Relative trajectory balance by josephdviviano · Pull Request #457 · GFNOrg/torchgfn

josephdviviano · 2025-12-16T07:40:55Z

I've read the .github/CONTRIBUTING.md file
My code follows the typing guidelines
I've added appropriate tests
I've run pre-commit hooks locally

Description

Adds method for pre-training a diffusion-based pf using either a learned backward or brownian bridge from provided samples using maximum likelihood, DiffusionMLE.
Adds a method for fine-tuning from a prior model to a posterior distribution using the Relative Trajectory Balance loss, RelativeTrajectoryBalanceGFlowNet.
Adds a script showing how to use both, train_diffusion_rtb.py.
Adds some simple tests.
Minor changes to existing diffusion modules to allow for additional features such as learned variance.

codecov · 2025-12-16T08:07:34Z

Codecov Report

❌ Patch coverage is 75.68389% with 80 lines in your changes missing coverage. Please review.
✅ Project coverage is 74.33%. Comparing base (91aae26) to head (97a2453).
⚠️ Report is 25 commits behind head on master.

Files with missing lines	Patch %	Lines
src/gfn/gym/diffusion_sampling.py	64.70%	27 Missing and 3 partials ⚠️
src/gfn/gflownet/mle.py	81.90%	12 Missing and 7 partials ⚠️
src/gfn/utils/modules.py	71.15%	10 Missing and 5 partials ⚠️
src/gfn/gflownet/trajectory_balance.py	70.27%	7 Missing and 4 partials ⚠️
src/gfn/estimators.py	95.12%	1 Missing and 1 partial ⚠️
src/gfn/gym/helpers/diffusion_utils.py	0.00%	2 Missing ⚠️
src/gfn/gflownet/base.py	80.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #457      +/-   ##
==========================================
+ Coverage   74.23%   74.33%   +0.09%     
==========================================
  Files          47       48       +1     
  Lines        6805     7112     +307     
  Branches      800      830      +30     
==========================================
+ Hits         5052     5287     +235     
- Misses       1449     1503      +54     
- Partials      304      322      +18

Files with missing lines	Coverage Δ
src/gfn/env.py	`81.29% <100.00%> (ø)`
src/gfn/gflownet/base.py	`78.88% <80.00%> (-0.19%)`	⬇️
src/gfn/estimators.py	`91.77% <95.12%> (-0.06%)`	⬇️
src/gfn/gym/helpers/diffusion_utils.py	`22.22% <0.00%> (-1.78%)`	⬇️
src/gfn/gflownet/trajectory_balance.py	`79.48% <70.27%> (-8.32%)`	⬇️
src/gfn/utils/modules.py	`72.80% <71.15%> (+0.05%)`	⬆️
src/gfn/gflownet/mle.py	`81.90% <81.90%> (ø)`
src/gfn/gym/diffusion_sampling.py	`48.75% <64.70%> (+4.89%)`	⬆️

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

josephdviviano · 2025-12-17T05:58:38Z

src/gfn/estimators.py

-            s_curr,
-            s_curr * self.dt / t_curr,
+            torch.zeros_like(s_curr),
+            s_curr * (1.0 - self.dt / t_curr),


@hyeok9855 this might have been the cause of the problem you had before in your code (learning slower), worth checking.

@hyeok9855 actually- this was my bug!!

josephdviviano · 2025-12-17T05:59:52Z

src/gfn/gflownet/base.py

+            recalculate_all_logprobs=recalculate_all_logprobs,
+        )
+
+    def trajectory_log_probs_backward(


leaving these here because they might come in handy, but I don't think they're actually needed right now in this implementation.

…gfn into relative_trajectory_balance

josephdviviano · 2025-12-18T02:38:57Z

@copilot how can I prevent codecov from failing CI? I want it to track coverage, but not block merge requests.

Copilot · 2025-12-18T02:39:04Z

@josephdviviano I've opened a new pull request, #458, to work on those changes. Once the pull request is ready, I'll request review from you.

