[Minor] PTQ should clean input tensor

After PTQ, input tensor are in Float32 because of calibration.

This is annoying as this means we always need to remove tensor or convert them to int32/int8 manually.

PTQ method should clean and remove input tensor afterwards. This ways a simple graphView setDatatype will be enough to use the graph.

changed milestone to %aidge v0.6.0

added Enhancement ⭐ label

assigned to @bhalimi

assigned to @noamzerah and unassigned @bhalimi

added Stuck ❌ label

Using the method resetInput() does not seem to resolve the issue:

    for (std::shared_ptr<Aidge::Node> input_node: graphView->inputNodes())
    {
        for (Aidge::IOIndex_t index = input_node->getFirstFreeDataInput();index < input_node->getNbFreeDataInputs(); index++)
        {
            std::static_pointer_cast<OperatorTensor>(input_node->getOperator())->resetInput(index);
        }
    }

What error do you get?

Resetting inputs should solve data type issue.

I encountered this error:

[ERROR] - Assertion failed: exists(key) in /data1/is156025/nz280189/aidge_install/include/aidge/utils/Registrar.hpp:79
[FATAL] - missing or invalid registrar key: [Float64, Default, [(1, 1), (3, 3), (224, 224), (224, 224)]], [Int32, Default, [(1, 1), (3, 3), (230, 230), (230, 230)]] for registrable object
[FATAL]   N5Aidge16OperatorImpl_cpuINS_6Pad_OpILh2EEEFvRKSt5arrayImLm4EENS_13PadBorderTypeEdS6_PKvPv

It essentially points to a registrar error, which means it expects an Float64 input tensor I think.

Can you share your script?

It looks like after quantization you set an input to float64 ...

This issue is still unresolved. Cleaning the input tensor does not seem to work for @axelfarr either. However, adding a casting node before all model inputs successfully prevents the error.

Have you tried to debug this to check when an input Tensor is set again?

When I debug, the input tensor is indeed in Float64. However, calling resetInput() does not seem to delete the input:

std::shared_ptr<Aidge::OperatorTensor> opTensor =std::static_pointer_cast<OperatorTensor>(input_node->getOperator());
        Aidge::DataType dt = opTensor->getInput(0)->dataType();
        Log::debug("node type is {} and name is {} and input 0 datatype is {}",input_node->type(),input_node->name(),dt);
        dt->resetInput(0);
        Log::debug("node type is {} and name is {} and input 0 datatype is {}",input_node->type(),input_node->name(),dt);

This give me the following output:

[DEBUG] - node type is PaddedConv2D and name is resnetv15_conv0_fwd_1 and input 0 datatype is
[DEBUG]   Float64
[DEBUG] - node type is PaddedConv2D and name is resnetv15_conv0_fwd_1 and input 0 datatype is
[DEBUG]   Float64

Maybe the issue only occurs when the first node of the graph is a metaOps. I can quantize a LeNet (first node is Conv2D) without triggering this issue, however running a Resnet18 (fist node is PaddedConv2D) will trigger the issue on the Pad operator...

As @axelfarr suggested, this issue is caused by resetInput(), which does not seem to work on metaoperator.

Let's fix this then! :D

mentioned in issue aidge_core#256 (closed)

@noamzerah Can you confirm this issue is closed ?

I think this issue is closed, but maybe we should ask @axelfarr?

It works for me !

closed