How to use d_x2 method in fMBT

Best Python code snippet using fMBT_python

gpu.py

Source:gpu.py

1from __future__ import division2from numbapro import cuda3import numpy as np4import numbapro.cudalib.cublas5import numpy.random6import math7import scipy.sparse.linalg8import scipy.sparse as sps9def fista(I, Phi, lambdav, L=None, tol=10e-6, max_iterations=200, display=True, verbose=False):10	"""11	I: Images 12 	Phi: Dictionary 13 	lambdav: Sparse Penalty 14 	L = Largest eigenvalue of Phi 15 	"""16	b = numbapro.cudalib.cublas.Blas()17	(m, n) = Phi.shape18	(m, batch) = I.shape19	if L == None:20		L = scipy.sparse.linalg.svds(Phi, 1, which='LM', return_singular_vectors=False)21		print "Max eigenvalue: ." + str(L)22	L = (L**2)*2 # L = svd(Phi) -> eig(2*(Phi.T*Phi))23	invL = 1/L24	t = 1.25	if sps.issparse(Phi):26		Phi = np.array(Phi.todense())27	d_I = cuda.to_device(np.array(I, dtype=np.float32, order='F'))28	d_Phi = cuda.to_device(np.array(Phi, dtype=np.float32, order='F'))29	d_Q = cuda.device_array((n, n), dtype=np.float32, order='F')30	d_c = cuda.device_array((n, batch), dtype=np.float32, order='F')31	d_x = cuda.to_device(np.array(np.zeros((n, batch), dtype=np.float32), order='F'))32	d_y = cuda.to_device(np.array(np.zeros((n, batch), dtype=np.float32), order='F'))33	d_x2 = cuda.to_device(np.array(np.zeros((n, batch), dtype=np.float32), order='F'))34	# Temporary array variables35	d_t = cuda.device_array((m, batch), dtype=np.float32, order='F')36	d_t2 = cuda.device_array(n*batch, dtype=np.float32, order='F')37	b.gemm('T', 'N', n, n, m, 1, d_Phi, d_Phi, 0, d_Q) 	# Q = Phi^T * Phi38	b.gemm('T', 'N', n, batch, m, -2, d_Phi, d_I, 0, d_c) # c = -2*Phi^T * y39	blockdim = 32, 3240	griddim = int(math.ceil(n/blockdim[0])), int(math.ceil(batch/blockdim[1]))41	blockdim_1d = 25642	griddim_1d = int(math.ceil(n*batch/blockdim_1d))43	start = l2l1obj(b, d_I, d_Phi, d_x, d_t, d_t2, lambdav, blockdim_1d, griddim_1d)44	obj2 = start45	for i in xrange(max_iterations):46		# x2 = 2*Q*y + c47		b.symm('L', 'U', n, batch, 2, d_Q, d_y, 0, d_x2)48		b.geam('N', 'N', n, batch, 1, d_c, 1, d_x2, d_x2)49		50		# x2 = y - invL * x251		b.geam('N', 'N', n, batch, 1, d_y, -invL, d_x2, d_x2)52		# proxOp()						53		l1prox[griddim, blockdim](d_x2, invL*lambdav, d_x2)54		t2 = (1+math.sqrt(1+4*(t**2)))/2.055		56		# y = x2 + ((t-1)/t2)*(x2-x)57		b.geam('N', 'N', n, batch, 1+(t-1)/t2, d_x2, (1-t)/t2, d_x, d_y)58		# x = x259		b.geam('N', 'N', n, batch, 1, d_x2, 0, d_x, d_x)60		t = t261		# update objective62		obj = obj263		obj2 = l2l1obj(b, d_I, d_Phi, d_x2, d_t, d_t2, lambdav, blockdim_1d, griddim_1d)64		if verbose:65			x2 = d_x2.copy_to_host()66			print "L1 Objective: " + str(obj2)67		if np.abs(obj-obj2)/float(obj) < tol:68			break69	x2 = d_x2.copy_to_host()70	if display:71		print "FISTA Iterations: " + str(i)72		print "L1 Objective: " +  str(lambdav*np.sum(np.abs(x2)) + np.sum((I-Phi.dot(x2))**2))73		print "Objective delta: " + str(obj2-start)74	return x275def l2l1obj(b, d_I, d_Phi, d_x2, d_t, d_t2, lambdav, blockdim, griddim):76	(m, n) = d_Phi.shape77	(m, batch) = d_I.shape78	b.gemm('N', 'N', m, batch, n, 1, d_Phi, d_x2, 0, d_t)79	b.geam('N', 'N', m, batch, 1, d_I, -1, d_t, d_t)80 	l2 = b.nrm2(d_t.ravel(order='F'))**281 	82 	gabs[griddim, blockdim](d_x2.ravel(order='F'), d_t2)83 	84 	l1 = lambdav*b.asum(d_t2)85 	return l2 + l186@cuda.jit('void(float32[:,:], float64, float32[:,:])')87def l1prox(A, t, C):88	""" l1 Proximal operator: C = np.fmax(A-t, 0) + np.fmin(A+t, 0) 89	A: coefficients matrix (dim, batch)90	t: threshold91	C: output (dim, batch) """92	i, j = cuda.grid(2)93	if i >= A.shape[0] or j >= A.shape[1]:94		return95	if A[i, j] >= t:96		C[i, j] = A[i, j] - t 97	elif A[i, j] <= -t: 98		C[i, j] = A[i, j] + t    99	else:100		C[i, j] = 0101	return102@cuda.jit('void(float32[:], float32[:])')103def gabs(x, y):104	i = cuda.grid(1)105	if i >= x.size:106		return107	if x[i] < 0:108		y[i] = -x[i]109	else:110		y[i] = x[i]...

gpu 2.py

Source:gpu 2.py

1from __future__ import division2from numbapro import cuda3import numpy as np4import numbapro.cudalib.cublas5import numpy.random6import math7import scipy.sparse.linalg8import scipy.sparse as sps9def fista(I, Phi, lambdav, L=None, tol=10e-6, max_iterations=200, display=True, verbose=False):10	"""11	I: Images 12 	Phi: Dictionary 13 	lambdav: Sparse Penalty 14 	L = Largest eigenvalue of Phi 15 	"""16	b = numbapro.cudalib.cublas.Blas()17	(m, n) = Phi.shape18	(m, batch) = I.shape19	if L == None:20		L = scipy.sparse.linalg.svds(Phi, 1, which='LM', return_singular_vectors=False)21		print "Max eigenvalue: ." + str(L)22	L = (L**2)*2 # L = svd(Phi) -> eig(2*(Phi.T*Phi))23	invL = 1/L24	t = 1.25	if sps.issparse(Phi):26		Phi = np.array(Phi.todense())27	d_I = cuda.to_device(np.array(I, dtype=np.float32, order='F'))28	d_Phi = cuda.to_device(np.array(Phi, dtype=np.float32, order='F'))29	d_Q = cuda.device_array((n, n), dtype=np.float32, order='F')30	d_c = cuda.device_array((n, batch), dtype=np.float32, order='F')31	d_x = cuda.to_device(np.array(np.zeros((n, batch), dtype=np.float32), order='F'))32	d_y = cuda.to_device(np.array(np.zeros((n, batch), dtype=np.float32), order='F'))33	d_x2 = cuda.to_device(np.array(np.zeros((n, batch), dtype=np.float32), order='F'))34	# Temporary array variables35	d_t = cuda.device_array((m, batch), dtype=np.float32, order='F')36	d_t2 = cuda.device_array(n*batch, dtype=np.float32, order='F')37	b.gemm('T', 'N', n, n, m, 1, d_Phi, d_Phi, 0, d_Q) 	# Q = Phi^T * Phi38	b.gemm('T', 'N', n, batch, m, -2, d_Phi, d_I, 0, d_c) # c = -2*Phi^T * y39	blockdim = 32, 3240	griddim = int(math.ceil(n/blockdim[0])), int(math.ceil(batch/blockdim[1]))41	blockdim_1d = 25642	griddim_1d = int(math.ceil(n*batch/blockdim_1d))43	start = l2l1obj(b, d_I, d_Phi, d_x, d_t, d_t2, lambdav, blockdim_1d, griddim_1d)44	obj2 = start45	for i in xrange(max_iterations):46		# x2 = 2*Q*y + c47		b.symm('L', 'U', n, batch, 2, d_Q, d_y, 0, d_x2)48		b.geam('N', 'N', n, batch, 1, d_c, 1, d_x2, d_x2)49		50		# x2 = y - invL * x251		b.geam('N', 'N', n, batch, 1, d_y, -invL, d_x2, d_x2)52		# proxOp()						53		l1prox[griddim, blockdim](d_x2, invL*lambdav, d_x2)54		t2 = (1+math.sqrt(1+4*(t**2)))/2.055		56		# y = x2 + ((t-1)/t2)*(x2-x)57		b.geam('N', 'N', n, batch, 1+(t-1)/t2, d_x2, (1-t)/t2, d_x, d_y)58		# x = x259		b.geam('N', 'N', n, batch, 1, d_x2, 0, d_x, d_x)60		t = t261		# update objective62		obj = obj263		obj2 = l2l1obj(b, d_I, d_Phi, d_x2, d_t, d_t2, lambdav, blockdim_1d, griddim_1d)64		if verbose:65			x2 = d_x2.copy_to_host()66			print "L1 Objective: " + str(obj2)67		if np.abs(obj-obj2)/float(obj) < tol:68			break69	x2 = d_x2.copy_to_host()70	if display:71		print "FISTA Iterations: " + str(i)72		print "L1 Objective: " +  str(lambdav*np.sum(np.abs(x2)) + np.sum((I-Phi.dot(x2))**2))73		print "Objective delta: " + str(obj2-start)74	return x275def l2l1obj(b, d_I, d_Phi, d_x2, d_t, d_t2, lambdav, blockdim, griddim):76	(m, n) = d_Phi.shape77	(m, batch) = d_I.shape78	b.gemm('N', 'N', m, batch, n, 1, d_Phi, d_x2, 0, d_t)79	b.geam('N', 'N', m, batch, 1, d_I, -1, d_t, d_t)80 	l2 = b.nrm2(d_t.ravel(order='F'))**281 	82 	gabs[griddim, blockdim](d_x2.ravel(order='F'), d_t2)83 	84 	l1 = lambdav*b.asum(d_t2)85 	return l2 + l186@cuda.jit('void(float32[:,:], float64, float32[:,:])')87def l1prox(A, t, C):88	""" l1 Proximal operator: C = np.fmax(A-t, 0) + np.fmin(A+t, 0) 89	A: coefficients matrix (dim, batch)90	t: threshold91	C: output (dim, batch) """92	i, j = cuda.grid(2)93	if i >= A.shape[0] or j >= A.shape[1]:94		return95	if A[i, j] >= t:96		C[i, j] = A[i, j] - t 97	elif A[i, j] <= -t: 98		C[i, j] = A[i, j] + t    99	else:100		C[i, j] = 0101	return102@cuda.jit('void(float32[:], float32[:])')103def gabs(x, y):104	i = cuda.grid(1)105	if i >= x.size:106		return107	if x[i] < 0:108		y[i] = -x[i]109	else:110		y[i] = x[i]...

