[Bug?] Casting int8-->float

[pauldog]

Describe the issue

I am getting a weird result using:

import torch
import torch.onnx as onnx

class ByteConverter(torch.nn.Module):
    def __init__(self):
        super(ByteConverter, self).__init__()

    def forward(self, input_array, scale, offset):
        # Convert input array to a PyTorch tensor
        input_tensor = input_array.to(torch.int8)

        # Apply scale and offset to input tensor
        input_tensor = (input_tensor.to(torch.float32) - offset) * scale

        # Convert input tensor to float16
        output_tensor = input_tensor.to(torch.float16)

        return output_tensor

# Instantiate an instance of ByteConverter with scale=1.0 and offset=0.0
model = ByteConverter()


# Define the input tensor
input_array = torch.tensor([1, 50, 100, 150, 200], dtype=torch.uint8)

# Define the scale and offset tensors
scale = torch.tensor([1])
offset = torch.tensor([0])

# Run the model with the input tensor, scale, and offset
output_tensor = model(input_array, scale, offset)

# Print the output tensor
print(output_tensor)


# Export the model to ONNX
input_array = torch.zeros((10), dtype=torch.uint8)  # Example input array
scale = torch.zeros((1), dtype=torch.float32)
offset = torch.zeros((1), dtype=torch.float32)
dynamic_axes = {"input_array": {0: "size"}}  # Specify dynamic axes for input_array
torch.onnx.export(
    model, 
    (input_array, scale, offset), 
    "model.onnx", 
    opset_version=14,
    input_names=["input_array", "scale", "offset"], 
    output_names=["output_tensor"], 
    dynamic_axes=dynamic_axes)

When I run this in Onnx runtime giving byte tensor of new byte[] { 3, 5, 100, 150, 200 }

I expect a float16 array of result [ 1., 50., 100., -106., -56.]

Instead I get the result of [3, 5, 100, 150, 200]

The expected result is that an int8 of -100 gets cast to a float of -100.0

To reproduce

run this python file to build the onnx and feed in a byte tensor, a scale=1 and offset=0.

Same results in both CPU and DML mode.

The graph seems fine:

Urgency

No response

Platform

Windows

OS Version

Windows 10

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.15

ONNX Runtime API

C#

Architecture

X64

Execution Provider

DirectML

Execution Provider Library Version

DirectML

[pauldog]

I am having to do a manual conversion using:

        input_tensor = input_array.to(torch.int16)
        input_tensor = input_tensor - (input_tensor>=128)*256

instead of .to(torch.int8)