Feature request

An implementation of the torch.nn.Unfold in a quantized way.

Motivation

The unfold operation requested is not done here with the torch._C.im2col but with the fhe_conv function to have lower compilation times with concrete-python afterwards.
Doing an unfold with for loops / indexing is very slow compilation wise as it generates a huge MLIR compared to a simple convolution with kernels of a single 1 and zeros

NOTE: This is not actually a bug.

Summary

What happened/what you expected to happen?

Not abe to install concrete-ml with pip. Requires Python >=3.8, <3.10 from Pypi concrete-ml 0.2.1.

EDIT: Sorry I printed the pip list from different python (3.9) version.

I do not have concrete-numpy installed.

Description

• versions affected: 0.1.0; 0.1.1; 0.2.0; 0.2.1.
• python version: Python 3.10.4
• pip version: 22.2.2
• config (optional: HW, OS): Ubuntu 20.04.1 LTS
• workaround (optional): if you’ve a way to workaround the issue
• proposed fix (optional): if you’ve a way to fix the issue

Step by step procedure someone should follow to trigger the bug:

$ pip install concrete-ml
ERROR: Ignored the following versions that require a different python version: 0.1.0 Requires-Python >=3.8,<3.10; 0.1.1 Requires-Python >=3.8,<3.10; 0.2.0 Requires-Python >=3.8,<3.10; 0.2.1 Requires-Python >=3.8,<3.10
ERROR: Could not find a version that satisfies the requirement concrete-ml (from versions: none)
ERROR: No matching distribution found for concrete-ml

Summary

Impossible to install latest version with pip

Description

• versions affected:
• python version: 3.8 3.9 3.10
• config (optional: HW, OS): Mac OS M1 / Mac OS intel

Step by step procedure someone should follow to trigger the bug:

Summary

The version provided through the docker image is not working on Apple M1 and M2 chips

Description

# Install
docker pull zamafhe/concrete-ml:latest

# Run the with:
docker run --rm -it -p 8888:8888 -v ~/host_folder:/data zamafhe/concrete-ml

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from concrete.ml.sklearn import LogisticRegression


# Lets create a synthetic data-set
x, y = make_classification(n_samples=100, class_sep=2, n_features=30, random_state=42)

# Split the data-set into a train and test set
X_train, X_test, y_train, y_test = train_test_split(
    x, y, test_size=0.2, random_state=42
)
# Now we train in the clear and quantize the weights
model = LogisticRegression(n_bits=8)
model.fit(X_train, y_train)

# We then compile on a representative set
model.compile(X_train) # --> This line kills the kernel and restart the notebook

Feature request

Disclaimer: I am very new to concrete-ml and FHE in general.

I am referring to the existing example for encrypted training. From my understanding the example works entirely on plaintext data. Or at least I would expect that the data is encrypted before it is fed into model.fit. In case I misunderstand the example I would be happy if you can clarify it for me. In case my understanding is correct, could you enhance the example to showcase the full workflow.

Motivation

I am very new to concrete ml and it is hard to understand when data is encrypted in the provided example for Encrypted training.

1.the bit in Brevitas model

2.the bit after fhe_compatibility

What is the difference between two bit？

Goal: install concrete-ml

Context
MacOS 10.14
Intel
python --version 3.8.2

Steps:
inside a virtual environment handled by conda

Just run pip install concrete-ml

Bug:

ERROR: Cannot install concrete-ml==1.0.0, concrete-ml==1.0.1, concrete-ml==1.0.2 and concrete-ml==1.0.3 because these package versions have conflicting dependencies.

The conflict is caused by:
    concrete-ml 1.0.3 depends on concrete-python==1.0.0
    concrete-ml 1.0.2 depends on concrete-python==1.0.0
    concrete-ml 1.0.1 depends on concrete-python==1.0.0
    concrete-ml 1.0.0 depends on concrete-python==1.0.0

To fix this you could try to:
1. loosen the range of package versions you've specified
2. remove package versions to allow pip attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/topics/dependency-resolution/#dealing-with-dependency-conflicts

Then, if you try to pip install concrete-python or with pip install concrete-python==1.0.0

ERROR: Could not find a version that satisfies the requirement concrete-python (from versions: none)
ERROR: No matching distribution found for concrete-python

ERROR: Could not find a version that satisfies the requirement concrete-python==1.0.0 (from versions: none)
ERROR: No matching distribution found for concrete-python==1.0.0

