Release v0.7.0 #336
Merged
Release v0.7.0 #336
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
**Context:** Now the default branch for Mr Mustard repository is develop. **Description of the Change:** This PR updates the CI checks so that - Every PR pointing to develop is checked using python 3.8 - Every PR pointing to develop is built and checked using python 3.8, 3.9, 3.10 so that code is surely releasable - adds codeowners file - adds `.coveragerc` file to avoid some non-sense codecov issues **Benefits:** CI checks now will be in sync with MrMustard development policies, this also reduces usage of CI checks (meaning less bills 💸 and less CO2 ☘️). Also, codecov should be less annoying from now. **Possible Drawbacks:** We'll have to wait for a full release cycle to check that all workflows are working as expected
**Context:**
Currently homodyne and heterodyne measurements on Mr Mustard require
user-specified outcome values.
**Description of the Change:**
This PR implements sampling for homodyne and heterodyne measurements:
when no measurement outcome value is specified (`result=None` for
homodyne and `x, y = None, None` for heterodyne), the outcome is sampled
from the measurement probability distribution and the conditional state
(conditional on the outcome) on the remaining modes is generated.
```python
import numpy as np
from mrmustard.lab import Homodyne, Heterodyne, TMSV, SqueezedVacuum
# conditional state from measurement
conditional_state1 = TMSV(r=0.5, phi=np.pi)[0, 1] >> Homodyne(quadrature_angle=np.pi/2, result=None)[1]
conditional_state2 = TMSV(r=0.5, phi=np.pi)[0, 1] >> Heterodyne(x=None, y=None)[1]
# outcome probability
outcome_prob1 = SqueezedVacuum(r=0.5) >> Homodyne(result=None)
outcome_prob2 = SqueezedVacuum(r=0.5) >> Heterodyne(x=None, y=None)
```
To do so:
- `tensorflow-probability==0.17.0` is added to Mr Mustard's
dependencies. [TensorFlow
Probability](https://www.tensorflow.org/probability/overview) provides
integration of probabilistic methods with automatic differentiation and
hardware acceleration (GPUs). Pytorch also provides similar
functionality through the
[`torch.distributions`](https://pytorch.org/docs/stable/distributions.html)
module.
- For gaussian states 𝛠, samples are drawn from the gaussian PDF
generated by a highly squeezed state ξ and the state of interest, i.e.,
`PDF=Tr[𝛠 ξ]`. Sampling from the distribution is implemented using
TensorFlow's multivariate normal distribution.
- For Homodyne only: To calculate the pdf for Fock states the
q-quadrature eigenfunctions `|x><x|` are used: `PDF=Tr[𝛠 |x><x|]`.
Sampling from the distribution uses TensorFlow's categorical
distribution. This case is inspired by the [implementation on
strawberryfields](https://github.com/XanaduAI/strawberryfields/blob/9a9a352b5b8cf7b2915e45d1538b51d7d306cfc8/strawberryfields/backends/tfbackend/circuit.py#L818-L926).
**Benefits:**
Sampling from homodyne measurements! 🎲
**Possible Drawbacks:**
Sampling for Fock states can be improved to reduce overheads on
execution: currently the pdf is recalculated every time a measurement is
performed leading to unnecessary slow-downs. To improve this, one can
use sampling algorithms that do not require the complete calculation of
the pdf, for example: using the cumulative distribution (however this is
highly dependent on normalization), rejection sampling,
Metropolis-Hasting, etc.
