Add logic to convert fp16 in parse instance norm, add flags for numpy in accuracy checker

src/onnx/parse_instancenorm.cpp

@@ -39,22 +39,39 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
     instruction_ref parse(const op_desc& opd,
                           const onnx_parser& parser,
                           onnx_parser::node_info info,
-                          std::vector<instruction_ref> args) const
+                          std::vector<instruction_ref> oargs) const
     {
         // y = scale * ( x - mean ) / sqrt ( variance + epsilon ) + bias
         // mean = reduce_mean({D1, D2, ... Dk}, x)
         // variance = reduce_mean({D1, D2, ... Dk}, (x - mean)^2)
-
-        float epsilon = 1e-5f;
+        bool convert_fp16 = true;
+        float epsilon     = 1e-5f;
         if(contains(info.attributes, "epsilon"))
         {
             epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
         }
+        auto dtype         = oargs[0]->get_shape().type();
+        auto literal_dtype = dtype;
+        std::vector<instruction_ref> args;
+        if(dtype == shape::half_type and convert_fp16)
+        {
+            args.push_back(info.add_instruction(
+                make_op("convert", {{"target_type", shape::float_type}}), oargs[0]));
+            args.push_back(info.add_instruction(
+                make_op("convert", {{"target_type", shape::float_type}}), oargs[1]));
+            args.push_back(info.add_instruction(
+                make_op("convert", {{"target_type", shape::float_type}}), oargs[2]));
+            literal_dtype = shape::float_type;
+        }
+        else
+        {
+            args = oargs;
+        }
+
         auto x     = args[0];
         auto scale = args[1];
         auto bias  = args[2];
         auto dims  = x->get_shape().lens();
-        auto dtype = x->get_shape().type();
         if(not contains(valid_types, dtype))
             MIGRAPHX_THROW(opd.op_name + ": invalid output type: " + std::to_string(dtype) +
                            ". Valid types are 1 (float), 10 (half), and 11 (double).");
@@ -72,7 +89,7 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
         auto l0              = info.add_instruction(make_op("sqdiff"), x, mean_bcast);
         auto variance        = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), l0);
         auto l1              = info.add_instruction(make_op("sub"), x, mean_bcast);
-        auto epsilon_literal = info.add_literal(literal{shape{dtype}, {epsilon}});
+        auto epsilon_literal = info.add_literal(literal{shape{literal_dtype}, {epsilon}});
         auto epsilon_bcast =
             info.add_instruction(make_op("multibroadcast", {{"out_lens", dims}}), epsilon_literal);
         auto variance_bcast =
@@ -85,8 +102,14 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
         ;
         auto bias_bcast =
             info.add_instruction(make_op("broadcast", {{"axis", 1}, {"out_lens", dims}}), bias);
-        auto l5 = info.add_instruction(make_op("mul"), l4, scale_bcast);
-        return info.add_instruction(make_op("add"), l5, bias_bcast);
+        auto l5  = info.add_instruction(make_op("mul"), l4, scale_bcast);
+        auto ret = info.add_instruction(make_op("add"), l5, bias_bcast);
+        if(dtype == shape::half_type and convert_fp16)
+        {
+            return info.add_instruction(make_op("convert", {{"target_type", shape::half_type}}),
+                                        ret);
+        }
+        return ret;
     }
 };
 

tools/accuracy/accuracy_checker.py

@@ -52,6 +52,8 @@ def parse_args():
     parser.add_argument('--fill0',
                         action='store_true',
                         help='fill all arguments with a value of 0')
+    parser.add_argument('--numpy', action='append', help='fill argument with numpy saved array', type=str)
+    parser.add_argument('--np_path', action='append', help='Path for the saved numpy array', type=str)
     parser.add_argument('--verbose',
                         action='store_true',
                         help='show verbose information (for debugging)')
@@ -81,7 +83,7 @@ def parse_args():
                         action='store_true',
                         default=False,
                         help='Turn on ort VERBOSE logging via session options')
-
+    parser.add_argument('--save-ort-res', dest="save_ort", type=str, help='Save output of ORT as numpy array at path provided by this argument')
     args = parser.parse_args()
 
     return args
@@ -186,7 +188,12 @@ def main():
             print(f'Parameter {name} -> {shape}')
         in_shape = shape.lens()
         in_type = shape.type_string()
-        if not args.fill1 and not args.fill0:
+        if name in args.numpy:
+            in_path = args.np_path[args.numpy.index(name)]
+            if args.verbose:
+                print("Loading numpy array for input name : {name}, from path : {in_path}")
+            test_input = np.load(in_path).astype(get_np_datatype(in_type))
+        elif not args.fill1 and not args.fill0:
             test_input = np.random.rand(*(in_shape)).astype(
                 get_np_datatype(in_type))
         elif not args.fill0:
@@ -216,6 +223,10 @@ def main():
 
         try:
             pred_fw = sess.run(None, ort_params)[-1]
+            if(args.save_ort): 
+                if args.verbose:
+                    print("saving ORT result as numpy array at location : {args.save_ort}")
+                np.save(args.save_ort, pred_fw)
         except Exception as e:
             if any(input_dims):
                 print(