Feedforward block cannot detect parameters outside initial layer.

[anthonyrtw]

Describe the bug
Feedforward operations are useful in QML since it allows us to encode certain parameters conditional on certain measurements. Currently, FeedForwardBlock rejects any perceval Experiment with parameters encoded outside the initial layer.

To Reproduce

from merlin.algorithms import FeedForwardBlock
from perceval import *

experiment = Experiment(3)
experiment.add(0, Unitary.random(3))
experiment.with_input([1, 1, 0])
experiment.add(0, Detector.pnr())

x_param = P("x")

feedforward = FFCircuitProvider(1, 0, Circuit(2))
feedforward.add_configuration([0], BS(x_param) // BS(P("A")))
feedforward.add_configuration([1], BS(x_param) // BS(P("B")))
experiment.add(0, feedforward)

feedforward = FeedForwardBlock(
    experiment,
    trainable_parameters=["A", "B"],
    input_parameters=["x"]
)

Observed results
Immediate ValueError.

ValueError: The first stage must use all of the input parameters. Create you own stages with variable input parameters with the partial measurement strategy instead

Expected behavior
FeedForward block should be able to inspect all of the configurations in the feedforward configurator and detect all parameters.

[CassNot]

test_feed_forward_partial_measurement_comparison.py

Hi @anthonyrtw !
Thank you for the report. This is currently a known limitation of FeedForwardBlock, and the new ValueError is intentional to give some hint before a stronger fix. This will be fixed with a different backend to FeedForwardBlock in Merlin 0.5.
This construction (with input after measurement) is rejected for now:

feedforward.add_configuration([0], BS(P("x")) // BS(P("A")))
feedforward.add_configuration([1], BS(P("x")) // BS(P("B")))

One proposition to overcome this limitation could be to build the stages manually with multiple QuantumLayers with partial measurement strategy (you decide which modes you measure:

1. first QuantumLayer with MeasurementStrategy.partial(...);
2. inspect the returned PartialMeasurement;
3. route each branch StateVector into a branch-specific QuantumLayer;
4. pass x to that branch layer.

I am attaching to my answer a python file to support my explanation. When I run it, I get:

Equivalent case: FeedForwardBlock and manual partial measurement agree.
FeedForwardBlock probabilities:
  (0, 0, 2): 0.13250732
  (0, 1, 1): 0.09640302
  (0, 2, 0): 0.04533275
  (1, 0, 1): 0.27336785
  (1, 1, 0): 0.25005433
  (2, 0, 0): 0.20233472
Manual partial-measurement probabilities:
  (0, 0, 2): 0.13250734
  (0, 1, 1): 0.09640302
  (0, 2, 0): 0.04533275
  (1, 0, 1): 0.27336785
  (1, 1, 0): 0.25005430
  (2, 0, 0): 0.20233475

Known limitation: FeedForwardBlock rejects branch-local input x.
  The first stage must use all of the input parameters. Create you own stages with variable input parameters with the partial measurement strategy instead

Manual partial-measurement layers accept branch-local input x.
  (0, 0, 2): 0.00993002
  (0, 1, 1): 0.67176980
  (0, 2, 0): 0.04788725
  (1, 0, 1): 0.21582304
  (1, 1, 0): 0.00341684
  (2, 0, 0): 0.05117322
  total probability: 1.00000024

If that circuit works for you, we could add it to our 0.4 documentation !