Imagine you developed a fancy model, living in /path/to/fancy_project, and you want to share your amazing results on the Dynabench model leaderboard. Dynalab makes it easy for you to do just that.
Dynalab is MIT-licensed.
Dynalab has been tested on Python 3.6+, on Mac OS and Ubuntu.
git clone https://github.com/facebookresearch/dynalab.git
cd dynalab
pip install -e .
You will also need to install docker. The Docker version we use is 20.10.5.
Run the following command to initialize your project folder for model upload:
$ cd /path/to/fancy_project # We will refer to this as the root path
$ dynalab-cli init -n <name_of_your_model>
Follow the prompts to configure this folder. You can find more information about the config of files by running dynalab-cli init -h. Make sure that all files you specified on initialization are physically inside this project folder, e.g. not soft linked from elsewhere, otherwise your may encounter errors later, and your model deployment may fail. You should assume no external internet access from the docker.
From now on, you should always run dynalab-cli from the root path, otherwise it will get confused and you may see weird errors.
If you don't already have a handler file, we will have created a template for you with instructions to fill at ./handler.py. The handler file defines how your model takes inputs, runs inference and returns a response. Follow the instructions in the template file to complete the handler.
For the expected model I/O format of your task, check the definitions here.
Now that you completed the handler file, run a local test to see if your code works correctly.
$ dynalab-cli test --local -n <name_of_your_model>
If your local test is successful, you'll see "Local test passed" on the prompt. You can then move on to the next step. Otherwise, fix your project according to the error prompt and re-run this step until the output is error free.
Exclude large files / folders
You may get an error if your project folder is too big (e.g. more than 2GB). You can reduce its size by excluding files / folders that are not relevant to your model (e.g. unused checkpoints). To do this, add the paths to the files / folders that you want to exclude into the config by running dynalab-cli init -n <name_of_your_model> --amend and update the exclude entry, e.g.
{
"exclude": ["checkpoints_folder", "config/irrelevant_config.txt"]
}
Remember not to exclude files / folders that are used by your model.
The integrated test will run the test inside a mock docker container to simulate the deployment environment, which is on Ubuntu 18.04 and uses Python 3.6 (see dockerfile for detailed version information). It may take some time to download dependencies in the docker.
$ dynalab-cli test -n <name_of_your_model>
If the integrated test is successful, you'll see "Integrated test passed" on the prompt. You can then proceed to the next step. Otherwise, please follow the on-screen instructions to check the log and fix your code / dependencies, and repeat this step until the output is error free.
If the integrated test is unsuccessful, it is possible that your machine lacks the resources to run the deployment environment in docker, or that you do not have sufficient resources allocated to docker. If this happens, your log file will show that workers crashed but will not include an error that references your handler.py. Uploading your model could result in fully functional behavior on our server in this scenario, even though the integrated test fails. However, we would strongly recommend running and passing the integrated test with more allocated resources prior to model upload.
Third party libraries
If your code uses third-party libraries, you may specify them via either requirements.txt or setup.py. Then call dynalab-cli init -n <name_of_your_model> --amend to update the corresponding entry in the config file
{
"requirements": true | false, # true if installing dependencies using requirements.txt
"setup": true | false # true if installing dependencies using setup.py
}
Some common libraries are pre-installed so you do not need to include them in your requirements.txt or setup.py, unless you need a different version. Please check the dockerfile for the supported libraries. At the moment, supported libraries include
torch==1.7.1
Extra model files There may be a config file, or self-defined modules that you want to read or import in your handler. There are two ways to do this.
- Include these files in the dynalab config and read / import them directly without worrying about paths. This also means that all file structure will be flattened. Firstly, run
dynalab-cli init -n <name_of_your_model> --amendand fill themodel_fileslist with the list of file paths inside the root directory, e.g.Then in the handler, to read a file, you can read the config by its name, i.e. no path needs to be specified{ "model_files": ["configs/model_config.json", "src/my_model.py"] }and directly import the module by its name, i.e. no path needs to be specifiedconfig = json.load("model_config.json")We recommend using this method to read files (e.g. configs, vocabularies) which is often flat-structured by its nature.import my_model - If you do not want to flatten the file structure (e.g. there may be too many dependencies involved), you do not need to add them to the dynalab config. First of all, notice there is a
ROOTPATHvariable available in your handler template. Suppose the file locations are the same as those specified above (configs/model_config.jsonandsrc/my_model.py), you will read the config byand import the module byconfig = json.load(os.path.join(ROOTPATH, "configs", "model_config.json"))We recommend using this method for importing self-defined modules.import sys sys.path.append(ROOTPATH) # you can uncomment this line in the handler template from src import my_model
Make sure you pass the integrated test in Step 4 before submitting the model, otherwise your model deployment might fail. You will first need to log in by running
$ dynalab-cli login
You will be taken to the Dynabench website where you'll see an API token (you'll be asked to log in there if you haven't). Click the "Copy" button on the webpage and paste that back in the terminal prompt.
To upload your model, run
$ dynalab-cli upload -n <name_of_your_model>
Follow the on-screen instructions for uploading the model. After the model is uploaded, it will enter our deployment queue, and you will receive an email when the deployment is done. If deployment is successful, your model will be evaluated on the datasets for that task, and you will be able to see the results on your model page. You can then publish the model for the results to be shown in the leaderboard.
Please create an issue.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent