`test_hmc_mcse` unit test suddenly failing to pass due to `aesara.gradient.DisconnectedInputError` error.
zoj613 opened this issue · comments
Description of your problem or feature request
CI tests are failing due to the tests for test_hmc_mcse raising a aesara.gradient.DisconnectedInputError exception.
Please provide a minimal, self-contained, and reproducible example.
python -m pytest tests/test_hmc.py::test_hmc_mcse -vPlease provide the full traceback of any errors.
tests/test_hmc.py:203:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
../../micromamba/envs/aehmc-dev/lib/python3.11/site-packages/aesara/scan/basic.py:864: in scan
raw_inner_outputs = fn(*args)
aehmc/hmc.py:132: in step
), updates = proposal_generator(
aehmc/hmc.py:222: in propose
), updates = integrate(q, p, potential_energy, potential_energy_grad, step_size)
aehmc/trajectory.py:92: in integrate
[q, p, energy, energy_grad], updates = aesara.scan(
../../micromamba/envs/aehmc-dev/lib/python3.11/site-packages/aesara/scan/basic.py:864: in scan
raw_inner_outputs = fn(*args)
aehmc/trajectory.py:87: in one_step
new_state = integrator(
aehmc/integrators.py:68: in one_step
potential_energy_grad = aesara.grad(potential_energy, position)
../../micromamba/envs/aehmc-dev/lib/python3.11/site-packages/aesara/gradient.py:608: in grad
handle_disconnected(elem)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
var = Elemwise{add,no_inplace}.0
def handle_disconnected(var):
message = (
"grad method was asked to compute the gradient "
"with respect to a variable that is not part of "
"the computational graph of the cost, or is used "
f"only by a non-differentiable operator: {var}"
)
if disconnected_inputs == "ignore":
pass
elif disconnected_inputs == "warn":
warnings.warn(message, stacklevel=2)
elif disconnected_inputs == "raise":
message = utils.get_variable_trace_string(var)
> raise DisconnectedInputError(message)
E aesara.gradient.DisconnectedInputError:
E Backtrace when that variable is created:
E
E File "/home/zoj/dev/aehmc/tests/test_hmc.py", line 203, in test_hmc_mcse
E trajectory, updates = aesara.scan(
E File "/home/zoj/micromamba/envs/aehmc-dev/lib/python3.11/site-packages/aesara/scan/basic.py", line 864, in scan
E raw_inner_outputs = fn(*args)
E File "/home/zoj/dev/aehmc/aehmc/hmc.py", line 132, in step
E ), updates = proposal_generator(
E File "/home/zoj/dev/aehmc/aehmc/hmc.py", line 222, in propose
E ), updates = integrate(q, p, potential_energy, potential_energy_grad, step_size)
E File "/home/zoj/dev/aehmc/aehmc/trajectory.py", line 92, in integrate
E [q, p, energy, energy_grad], updates = aesara.scan(
E File "/home/zoj/micromamba/envs/aehmc-dev/lib/python3.11/site-packages/aesara/scan/basic.py", line 864, in scan
E raw_inner_outputs = fn(*args)
E File "/home/zoj/dev/aehmc/aehmc/trajectory.py", line 87, in one_step
E new_state = integrator(
E File "/home/zoj/dev/aehmc/aehmc/integrators.py", line 65, in one_step
E position = position + a2 * step_size * kinetic_grad
../../micromamba/envs/aehmc-dev/lib/python3.11/site-packages/aesara/gradient.py:594: DisconnectedInputError
Versions and main components
- Aesara version: 2.8.12
- Aesara config (
python -c "import aesara; print(aesara.config)") - Python version: 3.11
- Operating system: Linux
- How did you install Aesara: conda
This appears to be due to the position term in https://github.com/aesara-devs/aehmc/blob/main/aehmc/integrators.py#L68 no longer being present in the potential_energy term.
Here's a graph for those two terms:
>>> aesara.dprint([position, potential_energy])
Elemwise{add,no_inplace} [id A]
|<TensorType(float64, (2,))> [id B]
|Elemwise{add,no_inplace} [id C]
|Elemwise{mul,no_inplace} [id D]
| |TensorConstant{(1,) of 0.5} [id E]
| |TensorConstant{(1,) of 1.0} [id F]
| |Elemwise{sub,no_inplace} [id G]
| | |<TensorType(float64, (2,))> [id H]
| | |Elemwise{mul,no_inplace} [id I]
| | |TensorConstant{(1,) of 0.5} [id E]
| | |<TensorType(float64, (2,))> [id J]
| |TensorConstant{[1. 2.]} [id K]
|Elemwise{mul,no_inplace} [id L]
|TensorConstant{(1,) of 0.5} [id E]
|TensorConstant{(1,) of 1.0} [id F]
|TensorConstant{[1. 2.]} [id K]
|Elemwise{sub,no_inplace} [id G]
Elemwise{neg,no_inplace} [id M]
|Sum{acc_dtype=float64} [id N]
|Check{0 < diag(Sigma)} [id O]
|Elemwise{sub,no_inplace} [id P]
| |Elemwise{sub,no_inplace} [id Q]
| | |Elemwise{mul,no_inplace} [id R]
| | | |Elemwise{mul,no_inplace} [id S]
| | | | |TensorConstant{-0.5} [id T]
| | | | |Subtensor{int64} [id U]
| | | | |TensorConstant{(1,) of 2} [id V]
| | | | |ScalarConstant{-1} [id W]
| | | |Elemwise{log,no_inplace} [id X]
| | | |TensorConstant{6.283185307179586} [id Y]
| | |Elemwise{mul,no_inplace} [id Z]
| | |TensorConstant{0.5} [id BA]
| | |Sum{axis=[0], acc_dtype=float64} [id BB]
| | |Elemwise{pow,no_inplace} [id BC]
| | |InplaceDimShuffle{0} [id BD]
| | | |SolveTriangular{lower=True, trans=0, unit_diagonal=False, check_finite=True} [id BE]
| | | |Elemwise{switch,no_inplace} [id BF]
| | | | |InplaceDimShuffle{x,x} [id BG]
| | | | | |All [id BH]
| | | | | |Elemwise{gt,no_inplace} [id BI]
| | | | | |ExtractDiag{offset=0, axis1=0, axis2=1, view=False} [id BJ]
| | | | | | |Cholesky{lower=True, destructive=False, on_error='nan'} [id BK]
| | | | | | |TensorConstant{[[1. 1.]
[1. 4.]]} [id BL]
| | | | | |InplaceDimShuffle{x} [id BM]
| | | | | |TensorConstant{0} [id BN]
| | | | |Cholesky{lower=True, destructive=False, on_error='nan'} [id BK]
| | | | |InplaceDimShuffle{x,x} [id BO]
| | | | |TensorConstant{1} [id BP]
| | | |InplaceDimShuffle{0} [id BQ]
| | | |Elemwise{sub,no_inplace} [id BR]
| | | |Elemwise{add,no_inplace} [id BS]
| | | | |<TensorType(float64, (2,))> [id B]
| | | | |Elemwise{mul,no_inplace} [id BT]
| | | | | |TensorConstant{(1,) of 0.5} [id E]
| | | | | |Elemwise{sub,no_inplace} [id G]
| | | | | |TensorConstant{[1. 2.]} [id K]
| | | | |Elemwise{mul,no_inplace} [id BU]
| | | | |TensorConstant{(1,) of 0.5} [id E]
| | | | |TensorConstant{[1. 2.]} [id K]
| | | | |Elemwise{sub,no_inplace} [id G]
| | | |TensorConstant{[0. 3.]} [id BV]
| | |InplaceDimShuffle{x} [id BW]
| | |TensorConstant{2} [id BX]
| |Sum{acc_dtype=float64} [id BY]
| |Elemwise{log,no_inplace} [id BZ]
| |ExtractDiag{offset=0, axis1=0, axis2=1, view=False} [id BJ]
|All [id BH]The graph for position isn't present in potential_energy (e.g. id A isn't a node in the graph for potential_energy starting at id M), which explains the error. It looks like the subgraph starting at the node with id BS in the potential_energy graph is a copy/clone of position. We need that subgraph to be identical to position in order for position to be considered an "input" of potential_energy.
It looks like AePPL is doing some unnecessary cloning somewhere.
This was fixed in aesara-devs/aeppl#248 and will be available soon in release v0.1.4.