fix(bnn_pynq): reload the best checkpoint for the QCDQ ONNX export too

[Anai-Guo]

What

In Trainer.train_model, the gate that reloads the best checkpoint before exporting ORs args.export_qonnx with itself:

best_path = os.path.join(self.checkpoints_dir_path, "best.tar")
if self.args.export_qonnx or self.args.export_qonnx:      # <- export_qcdq_onnx
    self.model = self.load_checkpoint(self.model, best_path, strict=True)
if self.args.export_qonnx:
    self.export_qonnx()
if self.args.export_qcdq_onnx:
    self.export_qcdq_onnx()

Why it matters

Running bnn_pynq_train.py --export_qcdq_onnx without --export_qonnx:

1. Trainer.__init__ creates the ONNX output dir (that check is spelled correctly).
2. Trainer.validate asserts JIT is disabled (also spelled correctly).
3. The reload above is skipped, because only export_qonnx is inspected.
4. self.export_qcdq_onnx() then exports self.model as it stands after the last epoch, not the best checkpoint.

So the QCDQ export silently ships the wrong weights, with no error and no warning. Every other place in this file that means "some ONNX export was requested" already writes both flags:

• trainer.py:94 — if args.export_qonnx or args.export_qcdq_onnx: (create the output dir)
• trainer.py:215 — if args.export_qonnx or args.export_qcdq_onnx: (assert JIT is off)

trainer.py:349 is the only one that does not, which is what this PR fixes.

Not affected

• --export_qonnx (alone or with --export_qcdq_onnx): condition was already true, stays true.
• Neither flag: condition was already false, stays false.

🤖 Generated with Claude Code