corrCUDA.py

Source:corrCUDA.py

1import numpy as np2import accelerate.cuda.blas as blas3import accelerate.cuda.fft as ft4from numba import cuda5def corr_td_single (x1,x2):6    c_12 = blas.dot(x1,x2)7    return c_128def best_grid_size(size, tpb):9    bpg = np.ceil(np.array(size, dtype=np.float) / tpb).astype(np.int).tolist()10    return tuple(bpg)11@cuda.jit('void(float32[:], float32[:])')12def mult_inplace(img, resp):13    i = cuda.grid(1)14    img[i] *= resp[i]15def corr_FD(x1,x2):16    threadperblock = 32, 817    blockpergrid = best_grid_size(tuple(reversed(x1.shape)), threadperblock)18    print('kernel config: %s x %s' % (blockpergrid, threadperblock))19    # Trigger initialization the cuFFT system.20    # This takes significant time for small dataset.21    # We should not be including the time wasted here22    #ft.FFTPlan(shape=x1.shape, itype=np.float32, otype=np.complex64)23    X1 = x1.astype(np.float32)24    X2 = x2.astype(np.float32)25    stream1 = cuda.stream()26    stream2 = cuda.stream()27    fftplan1 = ft.FFTPlan(shape=x1.shape, itype=np.float32,28                       otype=np.complex64, stream=stream1)29    fftplan2 = ft.FFTPlan(shape=x2.shape, itype=np.float32,30                       otype=np.complex64, stream=stream2)31    # pagelock memory32    with cuda.pinned(X1, X2):33        # We can overlap the transfer of response_complex with the forward FFT34        # on image_complex.35        d_X1 = cuda.to_device(X1, stream=stream1)36        d_X2 = cuda.to_device(X2, stream=stream2)37        fftplan1.forward(d_X1, out=d_X1)38        fftplan2.forward(d_X2, out=d_X2)39        print ('d_X1 is ',np.shape(d_X1),type(d_X1),np.max(d_X1))40        print ('d_X2 is ',np.shape(d_X2),type(d_X2),np.max(d_X2))41        stream2.synchronize()42        mult_inplace[blockpergrid, threadperblock, stream1](d_X1, d_X2)43        fftplan1.inverse(d_X1, out=d_X1)44        # implicitly synchronizes the streams45        c = d_X1.copy_to_host().real / np.prod(x1.shape)...

Automation Testing Tutorials

Learn to execute automation testing from scratch with LambdaTest Learning Hub. Right from setting up the prerequisites to run your first automation test, to following best practices and diving deeper into advanced test scenarios. LambdaTest Learning Hubs compile a list of step-by-step guides to help you be proficient with different test automation frameworks i.e. Selenium, Cypress, TestNG etc.