add models

export-onnx.py

@@ -162,7 +162,9 @@ def main():
         "neg_mean": neg_mean,
         "inv_stddev": inv_stddev,
         "model_type": "sense_voice_ctc",
-        "version": "1",
+        # version 1: Use QInt8
+        # version 2: Use QUInt8
+        "version": "2",
         "model_author": "iic",
         "maintainer": "k2-fsa",
         "vocab_size": vocab_size,
@@ -185,7 +187,10 @@ def main():
         model_input=filename,
         model_output=filename_int8,
         op_types_to_quantize=["MatMul"],
-        weight_type=QuantType.QInt8,
+        # Note that we have to use QUInt8 here.
+        #
+        # When QInt8 is used, C++ onnxruntime produces incorrect results
+        weight_type=QuantType.QUInt8,
     )
 
 

model.int8.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ecb4fc1f94df005d851ca3e72f966fd04bd69b2eb25025727b10c162d3df096e
-size 239233683
+oid sha256:c71f0ce00bec95b07744e116345e33d8cbbe08cef896382cf907bf4b51a2cd51
+size 239233841

model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:03b38fc8fde6a821f0953a980399c34eb38787a9a3094b73d956335049bc2612
+oid sha256:977016bd9c79f9eb343430b5cc305e07ab64d5212dff41b0dcfa1694bee9a8cb
 size 937617178