def test_elementwise():
A1 = np.float32(np.random.randn(100,100))
A2 = np.float32(np.random.randn(100,100))
B1 = gpu.array(A1)
B2 = gpu.array(A2)
t.assert_almost_equal(gpu.abs(B1).tocpu(), np.abs(A1), 3, "Abs not working")
#t.assert_almost_equal(gpu.log(B1).tocpu(), np.log(A1), 3, "Log not working")
#t.assert_almost_equal(gpu.sqrt(B1).tocpu(), np.sqrt(A1), 3, "Sqrt not working")
t.assert_almost_equal(gpu.pow(B1,2.0).tocpu(), np.power(A1,2.0), 3, "Pow not working")
t.assert_almost_equal(gpu.logistic(B1).tocpu(), 1.0/(1.0+np.exp(-A1)), 3, "Logistic not working")
t.assert_almost_equal(gpu.logistic_grad(B1).tocpu(), A1*(1.0-A1), 3, "Logistic grad not working")
t.assert_almost_equal(gpu.rectified_linear(B1).tocpu(), A1*(A1>0), 3, "Rectified not working")
t.assert_almost_equal(gpu.rectified_linear_grad(B1).tocpu(), (A1>0), 3, "Rectified grad not working")
t.assert_almost_equal(gpu.add(B1,B2).tocpu(), A1+A2, 3, "Add not working")
t.assert_almost_equal(gpu.sub(B1,B2).tocpu(), A1-A2, 3, "Sub not working")
t.assert_almost_equal(gpu.mul(B1,B2).tocpu(), A1*A2, 3, "Mul not working")
t.assert_almost_equal(gpu.div(B1,B2).tocpu(), A1/A2, 3, "Div not working")
t.assert_almost_equal(gpu.equal(B1,B2).tocpu(), np.equal(A1,A2), 3, "Equal not working")
t.assert_almost_equal(gpu.less(B1,B2).tocpu(), np.less(A1,A2), 3, "Less not working")
t.assert_almost_equal(gpu.greater(B1,B2).tocpu(), np.greater(A1,A2), 3, "Greater not working")
t.assert_almost_equal(gpu.greater_equal(B1,B2).tocpu(), np.greater_equal(A1,A2), 3, "Greater equal not working")
t.assert_almost_equal(gpu.less_equal(B1,B2).tocpu(), np.less_equal(A1,A2), 3, "Less equal not working")
t.assert_almost_equal(gpu.not_equal(B1,B2).tocpu(), np.not_equal(A1,A2), 3, "Not equal not working")
t.assert_almost_equal(gpu.squared_difference(B1,B2).tocpu(), (A1-A2)**2, 3, "Squared difference not working")
t.assert_almost_equal(gpu.copy(B1).tocpu(), A1.copy(), 3, "Copynot working")
def test_elementwise():
A1 = np.float32(np.random.randn(100, 100))
A2 = np.float32(np.random.randn(100, 100))
B1 = gpu.array(A1)
B2 = gpu.array(A2)
t.assert_almost_equal(
gpu.abs(B1).tocpu(), np.abs(A1), 3, "Abs not working")
#t.assert_almost_equal(gpu.log(B1).tocpu(), np.log(A1), 3, "Log not working")
#t.assert_almost_equal(gpu.sqrt(B1).tocpu(), np.sqrt(A1), 3, "Sqrt not working")
t.assert_almost_equal(
gpu.pow(B1, 2.0).tocpu(), np.power(A1, 2.0), 3, "Pow not working")
t.assert_almost_equal(
gpu.logistic(B1).tocpu(), 1.0 / (1.0 + np.exp(-A1)), 3,
"Logistic not working")
t.assert_almost_equal(
gpu.logistic_grad(B1).tocpu(), A1 * (1.0 - A1), 3,
"Logistic grad not working")
t.assert_almost_equal(
gpu.rectified_linear(B1).tocpu(), A1 * (A1 > 0), 3,
"Rectified not working")
t.assert_almost_equal(
gpu.rectified_linear_grad(B1).tocpu(), (A1 > 0), 3,
"Rectified grad not working")
t.assert_almost_equal(
gpu.add(B1, B2).tocpu(), A1 + A2, 3, "Add not working")
t.assert_almost_equal(
gpu.sub(B1, B2).tocpu(), A1 - A2, 3, "Sub not working")
t.assert_almost_equal(
gpu.mul(B1, B2).tocpu(), A1 * A2, 3, "Mul not working")
