im2im: Automatically Converting In-Memory Image Representations using A Knowledge(…)
summary
im2im is a Python library that builds a knowledge graph of 1:1 metadata nodes for in‑memory image representations (NumPy/OpenCV, Pillow, PyTorch, scikit‑image, GPU tensors) with edges annotated by code templates and measured conversion timings. It uses an A search with a cost tuple (binary information‑loss flag plus a constraint score and CPU/GPU penalty) to synthesize executable conversion code. Two block‑based VPL case studies and a brute‑force benchmark over all type pairs show negligible runtime overhead at realistic resolutions and fewer unnecessary conversions than IR‑based approaches; artifacts are on GitHub/Zenodo.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 2:56IM2IM’s knowledge graph
- 5:46Graph search and code synthesis
- 8:15Manual vs IM2IM comparison
- 9:41IM2IM applications
- 11:50Visual programming with IM2IM
- 12:31Summary and future work
- 13:14Q&A
related talks
28mMem2Img: Memory-Resident Malware Detection via Convolution Neural Network
01h 04mFrom Pixels to Words: Exploring Multi-Modal AI with Python
23mOptimize your network inference time with OpenVINO
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
21mKTransformers: Unleashing the Full Potential of CPU/GPU Hybrid Inference for M...
49mGraph Based Processing of Big Images
22mInterpreting Machine Learning Models with LIME
32mEverything you wanted to know about VLMs but were afraid to ask