workaround observer shapes seem to be important for EMA QAT model to have correct min_val/max_val

Summary:
Before this fix, the EMA GAN model will have inf min_val/max_val when converting QAT models to int8 torchscript model (as shown in f290518237).

https://pxl.cl/1MNx0

Reviewed By: yc-fb

Differential Revision: D23387923

fbshipit-source-id: 5c2119e2c5170e30c6059e7374c22e367fcd2b26

d2go/runner/default_runner.py

@@ -255,6 +255,14 @@ class Detectron2GoRunner(BaseRunner):
         if cfg.QUANTIZATION.QAT.ENABLED:
             # Disable fake_quant and observer so that the model will be trained normally
             # before QAT being turned on (controlled by QUANTIZATION.QAT.START_ITER).
+            if hasattr(model, "get_rand_input"):
+                model = setup_qat_model(
+                    cfg, model, enable_fake_quant=eval_only, enable_observer=True
+                )
+                imsize = cfg.INPUT.MAX_SIZE_TRAIN
+                rand_input = model.get_rand_input(imsize)
+                model(rand_input, {})
+            else:
                 model = setup_qat_model(
                     cfg, model, enable_fake_quant=eval_only, enable_observer=False
                 )