anirudhacharya commented on a change in pull request #11209: [MXNET-536]
implement var/std operators
URL: https://github.com/apache/incubator-mxnet/pull/11209#discussion_r217932556
##########
File path: tests/python/unittest/test_operator.py
##########
@@ -5702,6 +5702,98 @@ def test_softmax():
check_softmax_grad(default_context())
check_smoothed_softmax_grad(default_context())
+@with_seed()
+def test_variance():
+ def true_var(x, axis=None):
+ if len(x.shape) == 1:
+ return np.var(x, keepdims=True, axis=axis)
+ else:
+ return np.var(x, axis=axis)
+ def true_var_grad(x, ograd, axis=None):
+ if axis is None:
+ return 2 * (x - np.mean(x)) * ograd / np.prod(x.shape)
+ else:
+ denom = x.shape[axis]
+ ograd_shape = tuple(1 if i == axis else k for i, k in enumerate(x.shape))
+ return 2 * (x - np.mean(x, axis=axis, keepdims=True)) *
ograd.reshape(ograd_shape) / denom
+
+ for ndim in range(1, 6):
+ # check forward
+ shape = rand_shape_nd(ndim, 5)
+ data = rand_ndarray(shape=shape, stype='default')
+ data_np = data.asnumpy()
+ expected = true_var(data_np)
+ output = mx.nd.variance(data)
+ assert_almost_equal(output.asnumpy(), expected)
+ for axis in range(ndim):
+ expected = true_var(data_np, axis=axis)
+ output = mx.nd.variance(data, axis=axis)
+ assert_almost_equal(output.asnumpy(), expected)
+
+ # check backward
+ data = mx.sym.Variable('data')
+ var_sym = mx.sym.variance(data=data)
+ check_numeric_gradient(var_sym, [data_np], atol=1e-3)
+ ograd = np.random.random(size=(1,))
+ check_symbolic_backward(var_sym, [data_np], [ograd],
+ [true_var_grad(data_np, ograd)], atol=1e-6)
+ for axis in range(ndim):
+ data = mx.sym.Variable('data')
+ var_sym = mx.sym.variance(data=data, axis=axis)
+ check_numeric_gradient(var_sym, [data_np], atol=1e-3)
+ if ndim == 1:
+ ograd = np.random.random(size=(1,))
+ else:
+ ograd = np.random.random(size=tuple(x for i, x in
enumerate(data_np.shape) if i != axis))
+ check_symbolic_backward(var_sym, [data_np], [ograd],
+ [true_var_grad(data_np, ograd, axis=axis)], atol=1e-6)
+
+@with_seed()
+def test_std():
+ np.set_printoptions(threshold=np.nan)
+ def true_std(x, axis=None):
+ if len(x.shape) == 1:
+ return np.std(x, keepdims=True, axis=axis)
+ else:
+ return np.std(x, axis=axis)
+ def true_std_grad(x, ograd, axis=None):
+ if axis is None:
+ return (x - np.mean(x)) * ograd / true_std(x) / np.prod(x.shape)
+ else:
+ denom = x.shape[axis]
+ ograd_shape = tuple(1 if i == axis else k for i, k in enumerate(x.shape))
+ return (x - np.mean(x, axis=axis, keepdims=True)) *
ograd.reshape(ograd_shape) / true_std(x, axis=axis).reshape(ograd_shape) / denom
+
+ for ndim in range(1, 6):
+ # check forward
+ shape = rand_shape_nd(ndim, 4)
+ while not np.all(np.array(shape) > 1): # avoid length 1 in all dimension
+ shape = rand_shape_nd(ndim, 4)
+ data = rand_ndarray(shape=shape, stype='default')
+ data_np = data.asnumpy()
+ expected = true_std(data_np)
+ output = mx.nd.std(data)
+ assert_almost_equal(output.asnumpy(), expected)
+ for axis in range(ndim):
+ expected = true_std(data_np, axis=axis)
+ output = mx.nd.std(data, axis=axis)
Review comment:
please add a test for tuple values of axis.
----------------------------------------------------------------
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
