[RyanPCo]

Fixes two defects that would confound any backbone comparison: a frozen ResNet backbone still ran BatchNorm in train mode, and the pretrained backbone had to share one LR with the randomly initialised trunk.

Behaviour changes

None for shipped HPT configs. No HPT config freezes the ResNet, so no run to date hit the BatchNorm bug, and backbone_lr_scale defaults to 1.0, which builds the optimizer exactly as before.

What changed

• Frozen backbone stays in eval mode. requires_grad = False leaves BatchNorm in train mode, so a "frozen" backbone normalized with batch statistics and kept updating its running stats. ResNet.train() now puts a frozen backbone back in eval mode.
• Backbone LR group. The new model.backbone_lr_scale key puts the encoders' backbone parameters (backbone_parameters()) in their own param group at lr * scale.
  • At 1.0 the optimizer is a single group over trainer.model.parameters(), as before.
  • Any other value needs an explicit model.optimizer.lr and raises without one.
  • Frozen parameters and encoders that don't declare backbone_parameters() stay in the main group.
  • hpt_bc_mecka_6d_300M.yaml declares backbone_lr_scale: 1.0.
• Pretrained marker. ResNet declares _hpt_pretrained_attrs = ("net",), the same marker the text stems use, so a selective-init pass can find its pretrained weights. Nothing on this branch reads the marker yet; later branches in the stack do.

Usage

The RDT recipe (#680) is the first to use the two-group path: rdt_bc_mecka_6d_dinov3_vitb.yaml sets backbone_lr_scale: 0.1.

Tests

New tests/unit/test_vision_backbone.py: ResNet forward against a reference, BatchNorm mode and running stats under a frozen backbone, backbone_parameters(), the single-group default, the two-group split, frozen-parameter handling, and the base model config's backbone_lr_scale: 1.0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HjCQBqQCr2igFEMR7yUovw

[RyanPCo]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• hpt: never re-initialise pretrained stems; hash-check them; mask text padding #645 : 2 dependent PRs (#646 , #664 )
• obs: proprio history (K proprio steps, one token each, learned time embedding) #650
• loss: mask the repeat-last action padding out of the HPT loss #649
• resnet: keep a frozen backbone in eval mode; give backbones their own LR group #652 👈 (View in Graphite)
• augs: per-sample image augmentation for HPT train-mode augs #651
• eval: deterministic validation (seeded prompts + flow sampling); log-only grad-norm MAD #644
• data: mecka_fold_flagship_top3_hpt_6d, the 3-operator flagship train split #643
• val: subsampled metric loaders + pinned contiguous video loaders #642
• infra: mecka 6D / keypoint recipes, 300M models, defaults, and the rollout revert #640
• eval: layout-aware frame-invariant metrics, and a train-split viz head #659
• data: stop letting one bad episode or a rotation channel take a run down #658
• norm: guard degenerate channels, horizon pooling, and no rot6d normalization #657
• keypoints: the 144-D wrist-first hand-keypoint action becomes the human default #656
• mecka: fix the double-mirrored LEFT wrist frame, and correct existing zarrs #655
• 6d: PI goes 6D-only; Eva and Human gain the continuous-rotation modes #654
• 6d: continuous rotation representation, bimanual layouts, shared norm helpers #653
• train: give each Slurm slurm job gets its own cache dir to avoid conflict #638 : 1 other dependent PR (#639 )
• write git version into the run dir for reproducibility #637
• train: don't load PI base weights on resume and verify that checkpoint path exists #636
• configs: declare proprio and actions dims once per HPT model config #634 : 1 other dependent PR (#635 )
• cache norm stats for data configs to minimize recomputation #633
• import cloud packages lazily to support local training #631 : 1 other dependent PR (#632 )
• cache dataset resolver for faster training initialization #630
• validate episode hashes in data config before training #628
• modify pi camera map to match data #629
• add tests to ensure training works between every model and data vendor #626
• run CI tests on every PR #625
• add env variables for user-specific paths #624
• build: set pyproject and uv as the dependency managers #623
• fix(configs): fix paths for data hydra yaml configs #622
• refactor(algo): delete ACT #627 : 1 other dependent PR (#621 )
• tests: create dedicated test folder with unit and integration tests #620
• add pyright type checker #619
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[RyanPCo]

Automated adversarial review (read-only; findings were checked by a second skeptical pass before posting): 3 nit.

[RyanPCo]

nit (simplification): nn.Module.eval() recurses, so self.net.eval() already handles the ModuleList case. The isinstance branch can go: if self.freeze_backbone: self.net.eval().

[RyanPCo]

nit (simplification): groups is None exactly when scale == 1.0, and the two branches build the optimizer differently (instantiate with params plus a callable check, versus _partial_=True then a call). Use one path: factory = hydra.utils.instantiate(cfg.model.optimizer, _partial_=True), params = self._backbone_param_groups(lr, scale) if scale != 1.0 else self.trainer.model.parameters(), optimizer = factory(params=params).

[RyanPCo]

nit (comments): Drop the "0.15:" prefix: it refers to an internal ideas list.

[RyanPCo]

Review findings: status after the fixup pass. The fixes are squashed into this PR's commits, and the whole stack was force-pushed.

Fixed

• ModuleList special-case in train() is unnecessary: Now just self.net.eval().
• configure_optimizers has a groups = None sentinel and two instantiate paths: Single partial factory path; params are the backbone groups or all parameters.
• Private plan number in a test docstring: Dropped '0.15:'.