t.assert_almost_equal(
gpu.div(B1, B2).tocpu(), A1 / A2, 3, "Div not working")
t.assert_almost_equal(
gpu.equal(B1, B2).tocpu(), np.equal(A1, A2), 3, "Equal not working")
t.assert_almost_equal(
gpu.less(B1, B2).tocpu(), np.less(A1, A2), 3, "Less not working")
t.assert_almost_equal(
gpu.greater(B1, B2).tocpu(), np.greater(A1, A2), 3,
"Greater not working")
t.assert_almost_equal(
gpu.greater_equal(B1, B2).tocpu(), np.greater_equal(A1, A2), 3,
"Greater equal not working")
t.assert_almost_equal(
gpu.less_equal(B1, B2).tocpu(), np.less_equal(A1, A2), 3,
"Less equal not working")
t.assert_almost_equal(
gpu.not_equal(B1, B2).tocpu(), np.not_equal(A1, A2), 3,
"Not equal not working")
t.assert_almost_equal(
gpu.squared_difference(B1, B2).tocpu(), (A1 - A2)**2, 3,
"Squared difference not working")
t.assert_almost_equal(
gpu.copy(B1).tocpu(), A1.copy(), 3, "Copynot working")
def test_Transpose():
A = np.float32(np.random.rand(17,83))
B = gpu.array(A.T)
C = B.tocpu()
t.assert_array_equal(A.T,C,"Transpose error!")
A = np.float32(np.random.rand(100,1))
B = gpu.array(A.T)
C = B.tocpu()
t.assert_array_equal(A.T,C,"Transpose error!")
A = np.float32(np.random.rand(1,100))
B = gpu.array(A.T)
C = B.tocpu()
t.assert_array_equal(A.T,C,"Transpose error!")
def test_Transpose():
A = np.float32(np.random.rand(17, 83))
B = gpu.array(A.T)
C = B.tocpu()
t.assert_array_equal(A.T, C, "Transpose error!")
A = np.float32(np.random.rand(100, 1))
B = gpu.array(A.T)
C = B.tocpu()
t.assert_array_equal(A.T, C, "Transpose error!")
A = np.float32(np.random.rand(1, 100))
B = gpu.array(A.T)
C = B.tocpu()
t.assert_array_equal(A.T, C, "Transpose error!")
def slice_test():
A = np.float32(np.random.randn(100,100))
B = gpu.array(A)
C = gpu.slice(B,17,83,7,23).tocpu()
print( A)
t.assert_almost_equal(C, A[17:83,7:23], 3, "Slicing not working")
def slice_test():
A = np.float32(np.random.randn(100, 100))
B = gpu.array(A)
C = gpu.slice(B, 17, 83, 7, 23).tocpu()
print(A)
t.assert_almost_equal(C, A[17:83, 7:23], 3, "Slicing not working")
def test_dot():
A1 = np.float32(np.random.rand(2,4))
A2 = np.float32(np.random.rand(4,2))
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.dot(B1,B2)
C = B3.tocpu()
t.assert_array_almost_equal(np.dot(A1,A2),C,4,"array.tocpu not equal to init array!")
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.empty((2,2))
gpu.dot(B1,B2,B3)
t.assert_array_almost_equal(np.dot(A1,A2),B3.tocpu(),4,"array.tocpu not equal to init array!")
def test_dot():
A1 = np.float32(np.random.rand(2, 4))
A2 = np.float32(np.random.rand(4, 2))
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.dot(B1, B2)
C = B3.tocpu()
t.assert_array_almost_equal(np.dot(A1, A2), C, 4,
"array.tocpu not equal to init array!")
B1 = gpu.array(A1)
B2 = gpu.array(A2)
B3 = gpu.empty((2, 2))
gpu.dot(B1, B2, B3)
t.assert_array_almost_equal(np.dot(A1, A2), B3.tocpu(), 4,
"array.tocpu not equal to init array!")
def test_cpu():
A = np.ones((5,7))
B = gpu.ones((5,7))
C = B.tocpu()
t.assert_array_equal(A, C, "To GPU does not work!")
