mirror of
https://github.com/open-mmlab/mmdeploy.git
synced 2025-01-14 08:09:43 +08:00
* refactor SDK registry * fix lint * decouple transform logic and operations * data management * improve data management * improve data management * context management * fix ResizeOCR * fix operation fallback logic * fix MSVC build * clean-up * sync master * fix lint * Normalize - add `to_float`, merge `cvtcolor` operations * fix macOS build * rename * cleanup * fix lint * fix macOS build * fix MSVC build * support elena * fix * fix * optimize normalize * fix * fix MSVC build * simplify * profiler * use `throw_exception` * misc * fix typo
Demo
We provide a demo showing what our mmdeploy can do for general model deployment.
In demo_rewrite.py
, a resnet18 model from torchvision
is rewritten through mmdeploy tool. In our rewritten model, the forward function of resnet gets modified to only down sample the original input to 4x. Original onnx model of resnet18 and its rewritten are visualized through netron.
Prerequisite
Before we run demp_rewrite.py
, we need to install pyppeteer
through:
pip install pyppeteer
Demo results
The original resnet18 model and its modified one are visualized as follows. The left model is the original resnet18 while the right model is exported after rewritten.
Original resnet18 | Rewritten model |
---|---|
![]() |
![]() |