resize op convert to FP16 fail

[nistarlwc]

There is a model from tensorflow2onnx, the FP32 model can run successfully.

Then use float16_converter.convert_float_to_float16(onnx_model, keep_io_types=True) convert to FP16 model.
But the FP16 model can't create session, error:
onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Load model from model_fp16.onnx failed:Node (Resize__846) Op (Resize) [ShapeInferenceError] Either sizes or scales must be provided, but not both of them

The problem is similar with #266.
How to solve it ?

[nistarlwc]

@xiaowuhu @xadupre @skottmckay
Thanks a lot 🙏

[xadupre]

The model you shared cannot be accessed. Can you try something like:

import onnx
m1 = onnx.load("<first model>")
print([n for n in m1.graph.node if n.op_type == "Resize"])
m2 = onnx.load("<second model>")
print([n for n in m1.graph.node if n.op_type == "Resize"])

My guess is the operartor Resize changed in opset 13 but the tool was not tested for the new version.

[nistarlwc]

The model you shared cannot be accessed. Can you try something like:

import onnx
m1 = onnx.load("<first model>")
print([n for n in m1.graph.node if n.op_type == "Resize"])
m2 = onnx.load("<second model>")
print([n for n in m1.graph.node if n.op_type == "Resize"])

My guess is the operartor Resize changed in opset 13 but the tool was not tested for the new version.

无法访问您共享的模型。你可以尝试这样的东西：

导入onnx
m1=onnx。负载（“第一模型“）
print（[n] forninm1. graph. nodeifn. op_type==“调整大小”]）
m2=onnx。负载（“第二模型“）
print（[n] forninm1. graph. nodeifn. op_type==“调整大小”]）

我的猜测是operartor Resize在opset 13中发生了变化，但该工具没有在新版本中进行测试。

first model:

[input: "BiseNetV2/aggregation_branch/guided_aggregation_block/semantic_branch/3x3_conv_block/bn/FusedBatchNormV3:0"
input: "roi__879"
input: "roi__879"
input: "Concat__796:0"
output: "Resize__797:0"
name: "Resize__797"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "exclude_outside"
  i: 0
  type: INT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
domain: ""
, input: "BiseNetV2/aggregation_branch/guided_aggregation_block/aggregation_features/guided_semantic_features:0"
input: "roi__879"
input: "roi__879"
input: "Concat__845:0"
output: "Resize__846:0"
name: "Resize__846"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "exclude_outside"
  i: 0
  type: INT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
domain: ""
, input: "BiseNetV2/logits/1x1_conv_block/Conv2D:0"
input: "roi__879"
input: "roi__879"
input: "Concat__887:0"
output: "Resize__888:0"
name: "Resize__888"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "exclude_outside"
  i: 0
  type: INT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
domain: ""
]

[input: "Resize__797_input_cast_0"
input: "Resize__797_input_cast_1"
input: "Resize__797_input_cast_2"
input: "Concat__796:0"
output: "Resize__797_output_cast_0"
name: "Resize__797"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "exclude_outside"
  i: 0
  type: INT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
domain: ""
, input: "Resize__846_input_cast_0"
input: "Resize__846_input_cast_1"
input: "Resize__846_input_cast_2"
input: "Concat__845:0"
output: "Resize__846_output_cast_0"
name: "Resize__846"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "exclude_outside"
  i: 0
  type: INT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
domain: ""
, input: "Resize__888_input_cast_0"
input: "Resize__888_input_cast_1"
input: "Resize__888_input_cast_2"
input: "Concat__887:0"
output: "Resize__888_output_cast_0"
name: "Resize__888"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "exclude_outside"
  i: 0
  type: INT
}
attribute {
  name: "mode"
  s: "linear"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
domain: ""
]

The 2 models are converted from tensorflow-onn with opset 15.

Try to use opset 12, the fp16 model can create session, but don't have result when predict.

All model in GoogleDrive
Thanks a lot

[nistarlwc]

@xadupre
I try to train a pytorch model of Bisenet-V2, that can run fp16 model.
I compared the two onnx models of pytorch and tensorflow and found that tensorflow's RESIZE OP has two more parameters, scales and roi.

So I try to delete the parameters, use

        nodes = onnx_model.graph.node
        node_delete_list = []
        for i, node in enumerate(nodes[:]):
            node_op_type = node.op_type
            if node.op_type == "Resize":
                attrs = node.input
                for j, attr in enumerate(attrs[:]):
                    if 'const_empty_' in attr:
                        node_delete_list.append(attr)
                        nodes[i].input[j] = ''

The RESIZE OP are changed like:

The fp16 model can create session, but don't have result.
I have no idea how to solve this problem.