A = np.float32(np.random.rand(5,7))
B = gpu.array(A)
C = B.tocpu()
t.assert_array_equal(A, C, "To GPU does not work!")
def test_matrix_reductions():
A = np.float32(np.random.randn(100,100))
B = gpu.array(A)
C = gpu.sum(B)
t.assert_almost_equal(C, np.sum(A), 2, "Sum not working")
C = gpu.max(B)
t.assert_almost_equal(C, np.max(A), 3, "Max not working")
C = gpu.mean(B)
t.assert_almost_equal(C, np.mean(A), 3, "Max not working")
def test_matrix_reductions():
A = np.float32(np.random.randn(100, 100))
B = gpu.array(A)
C = gpu.sum(B)
t.assert_almost_equal(C, np.sum(A), 2, "Sum not working")
C = gpu.max(B)
t.assert_almost_equal(C, np.max(A), 3, "Max not working")
C = gpu.mean(B)
t.assert_almost_equal(C, np.mean(A), 3, "Max not working")
def test_cpu():
A = np.ones((5, 7))
B = gpu.ones((5, 7))
C = B.tocpu()
t.assert_array_equal(A, C, "To GPU does not work!")
A = np.float32(np.random.rand(5, 7))
B = gpu.array(A)
C = B.tocpu()
t.assert_array_equal(A, C, "To GPU does not work!")
def test_col_reductions():
A = np.float32(np.random.randn(100,110))
B = gpu.array(A)
C = gpu.col_sum(B).tocpu()
t.assert_almost_equal(C, np.sum(A,0), 3, "Colsum not working")
C = gpu.col_max(B).tocpu()
t.assert_almost_equal(C, np.max(A,0), 3, "Colmax not working")
C = gpu.col_mean(B).tocpu()
t.assert_almost_equal(C, np.mean(A,0), 3, "Colmean not working")
def test_row_reductions():
A = np.float32(np.random.randn(100, 110))
B = gpu.array(A)
C = gpu.row_sum(B).tocpu()
t.assert_almost_equal(C, np.sum(A, 1), 3, "Rowsum not working")
C = gpu.row_max(B).tocpu()
t.assert_almost_equal(C, np.max(A, 1), 3, "Rowmax not working")
C = gpu.row_mean(B).tocpu()
t.assert_almost_equal(C, np.mean(A, 1), 3, "Rowmean not working")
def test_col_reductions():
A = np.float32(np.random.randn(100, 110))
B = gpu.array(A)
C = gpu.col_sum(B).tocpu()
t.assert_almost_equal(C, np.sum(A, 0), 3, "Colsum not working")
C = gpu.col_max(B).tocpu()
t.assert_almost_equal(C, np.max(A, 0), 3, "Colmax not working")
C = gpu.col_mean(B).tocpu()
t.assert_almost_equal(C, np.mean(A, 0), 3, "Colmean not working")
def test_row_reductions():
A = np.float32(np.random.randn(100,110))
B = gpu.array(A)
C = gpu.row_sum(B).tocpu()
t.assert_almost_equal(C, np.sum(A,1), 3, "Rowsum not working")
C = gpu.row_max(B).tocpu()
t.assert_almost_equal(C, np.max(A,1), 3, "Rowmax not working")
C = gpu.row_mean(B).tocpu()
t.assert_almost_equal(C, np.mean(A,1), 3, "Rowmean not working")
def test_vectorwise():
def create_t_matrix(y, classes):
t = np.zeros((y.shape[0], classes+1))
for i in range(y.shape[0]):
t[np.int32(i), np.int32(y[i])] = 1.0
return t
A = np.float32(np.random.randn(2,4))
v = np.float32(np.random.randn(1,4))
labels = np.float32(np.random.randint(0,4,128)).reshape(128,1)
B = gpu.array(A)
V = gpu.array(v)
Y = gpu.array(labels)
t.assert_almost_equal(gpu.vector_add(B, V).tocpu(), A+v, 3, "Vec add not working")
t.assert_almost_equal(gpu.create_t_matrix(Y, 9).tocpu(), create_t_matrix(labels,9), 3, "Tmatrix not working")
def test_vectorwise():
def create_t_matrix(y, classes):
