.png&w=1920&q=75)
This project presents a high-performance, containerized Optical Character Recognition (OCR) pipeline built specifically for AMD GPU and NPU architectures. Developed as part of the AMD AI Challenge, the solution leverages the official ROCm PyTorch base environment to deliver fast and reliable text extraction from complex document images. The entire inference environment is fully self-contained within a public Docker image (lucass202/amd-ocr-challenge:v1), embedding all required Python packages and pre-trained weights without relying on external runtime downloads. Designed with strict memory management in mind, the pipeline utilizes FP16/BF16 optimizations and dynamic allocation to maintain a low VRAM footprint while guaranteeing reproducibility and seamless deployment across AMD workstations and cloud infrastructure.
Category tags: