[Relay] Improved `multiply_rewrite` for gluoncv_ssd quantization (#2848)

* improved `multiply_rewrite` for gluoncv_ssd quantization

* add a check to attach_simulated_quantize below to prevent quantizing lhs to INPUT if lhs_kind is INPUT;

diff --git a/python/tvm/relay/quantize/_annotate.py b/python/tvm/relay/quantize/_annotate.py
index c757468..08d56a3 100644 (file)
--- a/python/tvm/relay/quantize/_annotate.py
+++ b/python/tvm/relay/quantize/_annotate.py
@@ -166,9 +166,10 @@ def multiply_rewrite(ref_call, new_args, ctx):
 
     if lhs_kind is None and rhs_kind is None:
         return None
-    if lhs_kind == QAnnotateKind.ACTIVATION and rhs_kind is None:
+    if lhs_kind in [QAnnotateKind.ACTIVATION, QAnnotateKind.INPUT] and rhs_kind is None:
         # quantize lhs to INPUT field
-        lhs_expr = attach_simulated_quantize(lhs_expr, QAnnotateKind.INPUT)
+        if lhs_kind == QAnnotateKind.ACTIVATION:
+            lhs_expr = attach_simulated_quantize(lhs_expr, QAnnotateKind.INPUT)
         # quantize rhs to WEIGHT field
         rhs_expr = attach_simulated_quantize(rhs_expr, QAnnotateKind.WEIGHT)
         expr = _forward_op(ref_call, [lhs_expr, rhs_expr])