Thus, if you try another logic: pip install concrete-ml==1.0.3

Collecting concrete-ml==1.0.3
  Using cached concrete_ml-1.0.3-py3-none-any.whl (178 kB)
Collecting boto3<2.0.0,>=1.23.5 (from concrete-ml==1.0.3)
  Using cached boto3-1.26.145-py3-none-any.whl (135 kB)
Requirement already satisfied: brevitas==0.8.0 in /Users/mike/anaconda3/envs/fhe_chess/lib/python3.8/site-packages (from concrete-ml==1.0.3) (0.8.0)
INFO: pip is looking at multiple versions of concrete-ml to determine which version is compatible with other requirements. This could take a while.
ERROR: Could not find a version that satisfies the requirement concrete-python==1.0.0 (from concrete-ml) (from versions: none)
ERROR: No matching distribution found for concrete-python==1.0.0

Summary

I have created a image classification model in PyTorch and trained on unencrypted data. After that the model is exported using ONNX and import back and checked.
Now when the model is compiled with "compile_onnx_model", it gives following error.

Description

• versions affected:
• python version: 3.9.13
• config (optional: HW, OS):
• workaround (optional): if you’ve a way to workaround the issue
• proposed fix (optional): if you’ve a way to fix the issue

Step by step procedure someone should follow to trigger the bug:

print("Minimal POC to reproduce the bug")

concrete v:4.0.0 conflicts with python3.9

ERROR: Ignored the following versions that require a different python version: 4.0.0 Requires-Python >=3.6,<3.9; 4.0.1 Requires-Python >=3.6,<3.9; 4.1.0 Requires-Python >=3.6,<3.9
ERROR: Could not find a version that satisfies the requirement concrete-compiler<0.20.0,>=0.19.0 (from concrete-numpy) (from versions: 0.1.1, 0.1.2, 0.2.0, 0.3.0, 0.3.1, 0.4.0, 0.5.0, 0.6.0, 0.7.0, 0.8.0, 0.9.0, 0.10.0, 0.11.0, 0.12.0, 0.13.0, 0.14.0, 0.15.0, 0.16.0)
ERROR: No matching distribution found for concrete-compiler<0.20.0,>=0.19.0

Thanks for the awesome work! I was wondering if there is any way to run the library on GPUs?

Hello， can FHE inference run on cuda？

When I run the example of cifar_brevitas_finetiuning, cifar-100 can run normally, but once cifar-10 runs to the fhe_compatibility function, the program is killed, and my memory is 128G

Here is my memory usage for running the program for the dataset using cifar-10

Hello,
I have installed concrete-ml using pip and followed the steps to install pygraphiz (pygraphiz is working).

_ERROR: Could not find a version that satisfies the requirement concrete-compiler<0.5.0,>=0.4.0 (from concrete-numpy[full]<0.4.0,>=0.3.0->concrete-ml) (from versions: none)
ERROR: No matching distribution found for concrete-compiler<0.5.0,>=0.4.0 (from concrete-numpy[full]<0.4.0,>=0.3.0->concrete-ml)
_

I am using :
ubuntu 20.04.4
python 3.8.10

Summary

I have created a image classification model in PyTorch and trained on unencrypted data. After that the model is exported using ONNX and import back and checked.
Now when the model is compiled with "compile_onnx_model", it gives following error.

valueError: The following ONNX operators are required to convert the pytorch model to numpy but are not currently implemented: GlobalAveragePool

Description

• versions affected: v1.3.0
• python version: 3.10.6
• config (optional: HW, OS): Linux

Summary

There is a crash issue (in docker) which is being currently investigated in one of our main dependancies: Concrete-Numpy. This issue happens on Apple Silicon (M1/M2) with docker. We're working on it.

Description

• versions affected: Concrete-Numpy >= 0.9.0, ie Concrete-ML >= 0.6.0
• config (optional: HW, OS): macOS, M1/M2 chips
• workaround (optional): use an older version for now (ie, Concrete-ML 0.5.1)

Summary

What happened/what you expected to happen?
I wanted to check if the model gets compiled on my computer via concrete-ml library, but I faced following issue.

RuntimeError: Can't emit artifacts: Command failed:ld --shared -o /tmp/tmpgrs3bzdc/sharedlib.so /tmp/tmpgrs3bzdc.module-0.mlir.o /home/mahmoud/miniconda3/envs/concrete/lib/python3.10/site-packages/concrete_python.libs/libConcretelangRuntime-636b2a26.so -rpath=/home/mahmoud/miniconda3/envs/concrete/lib/python3.10/site-packages/concrete_python.libs --disable-new-dtags 2>&1
Code:32512
sh: 1: ld: not found

Description

• versions affected: 1.3
• python version: 3.10.13
• config (optional: HW, OS): Intel(R) Core(TM) i5-4200U CPU @ 1.60GHz, Ubuntu 22.04
• proposed fix (optional): I assume it maybe because of my CPU model, it is 4th generation of Intel CPUs.

