anirudh2290 closed pull request #12660: fix bug in prelu , issue 12061
URL: https://github.com/apache/incubator-mxnet/pull/12660
This is a PR merged from a forked repository.
As GitHub hides the original diff on merge, it is displayed below for
the sake of provenance:
As this is a foreign pull request (from a fork), the diff is supplied
below (as it won't show otherwise due to GitHub magic):
diff --git a/src/operator/leaky_relu-inl.h b/src/operator/leaky_relu-inl.h
index 1d2baa4b6c3..fe2668959af 100644
--- a/src/operator/leaky_relu-inl.h
+++ b/src/operator/leaky_relu-inl.h
@@ -301,7 +301,7 @@ class LeakyReLUOp : public Operator {
TShape result(dst.ndim());
int s = src.ndim() - 1;
for (int i = dst.ndim() - 1; i >= 0; i--) {
- if (s >= 0 && (dst[i] == src[s] || src[s] == 1)) {
+ if (s >= 0 && i <= 1 && (dst[i] == src[s] || src[s] == 1)) {
result[i] = src[s];
s--;
} else {
diff --git a/tests/python/unittest/test_operator.py
b/tests/python/unittest/test_operator.py
index 43c357808f1..6e1d1af01cb 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -772,48 +772,61 @@ def test_prelu():
def fprelu(x, gamma):
pos_indices = x > 0
out = x.copy()
- out = np.multiply(out, gamma)
+ if len(x.shape) == 4:
+ out = out.transpose(2,3,0,1)
+ out = np.multiply(out, gamma)
+ out = out.transpose(2,3,0,1)
+ else:
+ out = np.multiply(out, gamma)
out[pos_indices] = x[pos_indices]
return out
def fprelu_grad(x, y, gamma):
pos_indices = x > 0
- grad_x = np.multiply(np.ones(x.shape), gamma)
+ if len(x.shape) == 4:
+ grad_x = np.multiply(np.ones(x.shape).transpose(2,3,0,1), gamma)
+ grad_x = grad_x.transpose(2,3,0,1)
+ else:
+ grad_x = np.multiply(np.ones(x.shape), gamma)
grad_gam = np.zeros(gamma.shape)
copy_x = x.copy()
copy_x[pos_indices] = 0.0
grad_x[pos_indices] = 1.0
- if len(gamma.shape) > 1:
+ if len(gamma.shape) > 1 and len(x.shape) != 4:
grad_gam = copy_x
+ elif len(gamma.shape) > 1 and len(x.shape) == 4:
+ grad_gam = np.sum(copy_x, axis=(2,3))
elif gamma.shape[0] == 1:
grad_gam = np.sum(np.sum(copy_x))
- elif gamma.shape[0] > 1:
+ elif gamma.shape[0] > 1 and len(x.shape) != 4:
grad_gam = np.sum(copy_x, axis=0)
+ elif gamma.shape[0] > 1 and len(x.shape) == 4:
+ grad_gam = np.sum(copy_x, axis=(0,2,3))
return (grad_x, grad_gam)
- shape = (3,4)
x = mx.symbol.Variable("x")
gamma = mx.symbol.Variable("gamma")
- for dtype in [np.float16, np.float32, np.float64]:
- for gam in [np.array([0.1, 0.2, 0.3, 0.4], dtype=dtype)]:
- gam_full = np.array([gam, gam, gam])
- xa = np.random.uniform(low=-1.0,high=1.0,size=shape).astype(dtype)
- rtol = 1e-2
- atol = 1e-3
- eps = 1e-4
- xa[abs(xa) < eps] = 1.0
- y = mx.symbol.LeakyReLU(data=x, gamma=gamma, act_type='prelu')
- ya = fprelu(xa, gam)
- ya_full = fprelu(xa, gam_full)
- g_xa, g_gam = fprelu_grad(xa, ya, gamma=gam)
- g_xa_full, g_gam_full = fprelu_grad(xa, ya_full, gamma=gam_full)
- # Skip numeric check for float16 type to get rid of flaky behavior
- if dtype is not np.float16:
- check_numeric_gradient(y, [xa, gam], numeric_eps=eps,
rtol=rtol, atol=atol, dtype=dtype)
- check_numeric_gradient(y, [xa, gam_full], numeric_eps=eps,
rtol=rtol, atol=atol, dtype=dtype)
- check_symbolic_forward(y, [xa, gam], [ya], rtol=rtol, atol=atol,
dtype=dtype)
- check_symbolic_backward(y, [xa, gam], [np.ones(shape),
np.ones(gam.shape)], [g_xa, g_gam], rtol=rtol, atol=atol, dtype=dtype)
- check_symbolic_forward(y, [xa, gam_full], [ya_full], rtol=rtol,
atol=atol, dtype=dtype)
- check_symbolic_backward(y, [xa, gam_full], [np.ones(shape),
np.ones(gam_full.shape)],
- [g_xa_full, g_gam_full], rtol=rtol,
atol=atol, dtype=dtype)
+ for shape in [(3,4), (3,4,4,5)]:
+ for dtype in [np.float16, np.float32, np.float64]:
+ for gam in [np.array([0.1, 0.2, 0.3, 0.4], dtype=dtype)]:
+ gam_full = np.array([gam, gam, gam])
+ xa =
np.random.uniform(low=-1.0,high=1.0,size=shape).astype(dtype)
+ rtol = 1e-2
+ atol = 1e-3
+ eps = 1e-4
+ xa[abs(xa) < eps] = 1.0
+ y = mx.symbol.LeakyReLU(data=x, gamma=gamma, act_type='prelu')
+ ya = fprelu(xa, gam)
+ ya_full = fprelu(xa, gam_full)
+ g_xa, g_gam = fprelu_grad(xa, ya, gamma=gam)
+ g_xa_full, g_gam_full = fprelu_grad(xa, ya_full,
gamma=gam_full)
+ # Skip numeric check for float16 type to get rid of flaky
behavior
+ if dtype is not np.float16:
+ check_numeric_gradient(y, [xa, gam], numeric_eps=eps,
rtol=rtol, atol=atol, dtype=dtype)
+ check_numeric_gradient(y, [xa, gam_full], numeric_eps=eps,
rtol=rtol, atol=atol, dtype=dtype)
+ check_symbolic_forward(y, [xa, gam], [ya], rtol=rtol,
atol=atol, dtype=dtype)
+ check_symbolic_backward(y, [xa, gam], [np.ones(shape),
np.ones(gam.shape)], [g_xa, g_gam], rtol=rtol, atol=atol, dtype=dtype)
+ check_symbolic_forward(y, [xa, gam_full], [ya_full],
rtol=rtol, atol=atol, dtype=dtype)
+ check_symbolic_backward(y, [xa, gam_full], [np.ones(shape),
np.ones(gam_full.shape)],
+ [g_xa_full, g_gam_full], rtol=rtol,
atol=atol, dtype=dtype)
@with_seed()
def test_selu():
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
