zeekay
e037e8f7ea
The distribution called itself torus-ml while shipping a package directory named concrete, so the name a user installs and the name they import disagreed — and most of the tree already said torus.ml, which could not resolve against src/concrete. Moved src/concrete to src/torus and pointed the include at it. The dependency was Zama's concrete-python from their index. It is torus-fhe from ours, which is what luxfhe/torus already ships: that frontend declares torus-fhe and ships a torus/fhe module, so every 'from torus.fhe' in this tree was already naming our compiler and only the dependency lagged. Two distributions share one namespace, which is the shape the upstream this forked from used and the reason the split works: torus-fhe gives torus.fhe, torus-ml gives torus.ml. Swept every other spelling — identifiers, enum members, workflow variables, filenames, notebooks. LICENSE and NOTICE keep the upstream attribution a fork owes; the figures carry the old name in JPEG metadata a text sweep cannot reach.
ResNet18 with Fully Homomorphic Encryption
Overview
This project executes the ResNet18 image classification model using Fully Homomorphic Encryption (FHE) with Torus ML. The model is adapted for FHE compatibility and tested on a small subset of imagenet images.
Hybrid FHE Implementation
The project includes a hybrid FHE implementation (resnet_hybrid_fhe.py) that executes specific layers in FHE mode while keeping others in clear text. This approach significantly reduces execution time while maintaining model accuracy. The hybrid implementation:
- Executes all linear and convolutional layers (except the first conv layer) in FHE mode
- Uses 8-bit quantization for inputs with higher bit-width outputs
- Achieves equivalent accuracy to the clear text model
ResNet18
The ResNet18 model is adapted from torchvision the original https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py where the adaptive average pooling layer AdaptiveAvgPool2d (not yet supported by Torus ML) is replaced with a standard AvgPool2d layer as follows:
- self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+ self.avgpool = nn.AvgPool2d(kernel_size=7, stride=1, padding=0)
The rest is left unchanged.
Evaluation data-set
The model is evaluated on images from the ImageNet-1k data-set.
The ImageNetProcessor class in utils_resnet.py preprocesses the ImageNet validation set for model evaluation. It uses a subset of the validation data to ensure efficient processing and evaluation.
Usage
- Install a virtual Python environment and activate it:
python -m venv venv
source venv/bin/activate
- Install Torus ML:
pip install torus-ml
- Install the dependencies:
pip install -r requirements.txt
- Run the script:
python run_resnet18_fhe.py [--run_fhe] [--export_statistics] [--use_gpu] [--run_experiment] [--dataset_cache_dir <path>] [--num_images <number>]
The script run_resnet18_fhe.py accepts several command-line arguments to control its behavior:
-
--run_fhe: runs the actual FHE execution of the ResNet18 model. If not set, the script will run the model without FHE. -
--export_statistics: exports the circuit statistics after running the model. This can be useful for analyzing the performance and characteristics of the FHE execution. -
--use_gpu: utilizes the available GPU for FHE runtime, potentially speeding up the execution. If not set, the script will run on the CPU. -
--run_experiment: runs experiments with differentn_bitsandrounding_threshold_bitsconfigurations. This can help in finding the optimal settings for the model. -
--dataset_cache_dir <path>: specifies the path to the directory where the data-set is cached. If not provided, the data-set will be downloaded and cached in the default location. -
--num_images <number>: specifies the number of images to process in the FHE execution. The default value is 1. Increasing this number will process more images but may take longer to execute.
Example of output when running the script:
python resnet_fhe.py --run_fhe
Accuracy of the ResNet18 model on the images: 67.00%
Top-5 Accuracy of the ResNet18 model on the images: 87.00%
Compiling the model with compile_torch_model...
Model compiled successfully.
Quantized Model Accuracy of the FHEResNet18 on the images: 67.00%
Quantized Model Top-5 Accuracy of the FHEResNet18 on the images: 87.00%
FHE simulation Accuracy of the FHEResNet18 on the images: 66.00%
FHE simulation Top-5 Accuracy of the FHEResNet18 on the images: 87.00%
Processing 1 image(s)...
Image 1:
Running FHE execution...
FHE execution completed in 811.8710 seconds
FHE top 5 labels: 636, 588, 502, 774, 459
Running simulation...
Simulation top 5 labels: 636, 588, 502, 774, 459
Timings and Accuracy in FHE
CPU machine: 196 cores CPU machine (hp7c from AWS) GPU machine: 8xH100 GPU machine
Summary of the ImageNet results:
- Accuracy is evaluated on 100 images
- Runtime reported for one image
| w&a bits | p_error | Accuracy | Top-5 Accuracy | Runtime | Device |
|---|---|---|---|---|---|
| fp32 | - | 67% | 87% | - | - |
| 6/6 | 0.05 | 55% | 78% | 56 min | GPU |
| 6/6 | 0.05 | 55% | 78% | 1 h 31 min | CPU |
| 7/7 | 0.05 | 66% | 87% | 2 h 12 min | CPU |
| 8/hybrid | - | 67% | 87% | 12 min | CPU |
6/6 n_bits configuration: {"model_inputs": 8, "op_inputs": 6, "op_weights": 6, "model_outputs": 9}
7/7 n_bits configuration: {"model_inputs": 8, "op_inputs": 7, "op_weights": 7, "model_outputs": 9}
For each setting, we use a the following config for the rounding_threshold_bits: {"n_bits": 7, "method": "APPROXIMATE"}.