[fbsync] Document MobileNetV3 quantization approach (#5891)

Summary:
* Document MobileNetV3 quantization approach

* Change the URL.

Reviewed By: jdsgomes, NicolasHug

Differential Revision: D36095726

fbshipit-source-id: f9a6d1a35c53e890963bd9248d1d9c77e1e48fbb

Author: YosuaMichael

torchvision/models/quantization/mobilenetv3.py

@@ -140,6 +140,10 @@ def _mobilenet_v3_model(
     _replace_relu(model)
 
     if quantize:
+        # Instead of quantizing the model and then loading the quantized weights we take a different approach.
+        # We prepare the QAT model, load the QAT weights from training and then convert it.
+        # This is done to avoid extremely low accuracies observed on the specific model. This is rather a workaround
+        # for an unresolved bug on the eager quantization API detailed at: https://github.com/pytorch/vision/issues/5890
         model.fuse_model(is_qat=True)
         model.qconfig = torch.ao.quantization.get_default_qat_qconfig(backend)
         torch.ao.quantization.prepare_qat(model, inplace=True)