Skip to content

Latest commit

 

History

History
206 lines (136 loc) · 15.5 KB

CONTRIBUTING.md

File metadata and controls

206 lines (136 loc) · 15.5 KB
Raw

Introduction

Welcome to MONAI Model Zoo! We're excited you're here and want to contribute. This documentation is intended for individuals and institutions interested in contributing medical imaging models to MONAI Model Zoo.

Preparing a bundle

Please refer to MONAI Bundle Specification and the description of bundle config syntax to prepare your bundle.

The get started notebook is a step-by-step tutorial to help developers easily get started to develop a bundle. And bundle examples show the typical bundle for 3D segmentation, how to use customized components in a bundle, and how to parse bundle in your own program as "hybrid" mode, etc.

As for the path related varibles within config files (such as "bundle_root"), we suggest to use path that do not include personal information (such as "/home/your_name/").The following is an example of path using:

"bundle_root": "/workspace/data/<bundle name>".

Readme Format

A template on how to prepare a readme file is provided.

License Format

As described in README, please include LICENSE into the root directory of the bundle, and please also include docs/data_license.txt if there are any license conditions stated for data your bundle uses. Here is an example of the brats_mri_segmentation bundle.

Model naming

The name of a bundle is suggested to contain its characteristics, such as include the task type and data type. The following are some of the examples:

spleen_deepedit_annotation
spleen_ct_segmentation
chest_xray_classification
pathology_metastasis_detection

In addition, please also define a display name via editing the "name" tag in metadata.json, and it will be shown in the model browser.

Model storage

Github limits the size of files allowed in the repository (see About size limits on GitHub). Therefore, MONAI Model Zoo limits each single file to be no larger than 25MB.

Prepare a config file for large files

If a bundle has large files, please upload those files into a publicly accessible source, and provide a config file called large_files.yml (or .yaml, .json) that contains the corresponding download links. During the pull request, only the config file should be included (large files should be excluded). Please put the config file within the root directory of the bundle, and it should contain the following information:

  1. path, relative path of the large file in the bundle.
  2. url, URL link that can download the file.
  3. hash_val, (optional) expected hash value of the file.
  4. hash_type, (optional) hash type. Supprted hash type includes "md5", "sha1", "sha256" and "sha512".

The template is as follow, and you can also click here to see an actual example of spleen_ct_segmentation:

large_files:
  - path: "models/model.pt"
    url: "url-of-model.pt"
    hash_val: ""
    hash_type: ""
  - path: "models/model.ts"
    url: "url-of-model.ts"

Preferred Files and Keys

In order to be compatible with other apps such as MONAI FL, MONAI deploy and MONAI Label, except the format requirements from MONAI Bundle Specification, a bundle in MONAI Model Zoo should also contain the necessary files LICENSE, configs/metadata.json (click here for instance), and the following preferred files:

  1. models/model.pt (or a download link in the config file for large files, click here for instance)
  2. configs/inference.json (or .yaml, .json, click here for instance)

If your bundle does not have any of the preferred files, please add the bundle name into exclude_verify_preferred_files_list in ci/bundle_custom_data.py.

Except the requirements of files, there are also some requirements of keys within config files:

In inference config file (if exists), please include the following keys: bundle_root (root directory of your bundle), device (required device), network_def (definition of the network components), inferer, and if there are any output files, the directory key should be defined as output_dir (click here for instance).

In train config file (if exists), please follow the following requirements in order to maintain consistent naming format (click here for instance):

  1. Please include keys bundle_root, device and dataset_dir, and if there are any output files, the directory key should be defined as output_dir.
  2. If having train, components train#trainer, train#trainer#max_epochs, train#dataset, train#dataset#data, train#handlers should be defined.
  3. If having validate, components validate#evaluator, validate#handlers, validate#dataset, validate#dataset#data should be defined.
  4. In train and/or validate, please define preprocessing, postprocessing, inferer and key_metric if they are used.
  5. If ValidationHandler is used, please define the key val_interval and use it for the argument interval.

Verifying a bundle

We prepared several premerge CI tests to verify your bundle.

Prepare dependencies

For dependencies that support pip install command, please put them into optional_packages_version in configs/metadata.json (click here for instance), and the CI test program will extract and install all libraries directly before running tests. For dependencies that require multiple steps to install, please prepare a install script in ci/install_scripts, and put the bundle name and the script path into install_dependency_dict in ci/bundle_custom_data.py (click here for instance).

Necessary tests

  1. Check if necessary files are existing in your bundle.
  2. Check if keys naming are consistent with our requirements.
  3. If an existing bundle has been modified, check if version and changelog are updated.
  4. Check if metadata format is correct. You can also run the following command locally to verify your bundle before submitting a pull request:
python -m monai.bundle verify_metadata --meta_file configs/metadata.json --filepath eval/schema.json

Optional tests

Verify data shape and data type

Check if the input and output data shape and data type of network defined in the metadata are correct. You can also run the following command locally to verify your bundle before submitting a pull request.

python -m monai.bundle verify_net_in_out --net_id network_def --meta_file configs/metadata.json --config_file configs/inference.json

net_id is the ID name of the network component, config_file is the filepath (within the bundle) of the config file to get the network definition. Please modify the default values if needed.

If this test is not suitable for your bundle, please add your bundle name into exclude_verify_shape_list in ci/bundle_custom_data.py.