Step by step procedure someone should follow to trigger the bug:

from sklearn.model_selection import train_test_split
from sklearn.datasets import make_regression
from concrete.ml.sklearn import LinearRegression

X, y = make_regression(
    n_samples=200, n_features=8, n_targets=1, bias=5.0, noise=30.0, random_state=42
)
x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.4, random_state=42)

concrete_model = LinearRegression()
concrete_model.fit(x_train, y_train)
circuit = concrete_model.compile(x_train) # I get the error for this line of code
circuit.keygen()
y_pred = concrete_model.predict(x_test, execute_in_fhe=True)

Summary

When I attempt to run the LLM case, I the prosed error is ImportError: cannot import name 'enc_split' from 'utils' in LLM case.
The error appears in file quant_framework.py
and sequence that from utils import enc_split, max_fhe_relu, simple_slice

In detail, the error is :

ImportError Traceback (most recent call last)
Cell In[1], line 5
3 import numpy as np
4 from concrete.fhe.tracing import Tracer
----> 5 from quant_framework import DualArray, Quantizer
7 from concrete import fhe

File /test_for_docker_jupyter/quant_framework.py:7
5 import numpy as np
6 from concrete.fhe.tracing import Tracer
----> 7 from utils import enc_split, max_fhe_relu, simple_slice
9 EPSILON = 2**-11
11 import numpy as np

ImportError: cannot import name 'enc_split' from 'utils' (/opt/conda/envs/py39/lib/python3.9/site-packages/utils/init.py)

Description

• versions affected: concrete-ml 1.1
• python version: 3.9
• config (optional: HW, OS): Docker, Linux container

I am trying to run the logistic regression example provided in "A simple Concrete ML example with scikit-learn" on the concrete ml GitHub page.

However, while running the code, I am getting the following error:

predict() got an unexpected keyword argument 'fhe'

while executing the following statement:

y_pred_fhe = model.predict(X_test, fhe="execute")

I don't know why this is happening. I am running the code on Ubuntu 20.04 LTS. I have installed concrete ml through the "pip" method in Python 3.8.10

I am also receiving the following warning:

PkgResourcesDeprecationWarning: 0.23ubuntu1 is an invalid version and will not be supported in a future release
warnings.warn(

Hello.

Thank you for your great work!
I have a project to apply FHE on gene expression data for instance two of them

1. Yeast expression matrix, based on Tavazoie et al. 8224 rows and 17 columns, 4 bytes for each element, with -1 indicating a missing value.
2. Human expression matrix, based on Alizadeh et al. with 4026 rows and 96 columns, 4 bytes for each element, and 999 indicating a missing value.

and analyze encrypted input data with Biclustering algorithms like
Cheng and Church Algorithm .

**Biclustering can be performed with the module sklearn.cluster.bicluster ; more information can be found here.
**There are also wide range of Python libraries of biclustering algorithms like biclustlib.

I am very excited to contribute with Zama community to make this unsupervised machine learning algorithm run on concrete-ml and make gene expression data encrypted during process. I would be happy to provide more detailed information if needed.

Many thanks
Shokofeh

Feature request

I would like to request the support of the torch.nn.Unfold function, or a numpy equivalent.

Motivation

I am trying to do some convolutions that do different operations than regular convolutions on each kernels, therefore I would need to use a way of reorganising the images into each kernel.

I am doing it with numpy but it is quite inefficient:

def conv2D_unfold_deep(image, ratio, ksize=(1, 1), pad=0, stride=1):
    """Return the indices used in the 2D convolution"""

    nfilters, height, width = image.shape
    # arr = np.array(list(range(height * width))).reshape((image.shape))

    kernel_height, kernel_width = ksize

    output_height = int((height + 2 * pad - kernel_height) / stride) + 1
    output_width = int((width + 2 * pad - kernel_width) / stride) + 1

    out = fhe.zeros((nfilters//ratio, output_height, output_width, kernel_height * kernel_width*ratio))

    for f in range(0,nfilters,ratio):
        for h in range(output_height):
            for w in range(output_width):
                h_start = h * stride
                h_end = h_start + kernel_height
                w_start = w * stride
                w_end = w_start + kernel_width
                # Get the receptive_field
                # pad image # shape (B, 8, 10, 10)
                receptive_field = image[f:f+ratio, h_start:h_end, w_start:w_end].reshape(
                    (kernel_height * kernel_width * ratio))  # shape(B, 6) binaire
                # transform input 0/1 into int between [0 ; 2**n-1]
                out[f//ratio, h, w, :] = receptive_field  # output_var_unfold

    return out.astype(int)

which is very inefficient because of the three for loops.
I could do it with indexing but it is not supported by concrete:

import numpy as np
from concrete import fhe

coordinates = np.arange(0,100).reshape((10,10))
indexes = conv2D_unfold_deep(coordinates, ratio=1)
x = np.random.randint(0,2,(100))

