modeld: ort helpers (#34258)

* ort helpers

* import from ort helpers

* import that too

* linter

* linter

* linter

selfdrive/modeld/runners/onnxmodel.py

@@ -1,39 +1,12 @@
-import onnx
-import itertools
 import os
+import onnx
 import sys
 import numpy as np
 from typing import Any
 
 from openpilot.selfdrive.modeld.runners.runmodel_pyx import RunModel
+from openpilot.selfdrive.modeld.runners.ort_helpers import convert_fp16_to_fp32, ORT_TYPES_TO_NP_TYPES
 
-ORT_TYPES_TO_NP_TYPES = {'tensor(float16)': np.float16, 'tensor(float)': np.float32, 'tensor(uint8)': np.uint8}
-
-def attributeproto_fp16_to_fp32(attr):
-  float32_list = np.frombuffer(attr.raw_data, dtype=np.float16)
-  attr.data_type = 1
-  attr.raw_data = float32_list.astype(np.float32).tobytes()
-
-def convert_fp16_to_fp32(onnx_path_or_bytes):
-  if isinstance(onnx_path_or_bytes, bytes):
-    model = onnx.load_from_string(onnx_path_or_bytes)
-  elif isinstance(onnx_path_or_bytes, str):
-    model = onnx.load(onnx_path_or_bytes)
-
-  for i in model.graph.initializer:
-    if i.data_type == 10:
-      attributeproto_fp16_to_fp32(i)
-  for i in itertools.chain(model.graph.input, model.graph.output):
-    if i.type.tensor_type.elem_type == 10:
-      i.type.tensor_type.elem_type = 1
-  for i in model.graph.node:
-    if i.op_type == 'Cast' and i.attribute[0].i == 10:
-      i.attribute[0].i = 1
-    for a in i.attribute:
-      if hasattr(a, 't'):
-        if a.t.data_type == 10:
-          attributeproto_fp16_to_fp32(a.t)
-  return model.SerializeToString()
 
 def create_ort_session(path, fp16_to_fp32):
   os.environ["OMP_NUM_THREADS"] = "4"
@@ -56,7 +29,7 @@ def create_ort_session(path, fp16_to_fp32):
     options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
     provider = 'CPUExecutionProvider'
 
-  model_data = convert_fp16_to_fp32(path) if fp16_to_fp32 else path
+  model_data = convert_fp16_to_fp32(onnx.load(path)) if fp16_to_fp32 else path
   print("Onnx selected provider: ", [provider], file=sys.stderr)
   ort_session = ort.InferenceSession(model_data, options, providers=[provider])
   print("Onnx using ", ort_session.get_providers(), file=sys.stderr)

selfdrive/modeld/runners/ort_helpers.py

@@ -0,0 +1,36 @@
+import onnx
+import onnxruntime as ort
+import numpy as np
+import itertools
+
+ORT_TYPES_TO_NP_TYPES = {'tensor(float16)': np.float16, 'tensor(float)': np.float32, 'tensor(uint8)': np.uint8}
+
+def attributeproto_fp16_to_fp32(attr):
+  float32_list = np.frombuffer(attr.raw_data, dtype=np.float16)
+  attr.data_type = 1
+  attr.raw_data = float32_list.astype(np.float32).tobytes()
+
+def convert_fp16_to_fp32(model):
+  for i in model.graph.initializer:
+    if i.data_type == 10:
+      attributeproto_fp16_to_fp32(i)
+  for i in itertools.chain(model.graph.input, model.graph.output):
+    if i.type.tensor_type.elem_type == 10:
+      i.type.tensor_type.elem_type = 1
+  for i in model.graph.node:
+    if i.op_type == 'Cast' and i.attribute[0].i == 10:
+      i.attribute[0].i = 1
+    for a in i.attribute:
+      if hasattr(a, 't'):
+        if a.t.data_type == 10:
+          attributeproto_fp16_to_fp32(a.t)
+  return model.SerializeToString()
+
+
+def make_onnx_cpu_runner(model_path):
+  options = ort.SessionOptions()
+  options.intra_op_num_threads = 4
+  options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+  options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+  model_data = convert_fp16_to_fp32(onnx.load(model_path))
+  return ort.InferenceSession(model_data,  options, providers=['CPUExecutionProvider'])