WARNING: high error rate, more details with --display-optimizer-choice？ about concrete-ml HOT 16 CLOSED

Hello,
Would you be able to give us more info about your issue ? More precisely, would it be possible to :

• indicate to us which model you are working with
• set verbose=True when compiling your model (depending what you are doing: compile, compile_torch_model, etc.) and share us the printed details

Many thanks !

from concrete-ml.

In Concrete v1, homomorphic encryption wasn't simulated, computation graph was walked and evaluated (and some errors are added before TLUs to kinda simulate p_error). Here are the relevant pieces of code:

• https://github.com/zama-ai/concrete/blob/cb19989fe15970ef6d993e75c3a512f07c0a5408/frontends/concrete-python/concrete/fhe/compilation/circuit.py#L59-L72
• https://github.com/zama-ai/concrete/blob/cb19989fe15970ef6d993e75c3a512f07c0a5408/frontends/concrete-python/concrete/fhe/representation/graph.py#L53-L66
• https://github.com/zama-ai/concrete/blob/cb19989fe15970ef6d993e75c3a512f07c0a5408/frontends/concrete-python/concrete/fhe/representation/graph.py#L68-L188

With Concrete v2, homomorphic encryption is simulated! We've introduced a new backend to do that and new simulation is using the same compiler architecture, so it's much more accurate. Here is the relevant piece of code:

• https://github.com/zama-ai/concrete/blob/ec8cb214fa10adaf9aaf436d85df834d84eb4d14/frontends/concrete-python/concrete/fhe/compilation/circuit.py#L77-L116

Hope it helps!

from concrete-ml.

After updating the new version of concrete ml, this issue occurred. In previous versions, there was no such warning.
CifarInFhe.ipynb

from concrete-ml.

Thanks for reporting ! We are indeed getting the same warning and are thus currently investigating on what could be the issue 🙂

from concrete-ml.

Hello,
Thanks for reporting the warning!
We have fixed the issue on our end, but in the meantime, you can also make a fix on your side to continue.

def fhe_compatibility(model: Callable, data: DataLoader) -> Callable:
    qmodel = compile_brevitas_qat_model(
        model.to("cpu"),
        # Training
        torch_inputset=data,
        show_mlir=False,
        output_onnx_file="test.onnx",
    )
    return qmodel

Thanks !

from concrete-ml.

Hello, thank you for replying to me.
But I still have a problem.
Do you mean that the configuration and p_error parameters are not required？
Can you explain these two parameters to me？

from concrete-ml.

Hi, you can find an explanation of p_error in this documentation page. Here the solution proposed is to remove the custom p_error value that this example was setting.

You mention that you still have a problem, can you please elaborate ?

from concrete-ml.

Thanks，How fhe_simulation_inference function simulate Homomorphic encryption?Where can I see the details of simulation?

from concrete-ml.

To complement @umut-sahin's answer, Concrete-ML's currently works with Concrete v1 but we are planning on integrating v2 in our next release !

from concrete-ml.

So this code does not simulate Homomorphic encryption？It uses Concrete v1 ？cifar_brevitas_finetuning

from concrete-ml.

Yes it does provide some kind of FHE simulation using Concrete v1's technique (graph evaluation + noise simulation), fhe_simulation_inference does not execute the inference in FHE ! The technique is just not as accurate as its second version (documentation: https://docs.zama.ai/concrete/tutorials/simulation), which will come with Concrete ML's next release. Still, you should currently be able to simulate FHE execution with a very good accuracy.

from concrete-ml.

Hello @maxwellgodv , if you are satisfied with the above answer, please feel free to close this issue 🙂

from concrete-ml.

So is it to compile the quantitative model into an circuit, and then input the data into the circuit for prediction ？And, where can I find documents about circuit.
I don't have much knowledge about circuit.Thanks.

from concrete-ml.

There is no dedicated page for circuit, but there are pages for everything you can do with a circuit in https://docs.zama.ai/concrete!

Also, you can read the source code, every method is documented https://github.com/zama-ai/concrete/blob/main/frontends/concrete-python/concrete/fhe/compilation/circuit.py 🙂

And of course you can always ask directly to us, see https://www.zama.ai/community 😉

from concrete-ml.

Hello again @maxwellgodv, is it okay for you if we close this issue as the current discussion does not concern the initial problem anymore ? If you have more questions, we'll gladly answer you on our community forum as well as on the Concrete ML channel that you can find in FHE.org's discord server 🙂

Thanks again for the report !

from concrete-ml.

OK,thanks

from concrete-ml.