@fhe.compiler({"x":"encrypted"})
def f(x):
	x = x[indexes]
	return x

f(x)

returns :

ValueError: Tracer<output=EncryptedTensor<uint1, shape=(100,)>> cannot be indexed with [[[[ 0]    [ 1]    [ 2]    [ 3]    [ 4]    [ 5]    [ 6]    [ 7]    [ 8]    [ 9]]    [[10]    [11]    [12]    [13]    [14]    [15]    [16]    [17]    [18]    [19]]    [[20]    [21]    [22]    [23]    [24]    [25]    [26]    [27]    [28]    [29]]    [[30]    [31]    [32]    [33]    [34]    [35]    [36]    [37]    [38]    [39]]    [[40]    [41]    [42]    [43]    [44]    [45]    [46]    [47]    [48]    [49]]    [[50]    [51]    [52]    [53]    [54]    [55]    [56]    [57]    [58]    [59]]    [[60]    [61]    [62]    [63]    [64]    [65]    [66]    [67]    [68]    [69]]    [[70]    [71]    [72]    [73]    [74]    [75]    [76]    [77]    [78]    [79]]    [[80]    [81]    [82]    [83]    [84]    [85]    [86]    [87]    [88]    [89]]    [[90]    [91]    [92]    [93]    [94]    [95]    [96]    [97]    [98]    [99]]]]

Summary

I ran the following code in jupyter notebook related to Linear regression tutorial with concrete-ml and kernel crashes each time I run this code.
fhe_circuit.client.keygen(force=False)

Description

• versions affected:0.6.1
• python version:3.9.16
• OS:Linux mint una

Step by step procedure someone should follow to trigger the bug:

just run all the cells in jupyter notebook of linear regression with concrete-ml.

Summary

When develop a custom neural network and try to compile it using MNIST Dataset through compile_torch_model, I get an assertion error :
"AssertionError: Got the following unknown attributes: count_include_pad. Accepted attributes: ceil_mode, kernel_shape, pads, strides"

I was expecting compilation and generation of the circuit which I will then use to train my model.

• versions affected:
• python version: 3.10.12
• config : x86_64
• workaround (optional): remove the pooling layers but you are likely to run into other issues and above all, affect the accuracy of the network.

Step by step procedure someone should follow to trigger the bug:

I am unable to resolve following error while installing:
ERROR: Could not find a version that satisfies the requirement concrete-compiler<0.5.0,>=0.4.0 (from concrete-numpy[full]) (from versions: none)
ERROR: No matching distribution found for concrete-compiler<0.5.0,>=0.4.0

=============
I followed following 2 steps:

1. docker pull zamafhe/concrete-ml:latest
2. pip install concrete-ml>s1.txt

s1.txt has been attached
s1.txt

qmodel = fhe_compatibility(quant_vgg, data)

print(
f"With {param_c10['dataset_name']}, the circuit has a maximum bit-width of "
f"{qmodel.fhe_circuit.graph.maximum_integer_bit_width()}-bits."
)

Feature request

[edit by @bcm-at-zama]: info are given by @Daihongwei1900 here

Motivation

[edit by @bcm-at-zama]: info are given by @Daihongwei1900 here

I ran this codeopen("cifar10.graph", "w").write(str(quantized_numpy_module.fhe_circuit)) . Executed this code sentence. And I have obtained the file cifar10.graph，how to visualize this file.

I am trying to run the VGG like NN on CIFAR-10 example found at https://github.com/zama-ai/concrete-ml/tree/main/use_case_examples/cifar/cifar_brevitas_training.

When I am running the example once, I am getting correct results in terms of accuracy (during validation) being similar in clear and FHE predictions. However, when I am running the same code multiple times with same parameters, I am getting a huge difference in the accuracy of clear and FHE predictions. For example, in clear prediction I am getting accuracy of around 97%, which remains similar across the multiple iterations. On the other hand, the accuracy in FHE prediction drops to around 40% to 50% and varies in this range across multiple iterations. I am executing the scripts as obtained from the repository, while only changing the number of training epochs to 10 from default 1000.

I am unable to understand why this is happening. While it is understandable that multiple training iterations will result in models with varied accuracy, such significant drop in the accuracy, that too only in FHE prediction, does not make sense to me. Please let me know if any other code base for CIFAR-10 is available or not.

