def RandPosdefSpMatrix(dim):
"""
Generate random (non-singular) matrix
Arguments:
dim - int
matrix_sqrt - TpMatrix
logdet - float
Outputs:
matrix - SpMatrix
"""
while True:
tmp = Matrix(dim, dim)
tmp.set_randn_()
if tmp.cond() < 100: break
print(
"Condition number of random matrix large {}, trying again (this is normal)"
.format(tmp.cond()))
# tmp * tmp^T will give positive definite matrix
matrix = SpMatrix(dim)
matrix.add_mat2_(1.0, tmp, MatrixTransposeType.NO_TRANS, 0.0)
matrix_sqrt = TpMatrix(len(matrix))
matrix_sqrt = matrix_sqrt.cholesky_(matrix)
logdet_out = matrix.log_pos_def_det()
return matrix, matrix_sqrt, logdet_out
def test_nnet_decodable(self):
gen_config = NnetGenerationOptions()
configs = generate_config_sequence(gen_config)
nnet = Nnet()
for j, config in enumerate(configs):
print("Input config[{}]:".format(j))
print(config)
istrm = istringstream.from_str(config)
nnet.read_config(istrm)
num_frames = 5 + random.randint(1, 100)
input_dim = nnet.input_dim("input")
output_dim = nnet.output_dim("output")
ivector_dim = max(0, nnet.input_dim("ivector"))
input = Matrix(num_frames, input_dim)
set_batchnorm_test_mode(True, nnet)
set_dropout_test_mode(True, nnet)
input.set_randn_()
ivector = Vector(ivector_dim)
ivector.set_randn_()
priors = Vector(output_dim if random.choice([True, False]) else 0)
if len(priors) != 0:
priors.set_randn_()
priors.apply_exp_()
output1 = Matrix(num_frames, output_dim)
output2 = Matrix(num_frames, output_dim)
opts = NnetSimpleComputationOptions()
opts.frames_per_chunk = random.randint(5, 25)
compiler = CachingOptimizingCompiler(nnet)
decodable = DecodableNnetSimple(opts, nnet, priors, input, compiler,
ivector if ivector_dim else None)
for t in range(num_frames):
decodable.get_output_for_frame(t, output1[t])
opts = NnetSimpleLoopedComputationOptions()
info = DecodableNnetSimpleLoopedInfo.new_from_priors(
opts, priors, nnet)
decodable = DecodableNnetSimpleLooped(info, input,
ivector if ivector_dim else None)
for t in range(num_frames):
decodable.get_output_for_frame(t, output2[t])
if (not nnet_is_recurrent(nnet)
and nnet.info().find("statistics-extraction") == -1
and nnet.info().find("TimeHeightConvolutionComponent") == -1):
for t in range(num_frames):
self.assertTrue(approx_equal(output1[t], output2[t]))
def test__init__(self):
m = Matrix()
sb = SubMatrix(m)
m = Matrix(5, 5)
sb = SubMatrix(m)
for i in range(100):
m.set_randn_()
self.assertAlmostEqual(m.sum(), sb.sum())
m = DoubleMatrix()
sb = SubMatrix(m)
def testcopy_from_mat(self):
for i in range(10):
rows, cols = 10 * i, 5 * i
A = Matrix(rows, cols)
A.set_randn_()
B = CuMatrix.new_from_size(*A.shape)
B.copy_from_mat(A)
self.assertAlmostEqual(A.sum(), B.sum(), places=4)
A = CuMatrix.new_from_size(rows, cols)
A.set_randn()
B = CuMatrix.new_from_size(rows, cols)
B.copy_from_cu_mat(A)
self.assertAlmostEqual(A.sum(), B.sum(), places=4)