Add BiLipREN implementations in Jax and Pytorch#48
Open
yuruizhang06 wants to merge 4 commits into
Open
Conversation
RoverLiu
reviewed
Jul 1, 2026
RoverLiu
reviewed
Jul 1, 2026
RoverLiu
left a comment
RoverLiu
left a comment
Collaborator
There was a problem hiding this comment.
I did not explicitly look into ren related files (ren_base_jax, ren_composition_jax, ren_composition_torch, ren_jax, ren_torch). @nic-barbara , could you please have a look at it?
| return w_eq_final | ||
|
|
||
|
|
||
| def peaceman_rachford_layer(activation, D11, b, tol=1e-9, alpha=0.8, |
| return w_eq_final | ||
|
|
||
|
|
||
| def douglas_rachford_layer(activation, D11, b, tol=1e-9, alpha=0.6, |
| return total_elements | ||
|
|
||
|
|
||
| def compute_lipschitz_constants( |
Collaborator
There was a problem hiding this comment.
I am not sure if we actually need this function in robustnn package. This feels like an evaluation approach and is not related to model itself
Collaborator
Author
There was a problem hiding this comment.
When using BiLipREN, we might want to estimate the empirical Lip bounds from data as a baseline for choosing Lip bounds for BiLipREN
There was a problem hiding this comment.
Pull request overview
This PR adds bi-Lipschitz Recurrent Equilibrium Network (BiLipREN) implementations and supporting layers in both JAX and PyTorch, along with shared solver utilities and tests to validate forward/inverse behavior, bounds, and gradients.
Changes:
- Added JAX and PyTorch BiLipschitz REN implementations (including explicit inverse support) plus composition networks with (dynamic) unitary layers.
- Consolidated fixed-point / splitting solvers into a shared
robustnn/solvers.pymodule and updated MonLipNet imports accordingly. - Added new JAX and PyTorch test suites covering forward evaluation, inverse round-trips, empirical bi-Lipschitz bounds, and gradient flow.
Reviewed changes
Copilot reviewed 16 out of 16 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| test/test_bilipren.py | New JAX tests for BiLipschitzREN and CompositionREN, including inverse and bound checks. |
| test/test_bilipren_torch.py | New PyTorch tests for BiLipschitzREN and CompositionREN, including inverse and bound checks. |
| robustnn/utils.py | Removes trailing stray whitespace lines. |
| robustnn/solvers.py | New shared solver module (Davis–Yin + equilibrium-layer splitting solvers). |
| robustnn/solver_DYS.py | Removed; Davis–Yin solver moved into robustnn/solvers.py. |
| robustnn/ren_torch.py | New PyTorch REN + bi-Lipschitz parameterization and inverse evaluation utilities. |
| robustnn/ren_jax.py | Adds _get_qsr override hook and introduces BiLipschitzREN as a GeneralREN specialization. |
| robustnn/ren_composition_torch.py | New PyTorch composition of BiLipschitz REN blocks with (dynamic) unitary layers and inverses. |
| robustnn/ren_composition_jax.py | New JAX composition of BiLipschitz REN blocks with (dynamic) unitary layers and inverses. |
| robustnn/ren_base_jax.py | Adds full (non-triangular) equilibrium solver support with an IFT-based custom VJP + explicit inverse plumbing. |
| robustnn/orthogonal_torch.py | Trailing newline addition. |
| robustnn/orthogonal_jax.py | Trailing newline addition. |
| robustnn/monlipnet_torch.py | Updates import to use robustnn.solvers.DavisYinSplit. |
| robustnn/monlipnet_jax.py | Updates import to use robustnn.solvers.DavisYinSplit. |
| robustnn/dyn_orthogonal_torch.py | New PyTorch dynamic unitary (stateful orthogonal) layer. |
| robustnn/dyn_orthogonal_jax.py | New JAX dynamic unitary (stateful orthogonal) layer. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+82
to
+86
| for i in range(w_eq_bar.shape[0]): | ||
| ji = j_diag[i, ...] | ||
| y_bar_i = y_bar[i, ...] | ||
| w_grad = jnp.linalg.solve(I - (ji * D11.T), y_bar_i.T).T | ||
| w_eq_bar = w_eq_bar.at[i, ...].set(w_grad) |
Comment on lines
+67
to
+69
| def _equilibrium_ift_grad_fwd(activation, D11, v, w_eq): | ||
| I = jnp.identity(v.shape[-1]) | ||
| return w_eq, (D11, v, I) |
Comment on lines
+296
to
+304
| def _get_qsr(self): | ||
| n = self.input_size | ||
| I = torch.eye(n, dtype=self._dtype) | ||
| alpha1 = 2.0 * (self.mu * self.nu) / (self.mu + self.nu) | ||
| alpha2 = 2.0 / (self.mu + self.nu) | ||
| Q = -alpha2 * I | ||
| S = I.clone() | ||
| R = -alpha1 * I | ||
| return Q, S, R |
Comment on lines
+307
to
+311
| nu, nx, ny = self.input_size, self.state_size, self.output_size | ||
| Q, S, R = self._get_qsr() | ||
| I_ny = torch.eye(ny, dtype=Q.dtype) | ||
| I_nu = torch.eye(nu, dtype=Q.dtype) | ||
|
|
Comment on lines
+329
to
+333
| R1 = R + S @ D22 + D22.T @ S.T + D22.T @ Q @ D22 | ||
| mul_Q = torch.cat((self.C2, self.D21, torch.zeros((ny, nx), dtype=Q.dtype)), dim=1) | ||
| mul_R = torch.cat((C2_imp, D21_imp, self.B2.T), dim=1) | ||
| Gamma_Q = mul_Q.T @ Q @ mul_Q | ||
| Gamma_R = mul_R.T @ torch.linalg.solve(R1, mul_R) |
This file contains hidden or 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
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.
No description provided.