t = np.zeros((y.shape[0], classes + 1))
for i in range(y.shape[0]):
t[np.int32(i), np.int32(y[i])] = 1.0
return t
A = np.float32(np.random.randn(2, 4))
v = np.float32(np.random.randn(1, 4))
labels = np.float32(np.random.randint(0, 4, 128)).reshape(128, 1)
B = gpu.array(A)
V = gpu.array(v)
Y = gpu.array(labels)
t.assert_almost_equal(
gpu.vector_add(B, V).tocpu(), A + v, 3, "Vec add not working")
t.assert_almost_equal(
gpu.create_t_matrix(Y, 9).tocpu(), create_t_matrix(labels, 9), 3,
"Tmatrix not working")
def softmax_test():
def softmax(X):
#numerically stable softmax function
max_row_values = np.matrix(np.max(X,axis=1)).T
result = np.exp(X - max_row_values)
sums = np.matrix(np.sum(result,axis=1))
return result/sums
A = np.float32(np.random.randn(17,83))
B = gpu.array(A)
C = gpu.softmax(B).tocpu()
t.assert_almost_equal(C, softmax(A), 3, "Softmax not working")
def softmax_test():
def softmax(X):
#numerically stable softmax function
max_row_values = np.matrix(np.max(X, axis=1)).T
result = np.exp(X - max_row_values)
sums = np.matrix(np.sum(result, axis=1))
return result / sums
A = np.float32(np.random.randn(17, 83))
B = gpu.array(A)
C = gpu.softmax(B).tocpu()
t.assert_almost_equal(C, softmax(A), 3, "Softmax not working")
def test_get_view():
X = np.float32(np.random.rand(10,10))
Y = np.copy(X)
A = gpu.array(X)
B = gpu.get_view(A, rstart=0, rend=5)
t.assert_equal(X[0:5], B.tocpu(), "Get view not working!")
B = gpu.get_view(A, rstart=5, rend=10)
t.assert_equal(X[5:10], B.tocpu(), "Get view not working!")
gpu.sqrt(B, B)
t.assert_equal(Y[0:5], A.tocpu()[0:5], "Partial application to view not working!")
t.assert_equal(np.sqrt(Y[5:10]), A.tocpu()[5:10], "Partial application to view not working!")
def test_get_view():
X = np.float32(np.random.rand(10, 10))
Y = np.copy(X)
A = gpu.array(X)
B = gpu.get_view(A, rstart=0, rend=5)
t.assert_equal(X[0:5], B.tocpu(), "Get view not working!")
B = gpu.get_view(A, rstart=5, rend=10)
t.assert_equal(X[5:10], B.tocpu(), "Get view not working!")
gpu.sqrt(B, B)
t.assert_equal(Y[0:5],
A.tocpu()[0:5], "Partial application to view not working!")
t.assert_equal(np.sqrt(Y[5:10]),
A.tocpu()[5:10], "Partial application to view not working!")
def test_printmat():
X = np.float32(np.random.rand(2, 2))
A = gpu.array(X)
gpu.printmat(A)
print(X)
def test_togpu():
A = np.float32(np.random.rand(17, 83))
B = gpu.array(A)
C = B.tocpu()
t.assert_array_almost_equal(A, C, 4,
"array.tocpu not equal to init array!")
def argmax_test():
A = np.float32(np.random.randn(123, 83))
B = gpu.array(A)
C = gpu.argmax(B).tocpu()
t.assert_almost_equal(C, np.argmax(A, 1), 3, "Softmax not working")
def argmax_test():
A = np.float32(np.random.randn(123,83))
B = gpu.array(A)
C = gpu.argmax(B).tocpu()
t.assert_almost_equal(C, np.argmax(A,1), 3, "Softmax not working")
def test_togpu():
A = np.float32(np.random.rand(17,83))
B = gpu.array(A)
C = B.tocpu()
t.assert_array_almost_equal(A,C,4,"array.tocpu not equal to init array!")
def test_printmat():
X = np.float32(np.random.rand(2,2))
A = gpu.array(X)
gpu.printmat(A)
print( X)
