Neural Insights fixes (#999)

Signed-off-by: bmyrcha <[email protected]>

examples/onnxrt/nlp/bert/quantization/ptq_dynamic/README.md

@@ -41,13 +41,6 @@ bash prepare_model.sh --input_dir=./MRPC \
 
 ## Diagnosis
 Neural Compressor offers quantization and benchmark diagnosis. Adding `diagnosis` parameter to Quantization/Benchmark config will provide additional details useful in diagnostics.
-### Quantization diagnosis
-```
-config = PostTrainingQuantConfig(
-    diagnosis=True,
-    ...
-)
-``` 
 
 ### Benchmark diagnosis
 ```

examples/onnxrt/nlp/bert/quantization/ptq_dynamic/main.py

@@ -291,12 +291,6 @@ def result(self):
         default=False,
         help="whether quantize the model"
     )
-    parser.add_argument(
-        '--diagnose',
-        dest='diagnose',
-        action='store_true',
-        help='use Neural Insights to diagnose tuning and benchmark.',
-    )
     parser.add_argument(
         '--output_model',
         type=str,
@@ -379,8 +373,6 @@ def eval_func(model):
         return metric.result()
 
     if args.benchmark:
-        if args.diagnose and args.mode != "performance":
-            print("[ WARNING ] Diagnosis works only with performance benchmark.")
         model = onnx.load(args.model_path)
         if args.mode == 'performance':            
             from neural_compressor.benchmark import fit
@@ -389,7 +381,6 @@ def eval_func(model):
                 iteration=100,
                 cores_per_instance=4,
                 num_of_instance=1,
-                diagnosis=args.diagnose,
             )
             fit(model, conf, b_dataloader=dataloader)
         elif args.mode == 'accuracy':
@@ -424,7 +415,6 @@ def eval_func(model):
         from neural_compressor import quantization, PostTrainingQuantConfig
         config = PostTrainingQuantConfig(
             approach='dynamic',
-            diagnosis=args.diagnose,
         )
         q_model = quantization.fit(model, 
                                    config,

examples/onnxrt/nlp/bert/quantization/ptq_static/README.md

@@ -41,13 +41,6 @@ bash prepare_model.sh --input_dir=./MRPC \
 
 ## Diagnosis
 Neural Compressor offers quantization and benchmark diagnosis. Adding `diagnosis` parameter to Quantization/Benchmark config will provide additional details useful in diagnostics.
-### Quantization diagnosis
-```
-config = PostTrainingQuantConfig(
-    diagnosis=True,
-    ...
-)
-``` 
 
 ### Benchmark diagnosis
 ```

examples/onnxrt/nlp/bert/quantization/ptq_static/main.py

@@ -291,12 +291,6 @@ def result(self):
         default=False,
         help="whether quantize the model"
     )
-    parser.add_argument(
-        '--diagnose',
-        dest='diagnose',
-        action='store_true',
-        help='use Neural Insights to diagnose tuning and benchmark.',
-    )
     parser.add_argument(
         "--output_model",
         type=str,
@@ -386,8 +380,6 @@ def eval_func(model):
         return metric.result()
 
     if args.benchmark:
-        if args.diagnose and args.mode != "performance":
-            print("[ WARNING ] Diagnosis works only with performance benchmark.")
         model = onnx.load(args.model_path)
         if args.mode == "performance":            
             from neural_compressor.benchmark import fit
@@ -396,7 +388,6 @@ def eval_func(model):
                 iteration=100,
                 cores_per_instance=4,
                 num_of_instance=1,
-                diagnosis=args.diagnose,
             )
             fit(model, conf, b_dataloader=dataloader)
         elif args.mode == "accuracy":

neural_compressor/benchmark.py

@@ -535,6 +535,7 @@ def fit(model, conf, b_dataloader=None, b_func=None):
             profile(wrapped_model, conf, b_dataloader)
             update_neural_insights_workload(ni_workload_id, "success")
         except Exception as e:
+            logger.error(e)
             update_neural_insights_workload(ni_workload_id, "failure")
 
     logger.info("Start to run Benchmark.")