Co-authored-by: josephdviviano <4142570+josephdviviano@users.noreply.github.com>

Copilot

Pull request overview

This PR adds Relative Trajectory Balance (RTB) support for fine-tuning diffusion-based GFlowNets from a prior to a posterior distribution. It introduces a two-stage training pipeline: (1) maximum likelihood estimation (MLE) pre-training of a prior diffusion model using either a fixed Brownian bridge or learned backward policy, and (2) RTB-based fine-tuning that adapts the prior to match a posterior distribution weighted by a reward function.

Key changes include:

Implementation of RelativeTrajectoryBalanceGFlowNet for fine-tuning with the RTB loss
Implementation of MLEDiffusion trainer for prior pre-training via maximum likelihood
Support for learned variance in both forward and backward diffusion policies
New Gaussian mixture target distributions for RTB demonstrations (25-mode prior, 9-mode posterior)
Enhanced diffusion trajectory mask alignment with consistent terminal time detection

Reviewed changes

Copilot reviewed 15 out of 16 changed files in this pull request and generated 19 comments.

Show a summary per file

File	Description
`src/gfn/gflownet/trajectory_balance.py`	Adds `RelativeTrajectoryBalanceGFlowNet` class implementing the RTB objective
`src/gfn/gflownet/mle.py`	New MLE trainer for diffusion pre-training with backward sampling
`src/gfn/estimators.py`	Adds learned variance support and exploration noise handling to forward/backward estimators
`src/gfn/utils/modules.py`	Implements `DiffusionPISGradNetBackward` and adds learned variance to forward module
`src/gfn/gym/diffusion_sampling.py`	Adds `Grid25GaussianMixture` and `Posterior9of25GaussianMixture` targets; improves terminal state detection
`src/gfn/gym/helpers/diffusion_utils.py`	Adds `max_n_samples` parameter to visualization utility
`src/gfn/gflownet/base.py`	Adds helper methods for computing forward/backward trajectory log-probs separately
`src/gfn/gflownet/__init__.py`	Exports `RelativeTrajectoryBalanceGFlowNet`
`src/gfn/env.py`	Minor comment clarification on boolean masking behavior
`tutorials/examples/train_diffusion_rtb.py`	Complete end-to-end tutorial demonstrating MLE pre-training and RTB fine-tuning
`tutorials/examples/output/.gitignore`	Ignores generated checkpoint and visualization files
`testing/test_rtb.py`	Unit tests for RTB loss computation and gradient flow
`testing/test_environments.py`	Test verifying diffusion trajectory mask alignment
`testing/gym/test_diffusion_sampling_rtb.py`	Tests for new GMM target distributions
`testing/gflownet/test_mle_diffusion.py`	Comprehensive tests for MLE diffusion trainer
`.flake8`	Minor formatting (trailing newline)

Comments suppressed due to low confidence (2)

src/gfn/gym/diffusion_sampling.py:69

This call to BaseTarget.sample in an initialization method is overridden by SimpleGaussianMixture.sample.
This call to BaseTarget.sample in an initialization method is overridden by Grid25GaussianMixture.sample.
This call to BaseTarget.sample in an initialization method is overridden by Posterior9of25GaussianMixture.sample.
This call to BaseTarget.sample in an initialization method is overridden by Funnel.sample.
This call to BaseTarget.sample in an initialization method is overridden by ManyWell.sample.

            self.gt_xs = self.sample(n_gt_xs, seed)

src/gfn/gym/diffusion_sampling.py:70

This call to BaseTarget.log_reward in an initialization method is overridden by SimpleGaussianMixture.log_reward.
This call to BaseTarget.log_reward in an initialization method is overridden by Grid25GaussianMixture.log_reward.
This call to BaseTarget.log_reward in an initialization method is overridden by Posterior9of25GaussianMixture.log_reward.
This call to BaseTarget.log_reward in an initialization method is overridden by Funnel.log_reward.
This call to BaseTarget.log_reward in an initialization method is overridden by ManyWell.log_reward.

            self.gt_xs_log_rewards = self.log_reward(self.gt_xs)

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2025-12-18T02:51:06Z

src/gfn/gym/diffusion_sampling.py

+            os.makedirs("output", exist_ok=True)
+            plt.savefig(f"output/{prefix}simple_gmm.png")


The output directory is changed from "viz" to "output" across multiple target classes. While this change is consistent, it would be better to define the output directory as a constant or configuration parameter at the module level rather than hardcoding it in multiple places.

