pip install model-compression-toolkit

MCT supports various quantization methods as appears below.

For each flow, Quantization core utilizes various algorithms and hyper-parameters for optimal hardware-aware quantization results. For further details, please see Supported features and algorithms.

Required input:

• Floating point model - 32bit model in either .pt or .keras format
• Representative dataset - can be either provided by the user, or generated utilizing the Data Generation capability

High level features and techniques

MCT offers a range of powerful features to optimize models for efficient edge deployment. These supported features include:

Quantization Core Features

🏆 Mixed-precision search . Assigning optimal quantization bit-width per layer (for weights/activations)

📈 Graph optimizations. Transforming the model to be best fitted for quantization process.

🔎 Quantization parameter search . Minimizing expected quantization-noise during thresholds search using methods such as MSE, No-Clipping and MAE.

🧮 Advanced quantization algorithms . Enhancing quantization performance for advanced cases is available with some algorithms that can be applied, such as Shift negative correction, Outliers filtering and clustering.

Hardware-aware optimization

🎯 TPC (Target Platform Capabilities). Describes the target hardware’s constrains, for which the model optimization is targeted. See TPC Readme for more information.

Data-free quantization (Data Generation)

Generates synthetic images based on the statistics stored in the model's batch normalization layers, according to your specific needs, for when image data isn’t available. See Data Generation Library for more.

Structured Pruning

Reduces model size/complexity and ensures better channels utilization by removing redundant input channels from layers and reconstruction of layer weights. Read more (Pytorch API / Keras API).

Debugging and Visualization

🎛️ Network Editor (Modify Quantization Configurations) . Modify your model's quantization configuration for specific layers or apply a custom edit rule (e.g adjust layer's bit-width) using MCT’s network editor

🖥️ Visualization. Observe useful information for troubleshooting the quantized model's performance using TensorBoard. Read more.

🔑 XQuant (Explainable Quantization) . Get valuable insights regarding the quality and success of the quantization process of your model. The report includes histograms and similarity metrics between the original float model and the quantized model in key points of the model. The report can be visualized using TensorBoard.

Enhanced Post-Training Quantization (EPTQ)

As part of the GPTQ capability, we provide an advanced optimization algorithm called EPTQ. The specifications of the algorithm are detailed in the paper: "EPTQ: Enhanced Post-Training Quantization via Hessian-guided Network-wise Optimization" [4]. More details on how to use EPTQ via MCT can be found in the GPTQ guidelines.

Resources

• User Guide contains detailed information about MCT and guides you from installation through optimizing models for your edge AI applications.

• MCT's API Docs is seperated per quantization methods:

  • Post-training quantization | PTQ API docs
  • Gradient-based post-training quantization | GPTQ API docs
  • Quantization-aware training | QAT API docs

• Debug – modify optimization process or generate explainable report

• Release notes

Supported Versions

Currently, MCT is being tested on various Python, Pytorch and TensorFlow versions:

Results

### Pytorch We quantized classification networks from the torchvision library. In the following table we present the ImageNet validation results for these models: | Network Name | Float Accuracy | 8Bit Accuracy | Data-Free 8Bit Accuracy | |---------------------------|-----------------|-----------------|-------------------------| | MobileNet V2 [3] | 71.886 | 71.444 |71.29| | ResNet-18 [3] | 69.86 | 69.63 |69.53| | SqueezeNet 1.1 [3] | 58.128 | 57.678 || ### Keras MCT can quantize an existing 32-bit floating-point model to an 8-bit fixed-point (or less) model without compromising accuracy. Below is a graph of [MobileNetV2](https://keras.io/api/applications/mobilenet/) accuracy on ImageNet vs average bit-width of weights (X-axis), using single-precision quantization, mixed-precision quantization, and mixed-precision quantization with GPTQ. For more results, please see [1] ### Pruning Results Results for applying pruning to reduce the parameters of the following models by 50%: | Model | Dense Model Accuracy | Pruned Model Accuracy | |-----------------|----------------------|-----------------------| | ResNet50 [2] | 75.1 | 72.4 | | DenseNet121 [3] | 74.44 | 71.71 | ##