_ _
(_) | |
_ __ ___ ___ _ __ _ __ ___ __ _ _ __| |
| '_ ` _ \ / _ \ '__| '_ ` _ \ / _` | |/ _` |
| | | | | | __/ | | | | | | | (_| | | (_| |
|_| |_| |_|\___|_| |_| |_| |_|\__,_|_|\__,_|
Mermaid is a registration toolkit making use of automatic differentiation for rapid prototyping. It runs on the CPU and the GPU, though GPU acceleration only becomes obvious for large images or 3D volumes.
A basic installation requires the installation of a few python packages. Most of these packages can be installed via conda, but a few do not have conda installers, but can be installed via pip.
- conda install pytorch torchvision -c pytorch
- [for CPU only: conda install pytorch-cpu torchvision -c pytorch]
- conda install cffi
- conda install future
- conda install -c conda-forge itk
- conda install progressbar
- conda install sphinx
- conda install pandas
- conda install matplotlib
- conda install scipy
- conda install openpyxl
- conda install -c anaconda cython
- pip install pytorch-fft
- pip install pynrrd
IMPORTANT: pytorch 0.3.0 is required. The code will not run on 0.2.0
The mermaid.yaml file in the install directory also contains information for all the packages that need to be installed. We will provide an anaconda installer in the (hopefully) near future.
*WARNING: This does not work yet and has not been updated for pytorch 0.3.0
It is of course also possible to install everything into a virtual conda environment. For this, do the following:
- conda create -n mermaid-py27 python=2.7 pip
- source activate mermaid-py27
- cd install
- pip install -r requirements_python2_osx.txt [Pick the right one for your operating system]
- pip install -r requirements_pytorch_fft.txt
We are working on making the code compatible with python3, which is currently not supported. Once it is supported it will be possible to install doing the following:
- conda create -n mermaid-py36 python=3.6 pip
- source activate mermaid-py36
- cd install
- pip install -r requirements_python3_osx.txt [Pick the right one for your operating system]
- pip install -r requirements_pytorch_fft.txt
After everything is installed, compile the documentation (there are also more detailed installation instructions) by doing the following.
* cd mermaid
* cd docs
* make html
Lastly, to compile the spatial transformer code simply go to the pyreg/libraries directory
for gpu user: execute 'sh make_cuda.sh'
for cpu user: execute 'sh make_cpu.sh'
CUDA can be enabled by setting CUDA_ON to True in settings/compute_settings.json. All settings are contained in json configuration files in the settings directory.
There are a few example registration implementations in the demos directory. Actual applications are contained in the apps directory.
- Use float32 in most cases, float16 is not stable
- Most parts of the codes have been examined. In case of failure of the GPU code contact zyshen021@gmail.com; for all other failures contact mn@cs.unc.edu or open an issue in the github issue tracker.
-
12.30: added diffusion, curvature, and total variation registration models.
-
12.29: added multi Gaussian smoother
-
12.19
add result_analysis: an excel result analysis
1. detailed results: saved in fig_save_path/results.xlsx
** Sheet1: #total_filename * #metrics recorded by iteraton
** batch_0: #batch_filename * #metric_by_label recorded by iteration
** batch_1: same as batch_0
** ......
2. task results: saved in ../data/summary.xlsx
** Sheet1: task_name * #metrics recorded by iteration -
12.16
add cuda nearest neighbour interpolation, and many fixes -
12.14
add evalution method iou, recall, precison(overlapping), dice -
12.13
dataManger into the apps/mm-reg, the system works through
transfer to fileio when reading image in data_pool
add adaptive_padding in fileio.py, to be favourable by the guassian fft -
12.12
add dataManager, Registration net, dataloader for typical dataset. see sample in prepare data
add 'data_pool.py', a generalized dataset pool class
support inter intra personal dataset likeo asis, support inter-personal like lpba, IBSR, CUMU dataset
support 2D and 3D dataset
support labeled and unlabeled dataset
support 2D img, label slicing -
12.11
support oasis inter intra registration
support saving results, /data/saved_result/taskname, settings can be found in commandline parsers
support adaptiveNet (can be founded in smoother class) in llddmm mapping(tested), which embedding networks between m and v
(other possible combinations like [I_t, m and v], [I_0,m and v] also implemented but still underdevelop)
add m_1 into loss for avoid explosion ( maybe useful in small time step, cannot avoid explosion in large time_step) -
12.05: Added various momentum-based SVF models (SVFScalarMomentumImageNet, SVFScalarMomentumMapNet, SVFVectorMomentumImageNet, SVFVectorMomentumMapNet )
-
12.05: Added tests for the registration algorithms
-
12.02: Simplified interface. Petter parameters. Low-res capability (requires new arguent mapLowResFactor to be specified when calling the functions).
-
11.26: Created a separate IO class; use this for all file input/output
-
11.13: fix to adpat device
-
11.12: optimize, fix bugs,
-
10.31: unitest stn is added, see test_stn_cpu.py and test_stn_gpu.py
-
10.29: 1D, 2D and 3D cuda STN have been implemented. see pyreg.libraries.functions.stn_nd
-
stn = STNFunction_ND_BCXYZ(n_dim) stn.forward_stn(inputImage,inputGrids, output, n_dim, device_c, use_cuda=True) stn.backward_stn(inputImage, inputGrids, grad_input, grad_grids, grad_output, n_dim, device_c,use_cuda=True)
-
-
Important changes before 10.21:
-
add dataWarper.py, define some Variable Warppers, two are most common use
-
# AdaptVal(x): Adaptive Warper: used to adapt the data type, implemented on the existed Tensor/Variable # MyTensor: a adpative Tensor to create gpu, cpu and float16 Tensor
-
-
Changes are made in vector organization, batch is no longer calculated in loop
-
Memory efficient structure modification in "optimize" function
-
Real FFT is implemented for both speed and memory consideration, which needs the smoothing kernel should be symmetric
-
- make svf_quasi_momentum stable
- add smoothing unittest
- modify other testRegistration*
- add Tensorboard
- schedule the main function