 'F', because cuSolver and skcuda like 'F'
### Send laplacian_matrix back to GPU as pycuda.gpuarray.GPUArray, with order 'F'
laplacian_matrix = np.asanyarray(laplacian_matrix,
order='F',
dtype=np.float32)
laplacian_matrix_gpu = gpuarray.to_gpu(laplacian_matrix)
print("\n# Computing eigenvalues ...")
eigVals_current = linalg.eig(laplacian_matrix_gpu,
jobvl='N',
jobvr='N',
lib='cusolver')
eigVals_current = eigVals_current.get(
) #Convert GPUArray to numpy array and sort in increasing order
eigVals_current.sort()
eigVals_current = eigVals_current[:(maxNumClusters + 1)]
eigVals[:, (i + 1)] = eigVals_current
timeNow = datetime.datetime.now()
timeString = timeNow.strftime("%H:%M:%S %Y-%m-%d")
header = "##### Eigenvalues #####\n# " + timeString + "\n# First " + repr(
maxNumClusters + 1
) + " Eigenvalues of Laplacian.\n# Eigenvalues stored along columns. 0th column for no feature dropped, i'th column for i'th feature dropped (i > 0)\n#\n# Comment: " + str(
additional_comments)
np.savetxt(save_to_file,
eigVals,
header=header,
comments="",
delimiter=",")
print("\n#### Iterations complete\n")
print("\n# Eigenvalues saved to file " + save_to_file)
return eigVals
def eigh_gpu(X):
X=gpuarray.to_gpu(X)
eigen, w_gpu = cuda_linalg.eig(X, lib='cusolver')
return eigen.get(),w_gpu.get()
