BRECQ Data Extraction for DSC Accelerator
Budget: ₹1,500 – ₹12,500 INR
Project Brief: Extraction of Quantization Data from BRECQ for DSC Accelerator
I’m developing a Depthwise Separable Convolution (DSC) accelerator to run quantized MobileNetV2 models. For quantization, I’m using the BRECQ framework (GitHub link will be provided). I need a freelancer to analyze the code, extract specific data required for my hardware implementation, and document the extracted results.
Note: This is primarily a code analysis and data extraction task,and involves minor code modifications to support dataset compatibility or CPU-only execution.
Scope of Work
1. Extract the following from the BRECQ codebase: Quantized weights after post-training quantization.Intermediate output activations (especially quantized outputs at each layer).Quantization parameters per layer: scale factors (delta) and zero-point offsets.
2. Modify the code to: Run on a subset of ImageNet (e.g., Mini-ImageNet). If not natively supported, implement minimal changes to enable this.
Enable export of all extracted data in a structured format (JSON, CSV, or binary).
Ensure the pipeline can run in a CPU-only environment, removing or bypassing CUDA dependencies if needed.
Deliverables
• A well-structured package containing:
The extracted quantized weights, intermediate activations, and quantization parameters in a usable format (JSON, CSV, or binary).
A brief summary/report detailing:
Where the data was extracted from in the code. Any modifications made to enable extraction, dataset substitution, or CPU-only support.Scripts or tools used/generated for the extraction process.
Note: I will provide access to the BRECQ GitHub repository.
I’m developing a Depthwise Separable Convolution (DSC) accelerator to run quantized MobileNetV2 models. For quantization, I’m using the BRECQ framework (GitHub link will be provided). I need a freelancer to analyze the code, extract specific data required for my hardware implementation, and document the extracted results.
Note: This is primarily a code analysis and data extraction task,and involves minor code modifications to support dataset compatibility or CPU-only execution.
Scope of Work
1. Extract the following from the BRECQ codebase: Quantized weights after post-training quantization.Intermediate output activations (especially quantized outputs at each layer).Quantization parameters per layer: scale factors (delta) and zero-point offsets.
2. Modify the code to: Run on a subset of ImageNet (e.g., Mini-ImageNet). If not natively supported, implement minimal changes to enable this.
Enable export of all extracted data in a structured format (JSON, CSV, or binary).
Ensure the pipeline can run in a CPU-only environment, removing or bypassing CUDA dependencies if needed.
Deliverables
• A well-structured package containing:
The extracted quantized weights, intermediate activations, and quantization parameters in a usable format (JSON, CSV, or binary).
A brief summary/report detailing:
Where the data was extracted from in the code. Any modifications made to enable extraction, dataset substitution, or CPU-only support.Scripts or tools used/generated for the extraction process.
Note: I will provide access to the BRECQ GitHub repository.