Code Review: automated questions

[bes-dev]

Code Review Questions

The following questions and observations were raised during automated review of bes-dev/MobileStyleGAN.pytorch.

1. In train.py (~line 46), raise "Unknown export format." raises a plain string, not an exception — this is a TypeError in Python 3 and silently swallows the intended error message. Should be raise ValueError(...).
2. In core/models/synthesis_network.py SynthesisNetwork.forward, _style = style[..., 3*i+1:3*i+4, :] is computed but m(hidden, style, _noise) passes the full style tensor — the per-layer slice is never used, breaking W+ style mixing in the teacher network.
3. In core/models/utils.py NoiseManager.__init__, the guard if not None in noise checks whether any element is None and skips the entire list if so — likely intended to be a per-element check (if noise[i] is not None), causing all cached noise to be silently dropped whenever one entry is None.
4. In core/distiller.py validation_step, self.student(style, noise=gt["noise"]) is called twice — the first result is discarded and the student is run again for loss. One forward pass is redundant and doubles inference cost during validation.
5. In core/distiller.py validation_epoch_end, KID is computed by concatenating local-device outputs only with a # TODO: add all_gather for distributed mode comment. With multi-GPU DDP training the metric is incorrect and the TODO has no associated tracking issue.
6. In core/distiller.py, teacher networks mapping_net and synthesis_net are only set to .eval() but never have requires_grad_(False) applied. Gradients still flow into teacher parameters during student training, wasting memory and compute.
7. In core/distiller.py compute_mean_style, the batch_size parameter is accepted but ignored — torch.randn(4096, ...) is hardcoded. The method always uses 4096 samples regardless of the argument.
8. In core/loss/distiller_loss.py loss_g, gt["img"] and gt["freq"] tensors from the teacher are not detached before computing L1/L2/perceptual losses, potentially letting gradients flow into the frozen teacher synthesis network.
9. In core/models/modules/modulated_conv2d.py, ModulatedConv2d.get_demodulation uses self.style_inv (a randomly-initialised buffer, never updated) instead of the actual style vector. Demodulation is therefore style-independent, diverging from the StyleGAN2 weight-demodulation spec.
10. In core/model_zoo.py, json.load(open(zoo_path)) opens a file handle that is never closed. Should use with open(zoo_path) as f: json.load(f) to avoid resource leaks.
11. In core/models/modules/idwt_upsample.py, the class is named IDWTUpsaplme (typo — letters transposed). It is referenced by this name in mobile_synthesis_block.py, making refactoring error-prone.
12. In core/loss/non_saturating_gan_loss.py loss_d, both fake.detach() and real.detach() are called even though fake is already detached by the caller in Distiller.discriminator_step. The redundant .detach() on real could silently prevent R1 gradient computation if reg_d is mistakenly called through loss_d.
13. In core/distiller.py make_sample, the else branch (pure noise path) constructs style = self.mapping_net(var).view(1, self.wsize, ...) with batch-size 1 unconditionally, regardless of self.cfg.batch_size, so training occasionally runs with mismatched batch sizes without error.
14. In core/models/modules/noise_injection.py NoiseInjection.forward, the noise buffer is registered on first call only (if not hasattr(self, "noise") and self.trace_model). Subsequent CoreML/ONNX traces reuse stale noise, but the buffer is never reset between exports, risking silent reuse of wrong-shape noise across multiple export calls.
15. In requirements.txt, pytorch-lightning==1.0.2 is pinned to a 2020 release. ModelCheckpoint's filepath kwarg was deprecated and removed in later PL versions; the pinned version also predates security fixes. There is no upper bound or lock file to prevent breakage from transitive dependency updates.

---

Generated by automated code review pipeline.