@copilot open a new pull request to apply changes based on this feedback

Copilot · 2025-12-18T02:51:06Z

src/gfn/utils/modules.py

+class DiffusionPISGradNetBackward(nn.Module):
+    """Learnable backward correction module (PIS-style) for diffusion.
+
+    Produces mean and optional log-std corrections that are tanh-scaled by
+    `pb_scale_range` to stay close to the analytic Brownian bridge.
+    """
+
+    def __init__(
+        self,
+        s_dim: int,
+        harmonics_dim: int = 64,
+        t_emb_dim: int = 64,
+        s_emb_dim: int = 64,
+        hidden_dim: int = 64,
+        joint_layers: int = 2,
+        zero_init: bool = False,
+        clipping: bool = False,
+        gfn_clip: float = 1e4,
+        pb_scale_range: float = 0.1,
+        log_var_range: float = 4.0,
+        learn_variance: bool = True,
+    ) -> None:
+        super().__init__()
+        self.s_dim = s_dim
+        self.out_dim = s_dim + (1 if learn_variance else 0)
+        self.harmonics_dim = harmonics_dim
+        self.t_emb_dim = t_emb_dim
+        self.s_emb_dim = s_emb_dim
+        self.hidden_dim = hidden_dim
+        self.joint_layers = joint_layers
+        self.zero_init = zero_init
+        self.clipping = clipping
+        self.gfn_clip = gfn_clip
+        self.pb_scale_range = pb_scale_range
+        self.log_var_range = log_var_range
+        self.learn_variance = learn_variance
+
+        assert (
+            self.s_emb_dim == self.t_emb_dim
+        ), "Dimensionality of state embedding and time embedding should be the same!"
+
+        self.t_model = DiffusionPISTimeEncoding(
+            self.harmonics_dim, self.t_emb_dim, self.hidden_dim
+        )
+        self.s_model = DiffusionPISStateEncoding(self.s_dim, self.s_emb_dim)
+        self.joint_model = DiffusionPISJointPolicy(
+            self.s_emb_dim,
+            self.hidden_dim,
+            self.out_dim,
+            self.joint_layers,
+            self.zero_init,
+        )
+
+    def forward(self, preprocessed_states: torch.Tensor) -> torch.Tensor:
+        s = preprocessed_states[..., :-1]
+        t = preprocessed_states[..., -1]
+        s_emb = self.s_model(s)
+        t_emb = self.t_model(t)
+        out = self.joint_model(s_emb, t_emb)
+
+        if self.clipping:
+            out = torch.clamp(out, -self.gfn_clip, self.gfn_clip)
+
+        # Tanh-scale to stay near Brownian bridge; last dim (if present) is log-std corr.
+        drift_corr = torch.tanh(out[..., : self.s_dim]) * self.pb_scale_range
+        if self.learn_variance and out.shape[-1] == self.s_dim + 1:
+            log_std_corr = torch.tanh(out[..., [-1]]) * self.pb_scale_range
+            log_std_corr = torch.clamp(
+                log_std_corr, -self.log_var_range, self.log_var_range
+            )
+            out = torch.cat([drift_corr, log_std_corr], dim=-1)
+        else:
+            out = drift_corr
+
+        if torch.isnan(out).any():
+            raise ValueError("DiffusionPISGradNetBackward produced NaNs")
+
+        return out


The new DiffusionPISGradNetBackward class lacks dedicated test coverage. While it's used indirectly in the tutorial script, there are no unit tests that directly verify its behavior (e.g., output shapes, learned corrections, variance handling). Consider adding tests similar to those for DiffusionPISGradNetForward to ensure the backward module works correctly.