Co-authored-by: JacobHast <[email protected]>
Co-authored-by: elib20 <[email protected]>
**Context:** The rotation gate unitary is currently being calculated using the Choi isomorphisms which generates `nan` for angles of value zero where the identity matrix should be returned. **Description of the Change:** The rotation gate being diagonal in the Fock basis is easy to implement with better speed performance. This PR implements the `U` method of the rotation gate overriding the default way of calculating the unitaries of Gaussian transforms. **Benefits:** Fast computation of the rotation gate Fock representation which avoids invalid numerical outcomes. **Possible Drawbacks:** None
**Context:** [The Walrus PR #351](XanaduAI/thewalrus#351) implemented a more stable calculation of the displacement unitary. **Description of the Change:** The `Dgate` now uses The Walrus to calculate the unitary and gradients of the displacement gate. Before:  After:  **Benefits:** This provides better numerical stability for larger cutoff and displacement values. **Possible Drawbacks:** None **Related GitHub Issues:** None
------------------------------------------------------------------------------------------------------------ **Context:** Math interface **Description of the Change:** Adds `eye_like` function, which returns the identity matrix to match the shape and dtype of the argument. **Benefits:** Simpler to get the identity of the right size and dtype **Possible Drawbacks:** Not standard **Related GitHub Issues:** None
**Context:** Calculation of the Wigner function is already included in Mr Mustard but hidden in the graphics module. **Description of the Change:** This PR moves the Wigner function calculation to its own module and numbifies it. **Benefits:** Now you can calculate the Wigner function using ```python from mrmustard.utils.wigner import wigner_discretized wigner_discretized(dm, q, p) # dm is a density matrix ``` It should be faster as well because it uses numba jit. **Possible Drawbacks:** None
**Context:** There was a bug in the gradient computation of the Gate **Description of the Change:** Bug is fixed (there was a numpy array at some point that was breaking the chain rule) **Benefits:** It works **Possible Drawbacks:** None **Related GitHub Issues:** None
**Context:** settings are not easy do discover **Description of the Change:** the settings object has now a nice repr **Benefits:** users can see al the settings at once **Possible Drawbacks:** None I can think of **Related GitHub Issues:** None <img width="513" alt="Screen Shot 2022-11-08 at 9 55 08 AM" src="https://user-images.githubusercontent.com/8944955/200597522-66ebfa0f-dd87-4109-982f-c8d3231823f8.png"> Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** PR #143 introduced sampling for homodyne measurements and with it a bunch of functionality in the `utils/homodyne.py` module. Most of this functions are related to the sampling in fock representation and are not only useful for homodyne sampling — take for example the calculation of the quadrature distribution. **Description of the Change:** This PR refactors the homodyne module into: - `physics.fock` — functions related to the fock representation were moved into this module. Some of them are even split and written as separated functions such that they are available for use in other contexts (for example `quadrature_distribution` and `oscillator_eigenstate`). - `math.caching` — the cache decorator used for the Hermite polys is refactored into this new module, this decorator is not only applicable to the Hermite polys but also to any function taking a 1D tensor + int parameters. Now it can be used generically by any function with this same signature. The idea of this module is to contain caching functions of this kind. Also - Hermite polynomials have the modified flag removed (as per [this](#143 (review)) comment) and now only the regular polynomials are used. _Note:_ This PR only refactors code, meaning where the code is placed and _not_ what it does and _nor_ how it is done; there is no change to the logic whatsoever. **Benefits:** Pieces of the code that were there already are now reusable. The code for the sampling logic is more readable now. **Possible Drawbacks:** None **Related GitHub Issues:** None
**Context:** Currently marginals are calculated form the Wigner function. In cases in which not all features of the state are captured on the Wigner function, the marginals contain negativities hence not representing a true probability density. **Description of the Change:** This PR makes `mikkel_plot` - calculate marginals independently from the Wigner function thus ensuring that the marginals are physical even though the Wigner function might not contain all the features of the state within the defined window, - expose the `ticks`, `tick_labels` and `grid` arguments to customize the visualization — useful for example when checking that your state has peaks in those dreaded multiples of √(π) - return the figure and axes for further processing if needed **Benefits:** - Marginals in the visualization will always represent physical states _before_: note how this lion is not fully displayed in the Wigner function leading to negativities in the marginal distributions <img width="399" alt="image" src="https://user-images.githubusercontent.com/675763/202817119-1824ecc0-139b-42ef-8461-8eba3ebf5405.png"> _after_: although the big kittie is not displayed in its full glory its marginals show the true probability distribution <img width="390" alt="image" src="https://user-images.githubusercontent.com/675763/202817226-b06560e6-d481-4530-af0f-010cfafb6885.png"> - More configurability of the visualization ```python from matplotlib import cm ticks = [-2*np.sqrt(2),2*np.sqrt(2)] labels = [r"$-2\sqrt{\hbar}$", r"$2\sqrt{\hbar}$"] graphics.mikkel_plot(dm, xticks=ticks, yticks=ticks, xtick_labels=labels, ytick_labels=[r"$-2\sqrt{\hbar}$", r"$2\sqrt{\hbar}$"], grid=True, cmap=cm.PuOr) ```  - Users can take the figure and axes for post-processing or storing **Possible Drawbacks:** This will _marginally_ increase the computation time for the visualization but shouldn't be too impactful. **Related GitHub Issues:** None Co-authored-by: ziofil <[email protected]> Co-authored-by: Luke Helt <[email protected]>
**Context:** Optimizer is opaque **Description of the Change:** in the minimize function now we can pass a callback that will be executed at the end of each step (with `trainable_parameters` as argument) and the return is stored in `self.callback_history`. **Benefits:** Can do lots of things, e.g.  **Possible Drawbacks:** None **Related GitHub Issues:** None
**Context:** The `Rgate` and the `Dgate` don't work correctly in Fock representation in some cases. **Description of the Change:** 1. Fixed the `Rgate` and `Dgate` by removing the parser method and simplifying the code. 2. Added two functions in the fock module for explicitly applying an operator to a ket or to a dm (sandwich) which avoid constructing unitaries as large as the whole circuit. Now they are used in the `transform_fock` method of the `Transformation` class. 3. fixed a bug in the trace function 4. minor improvements **Benefits:** **Possible Drawbacks:** **Related GitHub Issues:** Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** sometimes it's useful to compute the fock representation using different cutoffs for input-output indices of the same mode **Description of the Change:** transformations (gates, circuit) now accept also a list of double (or quadruple for choi) length which specifies the cutoffs per index (rather than per mode) **Benefits:** saves runtime if one needs e.g. to input fock states into a gaussian circuit **Possible Drawbacks:** none **Related GitHub Issues:**
**Context:** norm is confusing field in the repr of `State` (it's the sqrt of the probability if the state is pure and the probability if it's mixed). **Description of the Change:** Replace norm with probability **Benefits:** Consistent meaning and also more practically useful **Possible Drawbacks:** some users may miss the good old norm? **Related GitHub Issues:**
**Context:** The application of a choi operator to a density matrix was resulting in a transposed dm **Description of the Change:** Fixes the order of the indices in the application of a choi operator to dm and ket **Benefits:** Correct result **Possible Drawbacks:** None **Related GitHub Issues:** None Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** Always forgetting to add entries to the CHANGELOG file? **Description of the Change:** Now github will remind you to do so — this PR implements a new CI workflow checking if an entry has been added to the CHANGELOG file. This check is not mandatory meaning it won't block the ability to merge the PR, however checks will appear as failed. In case no changelog entry is needed one can use the `no changelog` label to disable the check. This PR also adds the changelog entry for PR #188. **Benefits:** No more PRs without CHANGELOG entries **Possible Drawbacks:** None
**Context:** Setting a seed in MM is not trivial and often we need reproducible reults **Description of the Change:** The `settings` object now supports the `SEED` attribute, which is random unless it is set by the user. To unset it, just set it equal to `None`. **Benefits:** Easy to get reproducible results without messing with numpy.random. **Possible Drawbacks:** None? **Related GitHub Issues:** Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** The fock representation of a state is stored internally for future access. However, if in the meantime the cutoff is updated this is not taken into account, this PR solves this issue. **Description of the Change:** The cutoffs are applied to the internal fock representation upon access. **Benefits:** More correct implementation. **Possible Drawbacks:** None? **Related GitHub Issues:** None Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** Going forward we need a big refactoring of the methods to transform between representations. This PR is the first in this direction. **Description of the Change:** Introduced two new modules and refactored various methods **Benefits:** Allows for easier extension of representation methods **Possible Drawbacks:** Some