When I compile a model with BatchNorm2d layers, the following error occurs

The following tensors were expected to be quantized, but the values found during calibration do not appear to be quantized.

I am trying to run the fully connected neural network on MNIST example found at https://github.com/zama-ai/concrete-ml/blob/release/1.1.x/docs/advanced_examples/FullyConnectedNeuralNetworkOnMNIST.ipynb

I am not using the Jupyter Notebook but running the code locally by writing it in a .py file. The code is working fine and I am getting the expected result.

However, when I am trying to dump the model, either to a string using model.dumps() or a file using model.dump(f), I am getting the following error:

File "/home/user/.local/lib/python3.8/site-packages/concrete/ml/sklearn/qnn.py", line 566, in dump_dict
raise NotImplementedError(
NotImplementedError: Serializing a custom Callable object is not secure and is therefore disabled. Additionally, the serialization of skorch's different callback classes is not supported. Please set callbacks to 'disable'. Got None.

I also used these methods in the Logistic Regression example, found on the readme page, and they are working correctly as I am able to dump the model both into a string and a file.

As per the Serialization page, found at https://docs.zama.ai/concrete-ml/advanced-topics/serialization, this method should work for all built-in models so this should not be a problem caused by the selected model.

I am unable to understand why this is happening.

Hello，What is the difference between a graph and a circuit？

Hi team, I've noticed that ONNX operators have been slowly added over previous releases this past year (kudos!).

There are two operators missing when attempting to convert a pretrained torch ResNet18 model via compile_torch_model:

1. Expand
2. ReduceL2

Are these two operators missing due to limitations on the FHE side or because they are not commonly used?
Thanks!

Summary

I have implemented a GNN model, quantised using Brevitas, and compiled using 'compile_brevitas_qat_model'.
During the compilation, I face the following error: "ValueError: The following ONNX operators are required to convert the torch model to numpy but are not currently implemented: ScatterElements."

Would it be possible to add ScatterElements as a supported operator, or is there any suggested workarounds? Thank you!

I'm following the the decision tree tutorial https://docs.zama.ai/concrete-ml/stable/user/advanced_examples/DecisionTreeClassifier.html but I'm changing the dataset to another one, e.g., https://www.openml.org/search?type=data&status=active&tags.tag=uci&qualities.NumberOfClasses=%3D_2&id=1480

But I'm getting the following error

ValueError: The following ONNX operators are required to convert the torch model to numpy but are not currently implemented: Gather.
Available ONNX operators: Abs, Acos, Acosh, Add, Asin, Asinh, Atan, Atanh, Celu, Clip, Constant, Conv, Cos, Cosh, Div, Elu, Equal, Erf, Exp, Gemm, Greater, HardSigmoid, Identity, LeakyRelu, Less, Log, MatMul, Mul, Not, Relu, Reshape, Selu, Sigmoid, Sin, Sinh, Softplus, Sub, Tan, Tanh, ThresholdedRelu

when running .fit or .fit_benchmark on the model.

I imagine this operator https://github.com/onnx/onnx/blob/main/docs/Changelog.md#gather-1 is not yet implemented.

Thank you very much for the project. I would like to ask how i can get the encrypted test set after performing the inference in FHE.

Summary

Cannot compile concrete-ml AI models for example code

Description

• concrete-ml: 1.4.0
• python version: 3.8
• OS: Ubuntu 20.04 (virtual environment)
  or
• concrete-ml: 1.4.0
• python version: 3.10.12
• OS: Ubuntu 22.04 (virtual environment)

On either environment above, step by step procedure someone should follow to trigger the bug:

pip3 install concrete-ml

import concrete.ml
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from concrete.ml.sklearn import LogisticRegression
x, y = make_classification(n_samples=100, class_sep=2, n_features=30, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=42)
model = LogisticRegression(n_bits=8)
model.fit(X_train, y_train)
model.compile(X_train)

