Adding prototype flag on reference scripts

[datumbox]

Previously we've added in the reference scripts a --weights argument which operates as a feature switch for the prototype area (None vs fully weights name). When this was first introduced, we just wanted to confirm that the new model builders yielded the same results as the old ones for pre-trained models. Thus we only focused on inference.

Unfortunately the old approach of checking if weights was None won't work if we have models that haven't been graduated from the prototype area and need training. In this PR, we add the explicit feature switch --prototype which allows us to access the prototype model builders.

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 7a83575 (more details on the Dr. CI page):

---

✅ None of the CI failures appear to be your fault 💚

---

• 2/2 broken upstream at merge base 7d4bdd4 since Jan 14

---

🚧 2 ongoing upstream failures:

These were probably caused by upstream breakages that are not fixed yet.

• unittest_prototype since Jan 14 (adf8466)
  • 🔁 rerun
• unittest_linux_gpu_py3.8 since Jan 21 (7d4bdd4)
  • 🔁 rerun

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[datumbox]

I confirm that the arguments in the two classes take the parameters in different order. We are going to remove the old presets anyway and stick with the second one.

[yiwen-song]

Overall LGTM, can you launch 2 jobs to verify it loads prototype models properly?
1 for --prototype and --weights=None
1 for --prototype and --weights=ImageNet1K.xxx

[datumbox]

@sallysyw I've checked, see below. Please stamp to unblock.

torchrun --nproc_per_node=1 train.py --test-only --weights ConvNeXt_Tiny_Weights.ImageNet1K_V1 --model convnext_tiny 
ValueError: The weights parameter works only in prototype mode. Please pass the --prototype argument.

torchrun --nproc_per_node=1 train.py --test-only --prototype --weights ConvNeXt_Tiny_Weights.ImageNet1K_V1 --model convnext_tiny 
Test:  Acc@1 82.522 Acc@5 96.146

torchrun --nproc_per_node=1 train.py --test-only --prototype --model convnext_tiny 
Test:  Acc@1 0.092 Acc@5 0.444

torchrun --nproc_per_node=2 train.py --prototype --model convnext_tiny --batch-size 128 --opt adamw --lr 1e-3 --lr-scheduler cosineannealinglr --lr-warmup-epochs 5 --lr-warmup-method linear --auto-augment ta_wide --epochs 600 --random-erase 0.1 --label-smoothing 0.1 --mixup-alpha 0.2 --cutmix-alpha 1.0 --weight-decay 0.05 --norm-weight-decay 0.0 --train-crop-size 176 --model-ema --val-resize-size 232 --ra-sampler --ra-reps 4

Works as expected, output omitted.

[datumbox]

@sallysyw Thanks for the careful reviews, checked again everything and looks OK.