methods and functions have a new name and/or arguments **Related GitHub Issues:** None Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** Fixes a few bugs related to transformation of states in Fock representation **Description of the Change:** - adds missing default argument - adds a function for convenience - fixes a bug due to old code not removed in `Transformation.transform_fock` - adds 11 tests to avoid same type of bug - adds 2 parametrized tests (10 in total) to check correct application of unitaries and channels to kets and dm in Fock representation **Benefits:** Correct code **Possible Drawbacks:** None **Related GitHub Issues:** No issue but FTB simulation not running on develop branch
**Context:** Fixing a bug related to tensorflow not liking products of complex128 and float64 **Description of the Change:** cast where necessary **Benefits:** Code that runs **Possible Drawbacks:** None **Related GitHub Issues:** None
**Context:** new black version, wants to make new changes **Description of the Change:** black all the files that need to be blacked **Benefits:** to bring the develop branch up to date so that all the PRs that are open don't have to **Possible Drawbacks:** None **Related GitHub Issues:** None
**Context:** Adds `map_trainer` as the interface for distributing optimization workflows using ray so that things can happen in parallel -- replacing the need for `for` loops. _Code here has been personally looked at by Trudeau._ **Description of the Change:** Demo notebook in recipes. Documentation page for the `trainer` module is added with some examples, which is also in the docstring of `map_trainer` so that it can be conveniently read in Jupyter directly with shift+tab. I've also had some weird problem with the unit tests (all passing locally) running on github action where some times it would just hang. I've tried removing my new tests one by one and adding them back again one by one, and it somehow worked at the end... I changed the testing workflow file with direct `pip install .[ray]` instead of building wheel first, which I don't think is the reason. Thanks Trudeau I guess? **Benefits:** fast and simple experimentation -> more research done. **Possible Drawbacks:** more interface to introduce **Related GitHub Issues:**
**Context:** Circuit building is a bit obscure because there's no visual circuit representation **Description of the Change:** Adds a basic circuit drawer (adapted from pennylane). **Benefits:** Circuit visualizations **Possible Drawbacks:** Does not yet include initializations and measurements, but for this we need to refactor the `Circuit` class first. **Related GitHub Issues:** None --------- Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** Attempt to fix the (random) never ending tests on github actions. I have so far rerun the test 5 times without the issue reoccurring. 🤞 **Description of the Change:** Forces ray to init with 1 cpu for testing. **Benefits:** **Possible Drawbacks:** **Related GitHub Issues:** --------- Co-authored-by: Sebastián Duque Mesa <[email protected]>
**Context:** Tests could improve in MM **Description of the Change:** Add several new tests and improves hypothesis strategies **Benefits:** Better test suite **Possible Drawbacks:** More tests to maintain? Nah **Related GitHub Issues:** None
**Context:** Make PNR sampling faster for Gaussian circuits when using density matrices. This is done by applying the recurrence relation in a selective manner such that useless (off-diagonal) amplitudes from the Fock representation are not calculated. When all modes are detected, `math.hermite_renormalized` can be replaced by `math.hermite_renormalized_diagonal`. In case all but the first mode are detected, `math.hermite_renormalized_1leftoverMode` can be used. The complexity of these new methods is equal to performing a pure state simulation. The methods are differentiable, such that they can be used for defining costfunctions. **Description of the Change:** Adds the function `math.hermite_renormalized_diagonal` and `math.hermite_renormalized_1leftoverMode`. **Benefits:** Faster simulation and optimization of Gaussian circuits with PNR detectors when using density matrices. --------- Co-authored-by: Robbe De Prins (UGent-imec) <[email protected]> Co-authored-by: ziofil <[email protected]> Co-authored-by: ziofil <[email protected]>
**Context:** Fixing small bugs/typos before 0.4 release **Description of the Change:** - Threshold detector can now be correctly initialized - changes in settings.HBAR are correctly reflected everywhere - Interferometer can be placed on any set of modes - number of decimals are respected in circuit drawer **Benefits:** Fewer bugs/typos **Possible Drawbacks:** None **Related GitHub Issues:** None
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## main #336 +/- ##
===========================================
+ Coverage 71.62% 84.14% +12.51%
===========================================
Files 29 65 +36
Lines 2848 4775 +1927
===========================================
+ Hits 2040 4018 +1978
+ Misses 808 757 -51
Continue to review full report in Codecov by Sentry.