 #0 0x00007f13fccbb761 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x13a5761)
 #1 0x00007f13fccb9174 SignalHandler(int) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x13a3174)
 #2 0x00007f14a6ba8090 (/lib/x86_64-linux-gnu/libc.so.6+0x43090)
 #3 0x00007f1401063128 cxx::unwind::prevent_unwind::h027936808a60dbca (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x574d128)
 #4 0x00007f140105a7ba concrete_optimizer::dag::empty() (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x57447ba)
 #5 0x00007f140037f097 mlir::concretelang::optimizer::DagPass::runOnOperation() (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x4a69097)
 #6 0x00007f13fcbfe342 mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x12e8342)
 #7 0x00007f13fcbfe919 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x12e8919)
 #8 0x00007f13fcbff923 mlir::detail::OpToOpPassAdaptor::runOnOperationImpl(bool) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x12e9923)
 #9 0x00007f13fcbfe036 mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x12e8036)
#10 0x00007f13fcbfe919 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x12e8919)
#11 0x00007f13fcbff421 mlir::PassManager::run(mlir::Operation*) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x12e9421)
#12 0x00007f13fe8b1313 mlir::concretelang::pipeline::getFHEContextFromFHE[abi:cxx11](mlir::MLIRContext&, mlir::ModuleOp&, mlir::concretelang::optimizer::Config, std::function<bool (mlir::Pass*)>) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x2f9b313)
#13 0x00007f13fe88e58b mlir::concretelang::CompilerEngine::getConcreteOptimizerDescription(mlir::concretelang::CompilerEngine::CompilationResult&) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x2f7858b)
#14 0x00007f13fe8935bf mlir::concretelang::CompilerEngine::determineFHEParameters(mlir::concretelang::CompilerEngine::CompilationResult&) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x2f7d5bf)
#15 0x00007f13fe894653 mlir::concretelang::CompilerEngine::compile(mlir::ModuleOp, mlir::concretelang::CompilerEngine::Target, std::optional<std::shared_ptr<mlir::concretelang::CompilerEngine::Library> >) (.localalias) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x2f7e653)
#16 0x00007f13fe898c28 llvm::Expected<mlir::concretelang::CompilerEngine::Library> mlir::concretelang::compileModuleOrSource<mlir::ModuleOp>(mlir::concretelang::CompilerEngine*, mlir::ModuleOp, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, bool, bool, bool) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x2f82c28)
#17 0x00007f13fe8991c4 mlir::concretelang::CompilerEngine::compile(mlir::ModuleOp, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, bool, bool, bool, bool) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x2f831c4)
#18 0x00007f13fca990fd mlir::concretelang::LibrarySupport::compile(mlir::ModuleOp&, std::shared_ptr<mlir::concretelang::CompilationContext>&, mlir::concretelang::CompilationOptions) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x11830fd)
#19 0x00007f13fcaa13f9 library_compile_module(LibrarySupport_Py, mlir::ModuleOp, mlir::concretelang::CompilationOptions, std::shared_ptr<mlir::concretelang::CompilationContext>) (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/libConcretelangBindingsPythonCAPI.so+0x118b3f9)
#20 0x00007f13f91b3d35 (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/_concretelang.cpython-38-x86_64-linux-gnu.so+0x70d35)
#21 0x00007f13f918a1db (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/_concretelang.cpython-38-x86_64-linux-gnu.so+0x471db)
#22 0x00007f13f91689c7 (/home/weimin/.local/lib/python3.8/site-packages/mlir/_mlir_libs/_concretelang.cpython-38-x86_64-linux-gnu.so+0x259c7)
#23 0x00000000005d5499 PyCFunction_Call (/usr/bin/python3.8+0x5d5499)
#24 0x00000000005d6066 _PyObject_MakeTpCall (/usr/bin/python3.8+0x5d6066)
#25 0x00000000004e22b3 (/usr/bin/python3.8+0x4e22b3)
#26 0x000000000054c8a9 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x54c8a9)
#27 0x000000000054552a _PyEval_EvalCodeWithName (/usr/bin/python3.8+0x54552a)
#28 0x00000000005d5a23 _PyFunction_Vectorcall (/usr/bin/python3.8+0x5d5a23)
#29 0x0000000000547447 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x547447)
#30 0x000000000054552a _PyEval_EvalCodeWithName (/usr/bin/python3.8+0x54552a)
#31 0x00000000005d5a23 _PyFunction_Vectorcall (/usr/bin/python3.8+0x5d5a23)
#32 0x00000000005483b6 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x5483b6)
#33 0x00000000005d5846 _PyFunction_Vectorcall (/usr/bin/python3.8+0x5d5846)
#34 0x0000000000547447 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x547447)
#35 0x000000000054552a _PyEval_EvalCodeWithName (/usr/bin/python3.8+0x54552a)
#36 0x00000000005d5a23 _PyFunction_Vectorcall (/usr/bin/python3.8+0x5d5a23)
#37 0x0000000000579c7d (/usr/bin/python3.8+0x579c7d)
#38 0x00000000005d5fcf _PyObject_MakeTpCall (/usr/bin/python3.8+0x5d5fcf)
#39 0x000000000054ca58 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x54ca58)
#40 0x000000000054552a _PyEval_EvalCodeWithName (/usr/bin/python3.8+0x54552a)
#41 0x00000000004e1bd0 (/usr/bin/python3.8+0x4e1bd0)
#42 0x00000000005483b6 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x5483b6)
#43 0x000000000054552a _PyEval_EvalCodeWithName (/usr/bin/python3.8+0x54552a)
#44 0x00000000004e1bd0 (/usr/bin/python3.8+0x4e1bd0)
#45 0x000000000054c8a9 _PyEval_EvalFrameDefault (/usr/bin/python3.8+0x54c8a9)
#46 0x000000000054552a _PyEval_EvalCodeWithName (/usr/bin/python3.8+0x54552a)
#47 0x0000000000684327 PyEval_EvalCode (/usr/bin/python3.8+0x684327)
#48 0x0000000000673a41 (/usr/bin/python3.8+0x673a41)
#49 0x0000000000673abb (/usr/bin/python3.8+0x673abb)
#50 0x0000000000488acc (/usr/bin/python3.8+0x488acc)
#51 0x000000000048947a PyRun_InteractiveLoopFlags (/usr/bin/python3.8+0x48947a)
#52 0x0000000000674a19 PyRun_AnyFileExFlags (/usr/bin/python3.8+0x674a19)
#53 0x00000000004c41f6 (/usr/bin/python3.8+0x4c41f6)
#54 0x00000000006b43fd Py_BytesMain (/usr/bin/python3.8+0x6b43fd)
#55 0x00007f14a6b89083 __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x24083)
#56 0x00000000005da67e _start (/usr/bin/python3.8+0x5da67e)