neural_compressor/profiling/profiler/onnxrt_profiler/profiler.py

@@ -68,16 +68,22 @@ def profile_model(
         Returns:
             None
         """
+        import numpy as np
         import onnxruntime as ort
 
         graph = self.model
         onnx_options = create_onnx_config(ort, intra_num_of_threads, inter_num_of_threads)
         # Create a profile session
         sess_profile = ort.InferenceSession(graph.SerializePartialToString(), onnx_options)
-        input_tensor = sess_profile.get_inputs()[0]
+        input_tensors = sess_profile.get_inputs()
 
         for _, (inputs, _) in enumerate(self.dataloader):
-            input_dict = {input_tensor.name: inputs}
+            if not isinstance(inputs, np.ndarray) and len(input_tensors) != len(inputs):
+                raise Exception("Input data number mismatch.")
+            if len(input_tensors) == 1:
+                input_dict = {input_tensors[0].name: inputs}
+            else:
+                input_dict = {input_tensor.name: input_data for input_tensor, input_data in zip(input_tensors, inputs)}
             sess_profile.run(None, input_dict)
             break
 

neural_insights/components/diagnosis/diagnosis.py

@@ -76,6 +76,9 @@ def load_quantization_config(self) -> dict:
 
     def get_op_list(self) -> List[dict]:
         """Get OP list for model."""
+        check_module("numpy")
+        import numpy as np
+
         minmax_file_path = os.path.join(
             self.workload_location,
             "inspect_saved",
@@ -92,7 +95,7 @@ def get_op_list(self) -> List[dict]:
         for op_name, min_max in min_max_data.items():
 
             mse = self.calculate_mse(op_name, input_model_tensors, optimized_model_tensors)
-            if mse is None:
+            if mse is None or np.isnan(mse):
                 continue
             min = float(min_max.get("min", None))
             max = float(min_max.get("max", None))

neural_insights/components/graph/reader/onnxrt_reader.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 """Onnxrt Graph reader."""
 
+import uuid
 from typing import Any, Dict, List, Optional
 
 from neural_insights.components.graph.attribute import Attribute
@@ -48,25 +49,27 @@ def read(self) -> Graph:
         edges: dict = {}
 
         for node_def in onnx_nodes:
+            node_name = node_def.name
+            if node_name == "":
+                node_name = f"{node_def.op_type}_{uuid.uuid4()}"
+
             if self._should_hide_node(node_def):
-                self._hide_node(node_def)
+                self._hide_node(node_name)
                 continue
 
-            edges = self._add_edges_from_node(edges, node_def)
-
-            current_node_id = node_def.name
+            edges = self._add_edges_from_node(edges, node_def, node_name)
 
             graph.add_node(
                 Node(
-                    id=current_node_id,
+                    id=node_name,
                     label=node_def.op_type,
                     highlight=False,
                     properties={
-                        "name": node_def.name,
+                        "name": node_name,
                         "type": node_def.op_type,
                     },
                     attributes=self._convert_attributes(node_def),
-                    groups=self._get_group_names(current_node_id),
+                    groups=self._get_group_names(node_name),
                 ),
             )
 
@@ -102,9 +105,9 @@ def _node_def_has_not_hidden_inputs(self, node_def: Any) -> bool:
         )
         return len(not_hidden_input_ids) > 0
 
-    def _hide_node(self, node: Any) -> None:
+    def _hide_node(self, node_name: str) -> None:
         """Mark node as hidden."""
-        self._hidden_node_ids[node.name] = True
+        self._hidden_node_ids[node_name] = True
 
     def _is_node_id_hidden(self, node_id: str) -> bool:
         """Check if provided node_id is hidden."""
@@ -161,19 +164,19 @@ def _add_boundary_nodes(self, graph: Graph) -> Graph:
                 )
         return graph
 
-    def _add_edges_from_node(self, edges: dict, node: Any) -> dict:
+    def _add_edges_from_node(self, edges: dict, node: Any, node_name: str) -> dict:
         """Add update edges from node data."""
         for edge in node.input:
             edge_data = edges.get(edge, {})
             input_nodes = edge_data.get("inputs", [])
-            input_nodes.append(node.name)
+            input_nodes.append(node_name)
             edge_data.update({"inputs": input_nodes})
             edges.update({edge: edge_data})
 
         for edge in node.output:
             edge_data = edges.get(edge, {})
             output_nodes = edge_data.get("outputs", [])
-            output_nodes.append(node.name)
+            output_nodes.append(node_name)
             edge_data.update({"outputs": output_nodes})
             edges.update({edge: edge_data})