add QDQ format onnx model.

testdata/dnn/onnx/generate_quantized_onnx_models.py

@@ -5,9 +5,9 @@
 import torch.nn.functional as F
 import numpy as np
 import os
-import onnx
+import onnx # version >= 1.12.0
 import onnxruntime as rt
-from onnxruntime.quantization import quantize_static, CalibrationDataReader, QuantType
+from onnxruntime.quantization import quantize_static, CalibrationDataReader, QuantType, QuantFormat
 
 class DataReader(CalibrationDataReader):
     def __init__(self, model_path, batchsize=5):
@@ -20,16 +20,16 @@ def __init__(self, model_path, batchsize=5):
     def get_next(self):
         return next(self.enum_data_dicts, None)
 
-def quantize_and_save_model(name, input, model, act_type="uint8", wt_type="uint8", per_channel=False):
+def quantize_and_save_model(name, input, model, act_type="uint8", wt_type="uint8", per_channel=False, ops_version = 13, quanFormat=QuantFormat.QOperator):
     float_model_path = os.path.join("models", "dummy.onnx")
     quantized_model_path = os.path.join("models", name + ".onnx")
     type_dict = {"uint8" : QuantType.QUInt8, "int8" : QuantType.QInt8}
 
     model.eval()
-    torch.onnx.export(model, input, float_model_path, export_params=True, opset_version=12)
+    torch.onnx.export(model, input, float_model_path, export_params=True, opset_version=ops_version)
 
     dr = DataReader(float_model_path)
-    quantize_static(float_model_path, quantized_model_path, dr, per_channel=per_channel,
+    quantize_static(float_model_path, quantized_model_path, dr, quant_format=quanFormat, per_channel=per_channel,
                     activation_type=type_dict[act_type], weight_type=type_dict[wt_type])
 
     os.remove(float_model_path)
@@ -53,10 +53,16 @@ def quantize_and_save_model(name, input, model, act_type="uint8", wt_type="uint8
 
 input = Variable(torch.randn(1, 3, 10, 10))
 conv = nn.Conv2d(3, 5, kernel_size=3, stride=2, padding=1)
+# generate QOperator qunatized model
 quantize_and_save_model("quantized_conv_uint8_weights", input, conv)
 quantize_and_save_model("quantized_conv_int8_weights", input, conv, wt_type="int8")
 quantize_and_save_model("quantized_conv_per_channel_weights", input, conv, per_channel=True)
 
+# generate QDQ qunatized model
+quantize_and_save_model("quantized_conv_uint8_weights_qdq", input, conv, quanFormat=QuantFormat.QDQ)
+quantize_and_save_model("quantized_conv_int8_weights_qdq", input, conv, wt_type="int8", quanFormat=QuantFormat.QDQ)
+quantize_and_save_model("quantized_conv_per_channel_weights_qdq", input, conv, per_channel=True, quanFormat=QuantFormat.QDQ)
+
 input = Variable(torch.randn(1, 3))
 linear = nn.Linear(3, 4, bias=True)
 quantize_and_save_model("quantized_matmul_uint8_weights", input, linear)