Verify torchscript

Check the functionality of exporting the checkpoint to TorchScript file. You can also run the following command locally to verify your bundle before submitting a pull request.

python -m monai.bundle ckpt_export --net_id network_def --filepath models/model.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.json

After exporting your TorchScript file, you can check the evaluation or inference results based on it rather than model.pt with the following changes:

  1. Remove or disable CheckpointLoader in evaluation or inference config file if exists.
  2. Define network_def as: "$torch.jit.load(<your TorchScript file path>)".
  3. Execute evaluation or inference command.

If your bundle does not support TorchScript, please mention it in docs/README.md, and add your bundle name into exclude_verify_torchscript_list in ci/bundle_custom_data.py.

Verify TensorRT models

Check the functionality of exporting the checkpoint to TensorRT based models. TensorRT based models target NVIDIA GPUs via NVIDIA’s TensorRT Deep Learning Optimizer and Runtime. It can accelerate models' inference on NVIDIA GPU with speedup ratio up to 6x by converting models weight to float32 or float16 precision. In MONAI, models are compiled to TensorRT based models through Torch_TensorRT. To use the TensorRT conversion to accelerate model inference, you must have a NVIDIA GPU, install TensorRT and Torch-TensorRT. We have started testing most of our models on NVIDIA A100 80G GPU with Torch_TensorRT version >= 1.4.0 and TensorRT version >= 8.5.3. Please make sure your environment meets these minimum requirements. Or you can use the MONAI docker where the MONAI version is >= 1.2. Currently, only a subset of models in the MONAI model-zoo support TensorRT conversion. More models will be covered in the future with the new Torch-TensorRT and TensorRT version. The specific versions of these two libraries that support TensorRT conversion for the corresponding models can be found in the README file of the bundle. You can also run the following command locally to export your bundle to float32 precision or float16 precision.

python -m monai.bundle trt_export --net_id network_def --filepath models/model_trt.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.json --precision <fp32/fp16> --dynamic_batchsize "[min, opt, max]"

The other way to export a PyTorch model to a TensorRT engine-based TorchScript is through the ONNX-TensorRT. This way can solve the slowdown issue of some models. It can be simply achieved by adding an extra option --use_onnx "True" to the export command. Please notice that by default the ONNX-TensorRT way assigns only one output to a model. If a model has many outputs, please use the --onnx_output_names to specify the name of outputs.

python -m monai.bundle trt_export --net_id network_def --filepath models/model_trt.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.json --precision <fp32/fp16> --dynamic_batchsize "[min, opt, max]" --use_onnx "True" --onnx_output_names "['output_0', 'output_1', ..., 'output_N']"

After exported your TensorRT based models, you can check the evaluation or inference results based on it rather than model.pt with the following steps:

  1. Add a $import torch_tensorrt at the import part of inference.json file.
  2. Remove or disable CheckpointLoader in evaluation or inference config file if exists.
  3. Define network_def as: "$torch.jit.load(<your TensorRT model path>)".
  4. Set the amp parameter in the evaluator to false.

All above steps can be covered in an inference_trt.json like this example. Update the inference_trt.json based on your bundle and put it in the configs folder of the bundle. Then run the following command to execute the inference:

python -m monai.bundle run --config_file "['configs/inference.json', 'configs/inference_trt.json']"

If your bundle does not support TensorRT compilation, please mention it in docs/README.md.

Customized unit tests

It is recommended to prepare unit tests on your bundle. The main purpose is to ensure each config file is runnable. Please create a file called test_<bundle_name>.py in ci/unit_tests/ (like this example), and you can define the testing scope within the file. If multi-gpu config files are also need to be tested, please create a separate file called test_<bundle_name>_dist.py in the same directory (like this example).

Code format tests

If there are any .py files in your bundle, coding style is checked and enforced by flake8, black, isort and pytype. Before submitting a pull request, we recommend that all checks should pass by running the following command locally:

# optionally update the dependencies and dev tools
python -m pip install -U pip
python -m pip install -U -r requirements-dev.txt

# run the linting and type checking tools
./runtests.sh --codeformat

# try to fix the coding style errors automatically
./runtests.sh --autofix

Submitting pull requests

All code changes to the dev branch must be done via pull requests.

  1. Please create a new ticket from the issue list.
  2. create a new branch in your fork of the codebase named [ticket_id]-[task_name]. Ideally, the new branch should be based on the latest dev branch.
  3. Make changes to the branch (use detailed commit messages if possible).
  4. Create a new pull request from the task branch to the dev branch, with detailed descriptions of the purpose of this pull request.
  5. For any large files inside the bundle, please exclude them and provide the download links instead. Please follow the instructions mentioned above to prepare the necessary large_files.yml.
  6. Wait for reviews; if there are reviews, make point-to-point responses, make further code changes if needed.
  7. If there are conflicts between the pull request branch and the dev branch, pull the changes from the dev and resolve the conflicts locally.
  8. Reviewer and contributor may have discussions back and forth until all comments addressed.
  9. Wait for the pull request to be merged.

Validate and release

As for a pull request, a CI program will try to download all large files if mentioned and do several validations. If the pull request is approved and merged, the full bundle (with all large files if exists) will be archived and send to Releases.

Remove a bundle

If an existing bundle needs to be removed, please follow the steps mentioned above to do the delete manipulations during the created new branch in your fork, and then create a new pull request. However, bundles that are already stored in the release will not be changed if the corresponding source code is removed.