[BYOC][FIX] Infer types in MergeComposite (apache#5766)

If InferType isn't run between partitioning passes,
function calls are inserted which don't have a type.
This can result in failures for patterns which want
to check types.

This works around it simply by running InferType after
every partitioning.

Change-Id: Ie0887f0564a41eb0913bfe42a362e8effe9681b9

Author: mbaret

src/relay/transforms/merge_composite.cc

@@ -36,17 +36,24 @@ namespace tvm {
 namespace relay {
 namespace merge_composite {
 
+Function InferType(const Function& expr) {
+  auto mod = IRModule::FromExpr(expr);
+  mod = transform::InferType()(mod);
+  return Downcast<Function>(mod->Lookup("main"));
+}
+
 Expr MergeComposite(const Function& func, const Array<runtime::String>& pattern_names,
                     const Array<DFPattern>& patterns, const std::vector<PackedFunc>& checks) {
   CHECK_EQ(pattern_names.size(), patterns.size());
-  Expr merged_expr = func->body;
+  Function merged_func = func;
   // merge the patterns one-by-one in order
   for (size_t i = 0; i < patterns.size(); i++) {
     Map<String, ObjectRef> attrs;
     attrs.Set("Composite", pattern_names[i]);
-    merged_expr = PartitionPattern(patterns[i], merged_expr, attrs, checks[i]);
+    merged_func = Downcast<Function>(PartitionPattern(patterns[i], merged_func, attrs, checks[i]));
+    merged_func = InferType(merged_func);
   }
-  return Function(func->params, merged_expr, func->ret_type, func->type_params, func->attrs);
+  return std::move(merged_func);
 }
 
 }  // namespace merge_composite

tests/python/relay/test_pass_merge_composite.py

@@ -916,31 +916,63 @@ def before():
         x = relay.var('x', shape=(1, 10, 10, 10))
         w = relay.var('w', shape=(10, 10, 3, 3))
         b = relay.var('b', shape=(8,))
-        conv = relay.nn.conv2d(x,
+        add = relay.op.add(x, x)
+        relu = relay.nn.relu(add)
+        conv = relay.nn.conv2d(relu,
                                w,
                                kernel_size=(3, 3),
                                kernel_layout="OIHW",
                                data_layout="NHWC")
         bias = relay.nn.bias_add(conv, b)
-        relu = relay.nn.relu(bias)
-        return relay.Function([x, w, b], relu)
+        relu2 = relay.nn.relu(bias)
+        return run_opt_pass(relay.Function([x, w, b], relu2), relay.transform.InferType())
 
-    def expected():
-        x = relay.var('x')
-        w = relay.var('w')
-        b = relay.var('b')
-        conv = relay.nn.conv2d(x, w, kernel_size=(3, 3), kernel_layout="OIHW", data_layout="NHWC")
+    def expected_false():
+        x = relay.var('x', shape=(1, 10, 10, 10))
+        w = relay.var('w', shape=(10, 10, 3, 3))
+        b = relay.var('b', shape=(8, ))
+
+        x0 = relay.var('x')
+        y0 = relay.var('y')
+
+        add = relay.op.add(y0, y0)
+        relu = relay.nn.relu(add)
+        func = relay.Function([x0, y0], relu)
+        func = func.with_attr("PartitionedFromPattern", "add_nn.relu_")
+        func = func.with_attr("Composite", "add_relu")
+        call = relay.Call(func, [x, x])
+
+        conv = relay.nn.conv2d(call, w, kernel_size=(3, 3), kernel_layout="OIHW", data_layout="NHWC")
         bias = relay.nn.bias_add(conv, b)
-        relu = relay.nn.relu(bias)
-        func = relay.Function([x, w, b], relu)
-        func = func.with_attr("Composite", "conv_bias_relu")
-        func = func.with_attr("PartitionedFromPattern", "nn.conv2d_nn.bias_add_nn.relu_")
+        relu2 = relay.nn.relu(bias)
+        return relay.Function([x, w, b], relu2)
 
+    def expected_true():
         x = relay.var('x', shape=(1, 10, 10, 10))
         w = relay.var('w', shape=(10, 10, 3, 3))
         b = relay.var('b', shape=(8, ))
-        return relay.Function([x, w, b], func(x, w, b))
 
+        x0 = relay.var('x')
+        y0 = relay.var('y')
+
+        add = relay.op.add(y0, y0)
+        relu = relay.nn.relu(add)
+        func = relay.Function([x0, y0], relu)
+        func = func.with_attr("PartitionedFromPattern", "add_nn.relu_")
+        func = func.with_attr("Composite", "add_relu")
+        call = relay.Call(func, [x, x])
+
+        x2 = relay.var('x')
+        w1 = relay.var('w')
+        b1 = relay.var('b')
+        conv = relay.nn.conv2d(x2, w1, kernel_size=(3, 3), kernel_layout="OIHW", data_layout="NHWC")
+        bias = relay.nn.bias_add(conv, b1)
+        relu2 = relay.nn.relu(bias)
+        func = relay.Function([x2, w1, b1], relu2)
+        func = func.with_attr("Composite", "conv_bias_relu")
+        func = func.with_attr("PartitionedFromPattern", "nn.conv2d_nn.bias_add_nn.relu_")
+        call = relay.Call(func, [call, w, b])
+        return relay.Function([x, w, b], call)
 
     def _check_type_true(extract):
         conv = extract.args[0].args[0]
@@ -953,14 +985,16 @@ def _check_type_false(extract):
         return bool(typ.shape[0] != 1)
 
     pattern_table_false = [
+        ("add_relu", make_add_relu_pattern()),
         ("conv_bias_relu", make_conv_bias_relu_pattern(), _check_type_false)
     ]
-    check_result(pattern_table_false, before(), before())
+    check_result(pattern_table_false, before(), expected_false())
 
     pattern_table_true = [
+        ("add_relu", make_add_relu_pattern()),
         ("conv_bias_relu", make_conv_bias_relu_pattern(), _check_type_true)
     ]
-    check_result(pattern_table_true, before(), expected())
+    check_result(pattern_table_true, before(), expected_true())
 
 
 if __name__ == "__main__":