Fixing the backpath test for softmax that was accidentially dropped before.
Moving log_plus test to end of file
This commit is contained in:
Thilo Will
Родитель
220ac16c31
Коммит
0684fcfbd1
|
|
@ -112,42 +112,6 @@ def test_op_sigmoid(tensor, device_id, precision):
|
|||
precision=precision, clean_up=True, backward_pass=True,
|
||||
input_node=input_node)
|
||||
|
||||
LOG_PLUS_TESTCASES = [
|
||||
([ 1], [ 2], [[[m.log(m.exp( 1)+ m.exp( 2))]]], [[[m.exp( 1) / (m.exp( 1)+ m.exp( 2))]]], [[[m.exp( 2) / (m.exp( 1)+ m.exp( 2))]]]), # test case: first argument < seond argument
|
||||
([ -1], [ -2], [[[m.log(m.exp(-1)+ m.exp(-2))]]], [[[m.exp(-1) / (m.exp(-1)+ m.exp(-2))]]], [[[m.exp(-2) / (m.exp(-1)+ m.exp(-2))]]]), # test case: second argument > first argument
|
||||
([ 0], [100000], [[[100000]]], [[[0]]], [[[1]]]), # test case: check that we don't have overflow
|
||||
([ 100000], [ 0], [[[100000]]], [[[1]]], [[[0]]]), # test case: check that we don't have overflow
|
||||
([ 100000], [ 0,0], [[[100000, 100000]]], [[[2]]], [[[0,0]]]), # test case: broadcasting. Note the result for grad_x is two because of reduction in backward path
|
||||
]
|
||||
|
||||
@pytest.mark.parametrize("x, y, expected, grad_x, grad_y", LOG_PLUS_TESTCASES)
|
||||
def test_op_log_plus(x, y, expected, grad_x, grad_y, device_id, precision):
|
||||
|
||||
from .. import log_plus
|
||||
|
||||
# Forward pass test
|
||||
# ==================
|
||||
|
||||
a = I([x])
|
||||
b = I([y])
|
||||
|
||||
unittest_helper(log_plus(a,b), None, expected,
|
||||
device_id=device_id,
|
||||
precision=precision,
|
||||
clean_up=True, backward_pass=False)
|
||||
|
||||
# Backward pass tests
|
||||
# ==================
|
||||
op_node_a = log_plus(a, y)
|
||||
op_node_b = log_plus(x, b)
|
||||
|
||||
unittest_helper(op_node_a, None, grad_x, device_id=device_id,
|
||||
precision=precision, clean_up=True, backward_pass=True, input_node=a)
|
||||
|
||||
unittest_helper(op_node_b, None, grad_y, device_id=device_id,
|
||||
precision=precision, clean_up=True, backward_pass=True, input_node=b)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("batch",
|
||||
[
|
||||
[ # 2 samples having 4 classes
|
||||
|
|
@ -184,6 +148,29 @@ def test_op_softmax(batch, device_id, precision):
|
|||
precision=precision,
|
||||
clean_up=True, backward_pass=False)
|
||||
|
||||
# Backward pass test
|
||||
# ==================
|
||||
# The expected results for the backward pass is fi(1-fi) for i and -fi*fj
|
||||
# for element j!=i.
|
||||
def numpy_grad(x):
|
||||
grads = np.zeros((len(x), len(x)), dtype=PRECISION_TO_TYPE[precision])
|
||||
|
||||
for i in range(len(x)):
|
||||
# deriving wrt i-th element
|
||||
for j in range(len(x)):
|
||||
if i == j:
|
||||
grads[i, j] = x[i] * (1 - x[i])
|
||||
else:
|
||||
grads[i, j] = x[i] * (-x[j])
|
||||
|
||||
return grads.sum(axis=0)
|
||||
|
||||
expected = [[numpy_grad(numpy_op(sample))] for sample in batch]
|
||||
|
||||
unittest_helper(op_node, None, expected,
|
||||
device_id=device_id,
|
||||
precision=precision, clean_up=True, backward_pass=True,
|
||||
input_node=input_node)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("tensor", TENSORS)
|
||||
|
|
@ -511,3 +498,42 @@ def test_op_cond(flag, value_a, value_b, device_id, precision):
|
|||
result = cond(cond_as_const, value_a_as_const, value_b_as_input)
|
||||
unittest_helper(result, None, expected, device_id=device_id,
|
||||
precision=precision, clean_up=True, backward_pass=True, input_node=value_b_as_input)
|
||||
|
||||
|
||||
|
||||
LOG_PLUS_TESTCASES = [
|
||||
([ 1], [ 2], [[[m.log(m.exp( 1)+ m.exp( 2))]]], [[[m.exp( 1) / (m.exp( 1)+ m.exp( 2))]]], [[[m.exp( 2) / (m.exp( 1)+ m.exp( 2))]]]), # test case: first argument < seond argument
|
||||
([ -1], [ -2], [[[m.log(m.exp(-1)+ m.exp(-2))]]], [[[m.exp(-1) / (m.exp(-1)+ m.exp(-2))]]], [[[m.exp(-2) / (m.exp(-1)+ m.exp(-2))]]]), # test case: second argument > first argument
|
||||
([ 0], [100000], [[[100000]]], [[[0]]], [[[1]]]), # test case: check that we don't have overflow
|
||||
([ 100000], [ 0], [[[100000]]], [[[1]]], [[[0]]]), # test case: check that we don't have overflow
|
||||
([ 100000], [ 0,0], [[[100000, 100000]]], [[[2]]], [[[0,0]]]), # test case: broadcasting. Note the result for grad_x is two because of reduction in backward path
|
||||
]
|
||||
|
||||
@pytest.mark.parametrize("x, y, expected, grad_x, grad_y", LOG_PLUS_TESTCASES)
|
||||
def test_op_log_plus(x, y, expected, grad_x, grad_y, device_id, precision):
|
||||
|
||||
from .. import log_plus
|
||||
|
||||
# Forward pass test
|
||||
# ==================
|
||||
|
||||
a = I([x])
|
||||
b = I([y])
|
||||
|
||||
unittest_helper(log_plus(a,b), None, expected,
|
||||
device_id=device_id,
|
||||
precision=precision,
|
||||
clean_up=True, backward_pass=False)
|
||||
|
||||
# Backward pass tests
|
||||
# ==================
|
||||
op_node_a = log_plus(a, y)
|
||||
op_node_b = log_plus(x, b)
|
||||
|
||||
unittest_helper(op_node_a, None, grad_x, device_id=device_id,
|
||||
precision=precision, clean_up=True, backward_pass=True, input_node=a)
|
||||
|
||||
unittest_helper(op_node_b, None, grad_y, device_id=device_id,
|
||||
precision=precision, clean_up=True, backward_pass=True, input_node=b)
|
||||
|
||||
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче