Update tests to ensure that Torch backpropagation is supported #1598
Conversation
josh146
added
the
review-ready 👌
Codecov Report
@@ Coverage Diff @@
## master #1598 +/- ##
=======================================
Coverage 99.13% 99.13%
=======================================
Files 195 195
Lines 14104 14104
=======================================
Hits 13982 13982
Misses 122 122 Continue to review full report at Codecov.
|
There was a problem hiding this comment.
Looks good to me! 💯 Thank you for catching these @josh146 🎉
| @@ -2406,6 +2406,25 @@ def circuit(): | |||
| # evaluated one expval altogether | |||
| assert spy.call_count == 1 | |||
|
|
|||
| def test_do_not_split_analytic_torch(self, mocker): | |||
There was a problem hiding this comment.
Perhaps out of scope and an overall organizational question: wouldn't these tests belong to the dedicated files for each backprop device? So e.g., in test_default_qubit_torch.py?
There was a problem hiding this comment.
Yes, good point! The other tests are already in this file, however, so it would likely be out of scope to move all tests here.
| @@ -81,6 +88,9 @@ def circuit(a): | |||
|
|
|||
| def test_execution_with_interface(self, dev_name, diff_method): | |||
| """Test execution works with the interface""" | |||
| if diff_method == "backprop": | |||
There was a problem hiding this comment.
How come this case doesn't support backprop?
There was a problem hiding this comment.
Because this test explicitly tests the interface, but backprop requires/has no interface :)
|
|
||
| def test_jacobian_dtype(self, dev_name, diff_method, tol): | ||
| """Test calculating the jacobian with a different datatype""" | ||
| if diff_method == "backprop": | ||
| pytest.skip("Test does not support backprop") |
There was a problem hiding this comment.
Also curious: how come this doesn't work in backprop?
There was a problem hiding this comment.
Same reason as above; setting the dtype is a feature of the interface, however backprop does not use the interface
Co-authored-by: antalszava <antalszava@gmail.com>
Context: We now support backpropagation using Torch and
default.qubit, however a lot of our integration tests skipped this combination.Description of the Change: Enables testing this combination across our test suite.
Benefits: Better test coverage.
Possible Drawbacks: There is an odd issue I noticed;
return qml.probs(0), qml.probs(1)returns a(2, 2)shape array in TensorFlow, but a(4,)array in Torch+backprop. This could be due to differing behaviour in theDefaultQubitTorch._asarray()method.Related GitHub Issues: n/a