The OpenMMLab team released a new generation of training engine MMEngine at the World Artificial Intelligence Conference on September 1, 2022. It is a foundational library for training deep learning models. Compared with MMCV, it provides a universal and powerful runner, an open architecture with a more unified interface, and a more customizable training process.
At the same time, MMCV released 2.x release candidate version and will release 2.x official version on January 1, 2023. In version 2.x, it has the following changes:
(1) It removed the following components:
mmcv.fileiomodule, removed in PR #2179. FileIO module from mmengine will be used wherever required.mmcv.runner,mmcv.parallel,mmcv. engineandmmcv.device, removed in PR #2216.- All classes in
mmcv.utils(egConfigandRegistry) and many functions, removed in PR #2217. Only a few functions related to mmcv are reserved. mmcv.onnex,mmcv.tensorrtmodules and related functions, removed in PR #2225.
(2) It added the mmcv.transforms data transformation module.
(3) It renamed the package name mmcv to mmcv-lite and mmcv-full to mmcv in PR #2235. Also, change the default value of the environment variable MMCV_WITH_OPS from 0 to 1.
| MMCV < 2.0 | MMCV >= 2.0 |
|---|---|
# Contains ops, because the highest version of mmcv-full is less than 2.0.0, so there is no need to add version restrictions
pip install mmcv-full -f xxxx
# do not contain ops
pip install "mmcv < 2.0.0" |
# Contains ops
pip install "mmcv>=2.0.0rc1" -f xxxx
# Ops are not included, because the starting version of mmcv-lite is 2.0.0rc1, so there is no need to add version restrictions
pip install mmcv-lite |
Some ops have different implementations on different devices. Lots of macros and type checks are scattered in several files, which makes the code hard to maintain. For example:
if (input.device().is_cuda()) {
#ifdef MMCV_WITH_CUDA
CHECK_CUDA_INPUT(input);
CHECK_CUDA_INPUT(rois);
CHECK_CUDA_INPUT(output);
CHECK_CUDA_INPUT(argmax_y);
CHECK_CUDA_INPUT(argmax_x);
roi_align_forward_cuda(input, rois, output, argmax_y, argmax_x,
aligned_height, aligned_width, spatial_scale,
sampling_ratio, pool_mode, aligned);
#else
AT_ERROR("RoIAlign is not compiled with GPU support");
#endif
} else {
CHECK_CPU_INPUT(input);
CHECK_CPU_INPUT(rois);
CHECK_CPU_INPUT(output);
CHECK_CPU_INPUT(argmax_y);
CHECK_CPU_INPUT(argmax_x);
roi_align_forward_cpu(input, rois, output, argmax_y, argmax_x,
aligned_height, aligned_width, spatial_scale,
sampling_ratio, pool_mode, aligned);
}Registry and dispatcher are added to manage these implementations.
void ROIAlignForwardCUDAKernelLauncher(Tensor input, Tensor rois, Tensor output,
Tensor argmax_y, Tensor argmax_x,
int aligned_height, int aligned_width,
float spatial_scale, int sampling_ratio,
int pool_mode, bool aligned);
void roi_align_forward_cuda(Tensor input, Tensor rois, Tensor output,
Tensor argmax_y, Tensor argmax_x,
int aligned_height, int aligned_width,
float spatial_scale, int sampling_ratio,
int pool_mode, bool aligned) {
ROIAlignForwardCUDAKernelLauncher(
input, rois, output, argmax_y, argmax_x, aligned_height, aligned_width,
spatial_scale, sampling_ratio, pool_mode, aligned);
}
// register cuda implementation
void roi_align_forward_impl(Tensor input, Tensor rois, Tensor output,
Tensor argmax_y, Tensor argmax_x,
int aligned_height, int aligned_width,
float spatial_scale, int sampling_ratio,
int pool_mode, bool aligned);
REGISTER_DEVICE_IMPL(roi_align_forward_impl, CUDA, roi_align_forward_cuda);
// roi_align.cpp
// use the dispatcher to invoke different implementation depending on device type of input tensors.
void roi_align_forward_impl(Tensor input, Tensor rois, Tensor output,
Tensor argmax_y, Tensor argmax_x,
int aligned_height, int aligned_width,
float spatial_scale, int sampling_ratio,
int pool_mode, bool aligned) {
DISPATCH_DEVICE_IMPL(roi_align_forward_impl, input, rois, output, argmax_y,
argmax_x, aligned_height, aligned_width, spatial_scale,
sampling_ratio, pool_mode, aligned);
}
In order to flexibly support more backends and hardwares like NVIDIA GPUs and AMD GPUs, the directory of mmcv/ops/csrc is refactored. Note that this refactoring will not affect the usage in API. For related information, please refer to PR1206.
The original directory was organized as follows.
.
├── common_cuda_helper.hpp
├── ops_cuda_kernel.cuh
├── pytorch_cpp_helper.hpp
├── pytorch_cuda_helper.hpp
├── parrots_cpp_helper.hpp
├── parrots_cuda_helper.hpp
├── parrots_cudawarpfunction.cuh
├── onnxruntime
│ ├── onnxruntime_register.h
│ ├── onnxruntime_session_options_config_keys.h
│ ├── ort_mmcv_utils.h
│ ├── ...
│ ├── onnx_ops.h
│ └── cpu
│ ├── onnxruntime_register.cpp
│ ├── ...
│ └── onnx_ops_impl.cpp
├── parrots
│ ├── ...
│ ├── ops.cpp
│ ├── ops_cuda.cu
│ ├── ops_parrots.cpp
│ └── ops_pytorch.h
├── pytorch
│ ├── ...
│ ├── ops.cpp
│ ├── ops_cuda.cu
│ ├── pybind.cpp
└── tensorrt
├── trt_cuda_helper.cuh
├── trt_plugin_helper.hpp
├── trt_plugin.hpp
├── trt_serialize.hpp
├── ...
├── trt_ops.hpp
└── plugins
├── trt_cuda_helper.cu
├── trt_plugin.cpp
├── ...
├── trt_ops.cpp
└── trt_ops_kernel.cu
After refactored, it is organized as follows.
.
├── common
│ ├── box_iou_rotated_utils.hpp
│ ├── parrots_cpp_helper.hpp
│ ├── parrots_cuda_helper.hpp
│ ├── pytorch_cpp_helper.hpp
│ ├── pytorch_cuda_helper.hpp
│ └── cuda
│ ├── common_cuda_helper.hpp
│ ├── parrots_cudawarpfunction.cuh
│ ├── ...
│ └── ops_cuda_kernel.cuh
├── onnxruntime
│ ├── onnxruntime_register.h
│ ├── onnxruntime_session_options_config_keys.h
│ ├── ort_mmcv_utils.h
│ ├── ...
│ ├── onnx_ops.h
│ └── cpu
│ ├── onnxruntime_register.cpp
│ ├── ...
│ └── onnx_ops_impl.cpp
├── parrots
│ ├── ...
│ ├── ops.cpp
│ ├── ops_parrots.cpp
│ └── ops_pytorch.h
├── pytorch
│ ├── info.cpp
│ ├── pybind.cpp
│ ├── ...
│ ├── ops.cpp
│ └── cuda
│ ├── ...
│ └── ops_cuda.cu
└── tensorrt
├── trt_cuda_helper.cuh
├── trt_plugin_helper.hpp
├── trt_plugin.hpp
├── trt_serialize.hpp
├── ...
├── trt_ops.hpp
└── plugins
├── trt_cuda_helper.cu
├── trt_plugin.cpp
├── ...
├── trt_ops.cpp
└── trt_ops_kernel.cu