Is it possible to install concrete-ml through Jupyter Notebook using the pip command on a Windows machine?

Hi,

I am doing research on FHE with neural networks, and I am wondering whether you have tried or plan on doing model training with homomorphically encrypted data. If there are such capacities of the library, could you point me to the tutorial or give me some idea on how would one go about that?

Thank you and great work on the library so far!

Hi,
I'm trying to do Encrypted Inference with my own torch pre-trained model. I am using your client-server example as refence: https://github.com/zama-ai/concrete-ml/blob/release/0.6.x/docs/advanced_examples/ClientServer.ipynb

Description

The steps that I have done are the following: 

1. I have compiled my torch model using compile_torch_model 
2. Now when I use client_send_input_to_server_for_prediction(encrypted_input) it raises an error when it reaches this line inside the function:
   encrypted_prediction = FHEModelServer(self.server_dir.name).run( encrypted_input, serialized_evaluation_keys )
   Error output:
   RuntimeError: Number of dimensions did not match the number of expected dimensions

Could you give me some idea about how can I resolve it?
Thank you so much, this library is helping me a lot!

Hi, I saw a parameter (n_bits=3) when defining a model. Does it mean the security level? Or when I use it to encrypt inference, does it satisfy 128-bits security?

Hey, just wanted to note that https://huggingface.co/spaces/zama-fhe/encrypted_sentiment_analysis is displaying a build error.

Summary

I'm unable to view the documentation of the concrete ml models. I wanted to try and use logistic regression on India pima diabetes dataset.

On trying to read the docs from here , I get the error:

The 'zama-ai/concrete-ml' repository doesn't contain the 'docs/_apidoc/concrete.ml.sklearn.html' path in 'release/0.3.x'.

Summary

I'm unable to compile a model in my MacOS environment.

Description

Steps to reproduce:

• Install using

python -m venv venv
cd venv && source venv/bin/activate
pip install -U pip wheel setuptools
pip install concrete-ml

• Copy/paste the example from here.
• Use python example.py and get this error:

Traceback (most recent call last):
  File "example.py", line 21, in <module>
    model.compile(X_train)
  File "/Users/rmagner/Documents/Experiments/Concrete/venv/lib/python3.8/site-packages/concrete/ml/sklearn/base.py", line 1101, in compile
    circuit = self.quantized_module_.compile(
  File "/Users/rmagner/Documents/Experiments/Concrete/venv/lib/python3.8/site-packages/concrete/ml/quantization/quantized_module.py", line 358, in compile
    self.forward_fhe = compiler.compile(
  File "/Users/rmagner/Documents/Experiments/Concrete/venv/lib/python3.8/site-packages/concrete/numpy/compilation/compiler.py", line 373, in compile
    mlir = GraphConverter.convert(self.graph, virtual=self.configuration.virtual)
  File "/Users/rmagner/Documents/Experiments/Concrete/venv/lib/python3.8/site-packages/concrete/numpy/mlir/graph_converter.py", line 401, in convert
    GraphConverter._update_bit_widths(graph)
  File "/Users/rmagner/Documents/Experiments/Concrete/venv/lib/python3.8/site-packages/concrete/numpy/mlir/graph_converter.py", line 226, in _update_bit_widths
    raise RuntimeError(
RuntimeError: Function you are trying to compile cannot be converted to MLIR:

