def gradient_numba(result_1d, result_indrow_1d, result_indcol_1d, x_shape0,
x_shape1, x_data, x_indices, x_indptr, xw_1d, y_ind_1d,
y_indptr_1d, y_shape_2d):
sp_op.nonzero_numba(result_indrow_1d, result_indcol_1d, y_ind_1d,
y_indptr_1d, np.max(y_shape_2d),
np.ones((1), dtype=np.int64)[0])
# A, B = sp_op.nonzero(y)
# assert np.array_equal(A, result_indrow_1d)
# assert np.array_equal(B, result_indcol_1d)
row_vector = sigmoid(xw_1d)
row_vector[result_indrow_1d] -= 1
# result = sp_op.mult_col_raw_col_row_sum_raw(X, row_vector)
result_data = x_data.copy()
result_row = np.zeros(x_data.shape[0], dtype=np.int64)
result_col = np.zeros(x_data.shape[0], dtype=np.int64)
sp_op.mult_col_raw_col_row_sum_raw_numba(result_1d, row_vector.ravel(),
x_indptr, x_indices, result_data,
result_row, result_col, x_shape0,
x_shape1)
# assert np.array_equal(result, result_1d.reshape((result_1d.shape[0], 1)))
return result_1d
def gradient_sp_old(X, W, y):
"""
Gradient of log_likelihood optimised for sparse matrices.
@param X: Training examples
@param W: Weight vector
@param y: True Categories of the training examples X
@return: Gradient
"""
# sigm(XW) - y
sdotp = sigmoid(X.dot(W))
if y.nnz != 0:
# Originally it is required to compute
# s = -1 ^ (1 - y_n)
#
# which for : y_n = 1 -> s = 1
# and for : y_n = 0 -> s = -1
# Because y[ind] = 1, iff ind = y.nonzero()
resultrow, resultcol = sp_op.nonzero(y)
sdotp[resultrow] -= 1
# (sigm(XW) - y) * X,T
# data, row, col = mult_row_raw(X, sdotp)
result = sp_op.mult_col_raw_col_row_sum_raw(X, sdotp)
assert result.shape[1] == W.shape[1]
return result
def sigmoidOfList(numberList): outputList = [] while isEmpty(numberList) == False: e = numberList.pop(0) s = sigmoid(e) outputList.append(s) return outputList
def gradient(X, W, y):
"""
Gradient of log_likelihood
@param X: Training examples
@param W: Weight vector
@param y: True Categories of the training examples X
@return: Gradient
"""
if scipy.sparse.issparse(X) or scipy.sparse.issparse(
W) or scipy.sparse.issparse(y):
print("WARN : Use gradient_sp for sparse matrices")
return gradient_sp(X, W, y)
sig = (sigmoid(np.dot(X, W))).T - y
assert sig.shape == y.shape
inss = sig * X.T
assert inss.shape == X.T.shape
result = np.sum(inss, axis=1)
assert result.shape[0] == inss.shape[0]
return result
from mathutil import sigmoid print sigmoid(2.0) from mathutil import findQuadraticRoots findQuadraticRoots(2,5,0.5) findQuadraticRoots(1,1,1) findQuadraticRoots(1.0,-2.0,1.0) from mathutil import arsinh print arsinh(-4)
def sigmoidOfListV2(numberList): outputList = [] for e in numberList: s = sigmoid(e) outputList.append(s) return outputList