I added test coverage.

Copilot · 2025-12-18T02:51:06Z

src/gfn/estimators.py

+        eps = self.dt * _DIFFUSION_TERMINAL_TIME_EPS
+        is_final_step = (t_curr + self.dt) >= (1.0 - eps)
+        # TODO: The old code followed this convention (below). I believe the change
+        #       is slightly more correct, but I'd like to check this during review.
+        # (1.0 - t_curr) < self.dt * 1e-2  # Triggers when t_curr ≈ 1.0


This TODO comment suggests uncertainty about the correctness of the exit condition change. TODOs requesting review in production code should be resolved before merging. Either verify the correctness and remove the TODO, or if there's genuine uncertainty, add a test to validate the behavior.

Suggested change

eps = self.dt * _DIFFUSION_TERMINAL_TIME_EPS

is_final_step = (t_curr + self.dt) >= (1.0 - eps)

# TODO: The old code followed this convention (below). I believe the change

# is slightly more correct, but I'd like to check this during review.

# (1.0 - t_curr) < self.dt * 1e-2 # Triggers when t_curr ≈ 1.0

# Note: this replaces an older heuristic `(1.0 - t_curr) < self.dt * 1e-2`,

# using the shared `_DIFFUSION_TERMINAL_TIME_EPS` tolerance for consistency.

eps = self.dt * _DIFFUSION_TERMINAL_TIME_EPS

is_final_step = (t_curr + self.dt) >= (1.0 - eps)

Copilot · 2025-12-18T02:51:07Z

src/gfn/estimators.py

        """
        assert len(states.batch_shape) == 1, "States must have a batch_shape of length 1"
-        s_curr = states.tensor[:, :-1]
+        # s_curr = states.tensor[:, :-1]


The commented-out line appears to be dead code that should be removed. If it's intended for reference, consider moving it to a comment explaining why the change was made rather than leaving commented code.

Suggested change

# s_curr = states.tensor[:, :-1]

tutorials/examples/train_diffusion_rtb.py

src/gfn/gym/diffusion_sampling.py

src/gfn/estimators.py

Copilot · 2025-12-18T02:51:10Z

src/gfn/gflownet/mle.py

+        if self.debug and torch.isnan(logpf_sum).any():
+            raise ValueError("NaNs in logpf_sum during MLE loss.")
+
+        # TODO: Use included loss reduction helpers.


The TODO comment suggests using included loss reduction helpers, but the current implementation uses a simple conditional. Either implement the TODO or remove it if the current approach is acceptable.

Suggested change

# TODO: Use included loss reduction helpers.

src/gfn/utils/modules.py

Copilot · 2025-12-18T02:51:11Z

src/gfn/gflownet/trajectory_balance.py

+        self.prior_pf = prior_pf
+        self.beta = torch.tensor(beta)
+        self.logZ = logZ or nn.Parameter(torch.tensor(init_logZ))
+        self.debug = debug  # TODO: to be passed to base classes.


The TODO comment about passing debug flag to base classes should be resolved. Either implement the propagation of the debug flag to parent classes or remove the TODO if it's not necessary for this feature.

Suggested change

self.debug = debug # TODO: to be passed to base classes.

self.debug = debug

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Configure codecov to not block CI while tracking coverage

Copilot · 2025-12-18T04:01:13Z

@josephdviviano I've opened a new pull request, #459, to work on those changes. Once the pull request is ready, I'll request review from you.

Co-authored-by: josephdviviano <4142570+josephdviviano@users.noreply.github.com>

Refactor: Extract OUTPUT_DIR constant for visualization paths

josephdviviano · 2025-12-18T15:54:22Z

src/gfn/estimators.py

+        # This matches the exit condition in DiffusionSampling.step() and ensures the
+        # sampled action is marked as an exit action (-inf) so trajectory masks align
+        # correctly in get_trajectory_pbs.
+        eps = self.dt * _DIFFUSION_TERMINAL_TIME_EPS


Should revert change - am terminating one step too early.