%0 = [[ -3] [-1 ... -1] [  4]]        # ClearTensor<int8, shape=(30, 1)>
%1 = _input_0                         # EncryptedTensor<uint8, shape=(1, 30)>
%2 = -131                             # ClearScalar<int9>
%3 = add(%1, %2)                      # EncryptedTensor<int9, shape=(1, 30)>
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ only up to 8-bit integers are supported
%4 = subgraph(%3)                     # EncryptedTensor<int8, shape=(1, 30)>
%5 = matmul(%4, %0)                   # EncryptedTensor<int16, shape=(1, 1)>
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ only up to 8-bit integers are supported
%6 = subgraph(%5)                     # EncryptedTensor<uint8, shape=(1, 1)>
return %6

Subgraphs:

    %4 = subgraph(%3):

         %0 = -128                           # ClearScalar<int8>
         %1 = 127                            # ClearScalar<uint7>
         %2 = 3                              # ClearScalar<uint2>
         %3 = 0.02919340787200249            # ClearScalar<float64>
         %4 = 0.02919340787200249            # ClearScalar<float64>
         %5 = input                          # EncryptedTensor<uint2, shape=(1, 30)>
         %6 = multiply(%4, %5)               # EncryptedTensor<float64, shape=(1, 30)>
         %7 = true_divide(%6, %3)            # EncryptedTensor<float64, shape=(1, 30)>
         %8 = add(%7, %2)                    # EncryptedTensor<float64, shape=(1, 30)>
         %9 = rint(%8)                       # EncryptedTensor<float64, shape=(1, 30)>
        %10 = clip(%9, %0, %1)               # EncryptedTensor<float64, shape=(1, 30)>
        %11 = astype(%10, dtype=int_)        # EncryptedTensor<uint1, shape=(1, 30)>
        return %11

    %6 = subgraph(%5):

         %0 = 0                                 # ClearScalar<uint1>
         %1 = 255                               # ClearScalar<uint8>
         %2 = 0                                 # ClearScalar<uint1>
         %3 = 0.003919792729613519              # ClearScalar<float64>
         %4 = 1.0                               # ClearScalar<float64>
         %5 = 1.0                               # ClearScalar<float64>
         %6 = [0.11045581]                      # ClearTensor<float32, shape=(1,)>
         %7 = 0.0003101464749011019             # ClearScalar<float64>
         %8 = [[30]]                            # ClearTensor<uint5, shape=(1, 1)>
         %9 = 0                                 # ClearScalar<uint1>
        %10 = input                             # EncryptedTensor<uint5, shape=(1, 1)>
        %11 = astype(%10, dtype=float32)        # EncryptedTensor<float32, shape=(1, 1)>
        %12 = add(%11, %9)                      # EncryptedTensor<float32, shape=(1, 1)>
        %13 = add(%12, %8)                      # EncryptedTensor<float64, shape=(1, 1)>
        %14 = multiply(%7, %13)                 # EncryptedTensor<float64, shape=(1, 1)>
        %15 = add(%14, %6)                      # EncryptedTensor<float64, shape=(1, 1)>
        %16 = negative(%15)                     # EncryptedTensor<float64, shape=(1, 1)>
        %17 = exp(%16)                          # EncryptedTensor<float64, shape=(1, 1)>
        %18 = add(%5, %17)                      # EncryptedTensor<float64, shape=(1, 1)>
        %19 = true_divide(%4, %18)              # EncryptedTensor<float64, shape=(1, 1)>
        %20 = true_divide(%19, %3)              # EncryptedTensor<float64, shape=(1, 1)>
        %21 = add(%20, %2)                      # EncryptedTensor<float64, shape=(1, 1)>
        %22 = rint(%21)                         # EncryptedTensor<float64, shape=(1, 1)>
        %23 = clip(%22, %0, %1)                 # EncryptedTensor<float64, shape=(1, 1)>
        %24 = astype(%23, dtype=int_)           # EncryptedTensor<uint1, shape=(1, 1)>
        return %24

Note I've also tried running in JupyterLab and get the same error when hitting the model.compile step. Everything else up to that works as expected.

The documentation is a little conflicting. This page says this method should work just fine on MacOS, but this page says glibc is required, which isn't supported on MacOS as far as I'm aware. So is anything further required to get this running on MacOS, or is something else misconfigured?

Here are some environment stats:

$ ~ gcc --version
Apple clang version 14.0.0 (clang-1400.0.29.202)
Target: x86_64-apple-darwin21.6.0
Thread model: posix
InstalledDir: /Library/Developer/CommandLineTools/usr/bin