|
SamFerracin
force-pushed
the
release-0.7
branch
2 times, most recently
from
9cd7586 to
ab42599
Compare
SamFerracin
force-pushed
the
release-0.7
branch
from
7c50f44 to
fd4323f
Compare
**Context:** In the latest MrM release, `scipy` was unpinned. Syncing the `develop` branch with this latest release causes pytest errors due to improper type handling on `scipy`'s side. **Description of the Change:** - Syncing `develop` with latest `main` - Fixing scipy error by adding a `dtype` parameter to `math.sqrtm` so that it never returns arrays of an unsupported type (specifically, `complex256`) - Updating dependencies - Modifying the CHANGELOG in preparation for release --------- Co-authored-by: zy <[email protected]>
**Context:** The previous solution based on `logging` does not appear to work on every OS
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
New features
Added a new interface for backends, as well as a
numpybackend (which is now default). Users can runall the functions in the
utils,math,physics, andlabwith both backends, whiletrainingrequires using
tensorflow. Thenumpybackend provides significant improvements both in importtime and runtime. (#301)
Added the classes and methods to create, contract, and draw tensor networks with
mrmustard.math.(#284)
Added functions in physics.bargmann to join and contract (A,b,c) triples.
(#295)
Added an Ansatz abstract class and PolyExpAnsatz concrete implementation. This is used in the Bargmann representation.
(#295)
Added
complex_gaussian_integralandreal_gaussian_integralmethods.(#295)
Added
Bargmannrepresentation (parametrized by Abc). Supports all algebraic operations and CV (exact) inner product.(#296)
Breaking changes
Removed circular dependencies by:
graphics.py--movedProgressBartotrainingandmikkel_plottolab.circuit_drawerandwignertophysics.xptensortomath.(#289)
Created
settings.pyfile to hostSettings.(#289)
Moved
settings.py,logger.py, andtyping.pytoutils.(#289)
Removed the
Mathclass. To use the mathematical backend, replacefrom mrmustard.math import Math ; math = Math()withimport mrmustard.math as mathin your scripts.
(#301)
The
numpybackend is now default. To switch to thetensorflowbackend, add the line
math.change_backend("tensorflow")to your scripts.(#301)
Improvements
Calculating Fock representations and their gradients is now more numerically stable (i.e. numerical blowups that
result from repeatedly applying the recurrence relation are postponed to higher cutoff values).
This holds for both the "vanilla strategy" (#274) and for the
"diagonal strategy" and "single leftover mode strategy" (#288).
This is done by representing Fock amplitudes with a higher precision than complex128 (countering floating-point errors).
We run Julia code via PyJulia (where Numba was used before) to keep the code fast.
The precision is controlled by
setting settings.PRECISION_BITS_HERMITE_POLY. The default value is128,which uses the old Numba code. When setting to a higher value, the new Julia code is run.
Replaced parameters in
trainingwithConstantandVariableclasses.(#298)
Improved how states, transformations, and detectors deal with parameters by replacing the
Parametrizedclass withParameterSet.(#298)
Includes julia dependencies into the python packaging for downstream installation reproducibility.
Removes dependency on tomli to load pyproject.toml for version info, uses importlib.metadata instead.
(#303)
(#304)
Improves the algorithms implemented in
vanillaandvanilla_vjpto achieve a speedup.Specifically, the improved algorithms work on flattened arrays (which are reshaped before being returned) as opposed to multi-dimensional array.
(#312)
(#318)
Adds functions
hermite_renormalized_batchandhermite_renormalized_diagonal_batchto speed up calculatingHermite polynomials over a batch of B vectors.
(#308)
Added suite to filter undesired warnings, and used it to filter tensorflow's
ComplexWarnings.(#332)
Bug fixes
shapeinput parameters to all methodsUin thegates.pyfile.(#291)
atolin purity evaluation for Gaussian states.(#294)
(#305)
np.emptywithnp.zerosto fix instabilities.(#309)