josephdviviano added 4 commits December 15, 2025 07:23

rtb prototype

ee17515

sync of RTB prototype

85190b0

ignore outputs

6ecd0ca

rtb finetune first pass

08258ad

josephdviviano requested a review from hyeok9855 December 16, 2025 07:40

josephdviviano self-assigned this Dec 16, 2025

josephdviviano added 4 commits December 16, 2025 03:15

loss bugfixes

6ea54cd

prior learning works, still working on finetune step

a0666da

RTB is working - next step is to factorize

f771c2c

refactored MLE pipeline

51ae63d

josephdviviano marked this pull request as ready for review December 17, 2025 05:52

josephdviviano commented Dec 17, 2025

View reviewed changes

josephdviviano added 6 commits December 17, 2025 01:04

added the MLE trainer

4aebf20

cleaned up training script

f685b86

shrunk script for clarity

e647877

no change

06ccc10

fixed backward bug

bfbbc22

Merge branch 'master' into relative_trajectory_balance

8f8cadb

josephdviviano requested review from saleml and younik December 18, 2025 00:27

josephdviviano added 2 commits December 17, 2025 19:37

changed back to old line lengths

bb2cb45

Merge branch 'relative_trajectory_balance' of github.com:GFNOrg/torch…

69efb35

…gfn into relative_trajectory_balance

josephdviviano requested a review from Copilot December 18, 2025 02:38

Copilot started reviewing on behalf of josephdviviano December 18, 2025 02:38 View session

Initial plan

dd2c4e4

Copilot AI mentioned this pull request Dec 18, 2025

Configure codecov to not block CI while tracking coverage #458

Merged

Configure codecov to not block CI/merge requests

c8eb351

Co-authored-by: josephdviviano <4142570+josephdviviano@users.noreply.github.com>

Copilot AI reviewed Dec 18, 2025

View reviewed changes

josephdviviano and others added 10 commits December 17, 2025 22:13

Update src/gfn/estimators.py

63a4bd1

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update tutorials/examples/train_diffusion_rtb.py

46807ed

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update tutorials/examples/train_diffusion_rtb.py

b40655b

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update tutorials/examples/train_diffusion_rtb.py

74b8a60

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/gfn/utils/modules.py

7deba0e

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/gfn/gflownet/trajectory_balance.py

a2f5a6c

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/gfn/gym/diffusion_sampling.py

ee22f3c

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Merge pull request #458 from GFNOrg/copilot/sub-pr-457

f8d5c5b

Configure codecov to not block CI while tracking coverage

added diffusion tests

ad457f6

Initial plan

f439a78

Copilot AI mentioned this pull request Dec 18, 2025

Refactor: Extract OUTPUT_DIR constant for visualization paths #459

Merged

Copilot AI and others added 2 commits December 18, 2025 04:05

Refactor: Define OUTPUT_DIR constant for visualization output directory

b141f3f

Co-authored-by: josephdviviano <4142570+josephdviviano@users.noreply.github.com>

Merge pull request #459 from GFNOrg/copilot/sub-pr-457-again

97a2453

Refactor: Extract OUTPUT_DIR constant for visualization paths

josephdviviano commented Dec 18, 2025

View reviewed changes

		os.makedirs("output", exist_ok=True)
		plt.savefig(f"output/{prefix}simple_gmm.png")

	self.debug = debug # TODO: to be passed to base classes.
	self.debug = debug

Conversation

josephdviviano commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Uh oh!

codecov bot commented Dec 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

josephdviviano Dec 17, 2025

Choose a reason for hiding this comment

Uh oh!

josephdviviano Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

josephdviviano Dec 17, 2025

Choose a reason for hiding this comment

Uh oh!

josephdviviano commented Dec 18, 2025

Uh oh!

Copilot AI commented Dec 18, 2025

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

josephdviviano Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

josephdviviano Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Copilot AI Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI commented Dec 18, 2025

Uh oh!

josephdviviano Dec 18, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

josephdviviano commented Dec 16, 2025 •

edited

Loading

codecov bot commented Dec 16, 2025 •

edited

Loading