aaltonenzhang opened a new issue #7258:
URL: https://github.com/apache/tvm/issues/7258
I have a tflite model with float32 feature map and int8 weights, when I look
into convert_conv() in tflite frontend, I found that _qnn.op.conv2d is
constructed only if feature map is int8(uint8) typed. So, I tried to modify
code like:
>
if (input_tensor.qnn_params or weight_tensor.qnn_params):
qnn_conv2d_params = dict(params)
qnn_conv2d_params["input_zero_point"] =
input_tensor.qnn_params["zero_point"] if input_tensor.qnn_params else
relay.const(0, "int32")
qnn_conv2d_params["kernel_zero_point"] =
weight_tensor.qnn_params["zero_point"] if weight_tensor.qnn_params else
relay.const(0, "int32")
qnn_conv2d_params["out_dtype"] = "int32" if
(input_tensor.qnn_params and weight_tensor.qnn_params) else "float32"
qnn_conv2d_params["input_scale"] =
input_tensor.qnn_params["scale"] if input_tensor.qnn_params else relay.const(1,
"float32")
qnn_conv2d_params["kernel_scale"] =
weight_tensor.qnn_params["scale"] if weight_tensor.qnn_params else
relay.const(1, "float32")
out = _qnn.op.conv2d(in_expr, weight_expr, **qnn_conv2d_params)
else:
out = _op.nn.conv2d(in_expr, weight_expr, **params)
And I found that another check in QnnConv2DRel() has failed: "Expected qnn
conv2d type(int8, uint8) for input but was float32".
Is it ok to simply modify code like I mentioned above and is it ok to ignore
the ICHECK()? I don't know are there any other inner constraints for the
mix-precision issue, and I hope somebody could support this feature officially.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
