kshitij12345 commented on a change in pull request #15611: [MXNET-978] n-th
order gradient test support
URL: https://github.com/apache/incubator-mxnet/pull/15611#discussion_r306132123
##########
File path: tests/python/unittest/test_higher_order_grad.py
##########
@@ -148,30 +165,48 @@ def grad_grad_op(x):
shape = rand_shape_nd(dim)
array = random_arrays(shape)
check_second_order_unary(array, sigmoid, grad_grad_op)
+ # TODO(kshitij12345): Remove
+ check_nth_order_unary(array, sigmoid, [grad_op, grad_grad_op], [1, 2])
+ check_nth_order_unary(array, sigmoid, grad_grad_op, 2)
def check_second_order_unary(x, op, grad_grad_op):
+ check_nth_order_unary(x, op, grad_grad_op, 2)
+
+
+def check_nth_order_unary(x, op, grad_ops, orders):
+ if isinstance(orders, int):
+ orders = [orders]
+ grad_ops = [grad_ops]
+
x = nd.array(x)
- grad_grad_x = grad_grad_op(x)
x.attach_grad()
- # Manual head_grads.
- y_grad = nd.random.normal(shape=x.shape)
- head_grad_grads = nd.random.normal(shape=x.shape)
+ order = max(orders)
+ expected_grads = [grad_op(x) for grad_op in grad_ops]
+ computed_grads = []
+ head_grads = []
# Perform compute.
with autograd.record():
y = op(x)
- x_grad = autograd.grad(heads=y, variables=x, head_grads=y_grad,
- create_graph=True, retain_graph=True)[0]
- x_grad.backward(head_grad_grads)
-
- # Compute expected values.
- expected_grad_grad = grad_grad_x.asnumpy() * head_grad_grads.asnumpy() * \
- y_grad.asnumpy()
-
- # Validate the gradients.
- assert_almost_equal(expected_grad_grad, x.grad.asnumpy())
+ for current_order in range(1, order+1):
+ head_grad = nd.random.normal(shape=x.shape)
+ y = autograd.grad(heads=y, variables=x, head_grads=head_grad,
Review comment:
One option I can think of is, at the start of for loop we'll have
`computed_grad = y` (which is deceiving) and replace `y` by `computed_grad` in
the for loop.
Up for suggestions.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to 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
