Delay import onnxruntime to avoid ImportError when onnxruntime is missing and not needed

.azure-pipelines/linux-conda-CI.yml

@@ -37,6 +37,10 @@ jobs:
         python.version: '3.9'
         ONNX_PATH: onnx==1.12.0
 
+      Python310_112:
+        python.version: '3.10'
+        ONNX_PATH: onnx==1.12.0
+
 
     maxParallel: 4
 
@@ -56,12 +60,22 @@ jobs:
       conda install -c conda-forge protobuf
       conda install -c conda-forge numpy
       pip install -r requirements.txt
+      pip install pytest
+    displayName: 'Install dependencies'
+
+  - script: |
+      python -c "import onnxconverter_common"
+    displayName: 'Check import without onnxruntime'
+
+  - script: |
       pip install onnxruntime
       pip install onnxmltools
       pip install $(ONNX_PATH)
-      pip install pytest
+    displayName: 'Install onnxruntime, onnxmltools'
+
+  - script: |
       pip install -e .
-    displayName: 'Install dependencies'
+    displayName: 'Install onnxconverter_common'
 
   - script: |
       pip install flake8

onnxconverter_common/auto_mixed_precision.py

@@ -26,7 +26,6 @@ def validate(res1, res2):
 
 """
 
-import onnxruntime as ort
 import onnx
 import numpy as np
 from onnxconverter_common import float16
@@ -127,6 +126,8 @@ def add_missing_dtypes_using_ort(model, feed_dict, outputs_per_iter=100):
 
 
 def get_tensor_values_using_ort(model, input_feed, output_names=None, sess_options=None):
+    # delayed import to avoid taking a strong dependancy on onnxruntime
+    import onnxruntime as ort
     if output_names is None:
         sess = ort.InferenceSession(model.SerializeToString(), sess_options, providers=['CUDAExecutionProvider'])
         return sess.run(None, input_feed)

onnxconverter_common/auto_mixed_precision_model_path.py

@@ -12,7 +12,6 @@
 
 import copy
 import numpy as np
-import onnxruntime as ort
 import onnx
 import os
 import uuid
@@ -115,7 +114,8 @@ def validate(res1, res2):
         kwargs["is_final_model"] = False
         result = _convert_and_check_inference_result(**kwargs)
         if not result:
-            raise ValueError("Validation failed for model with nothing converted to fp16.")
+            raise ValueError("Validation failed for model with nothing converted to fp16. "
+                             "Given parameters %r." % kwargs)
 
         final_block_list = _find_nodes_blocking_fp16(**kwargs)
 
@@ -245,6 +245,8 @@ def _convert_and_check_inference_result(**kwargs):
 
 
 def inference(model_path, input_feed, providers=None):
+    # delayed import to avoid taking a strong dependancy on onnxruntime
+    import onnxruntime as ort
     sess = ort.InferenceSession(model_path, None, providers=providers)
     output = sess.run(None, input_feed)
     return output