| name | SINGA version | SINGA commit | parameter_url | parameter_sha1 | data_url | data_md5 | license | gist_id |
|---|---|---|---|---|---|---|---|---|
VGG model on Cifar-10 |
1.0.0 |
7956019cf326c5f84401551f31f1e597fba77d46 |
52fe5e237e9efa42f9d4ed081a5a2ee9bcd22c8e |
c58f30108f718f92721af3b95e74349a |
non-commercial |
51b10a13fbc3427934591f0e135f40a3 |
This example provides the training and serving scripts for VGG over CIFAR-10 data. The best validation accuracy (without data augmentation) we achieved was about 92%.
The folder structure for an example is as follows where README.md is required and other files are optional.
- README.md. Every example should have a README.md file for the model description, SINGA version and running instructions.
- train.py. The training script. Users should be able to run it directly by
python train.py. It is optional if the model is shared only for prediction or serving tasks. - serve.py. The serving script. It is typically used in the cloud mode, where users can submit the query via the web front end provided by Rafiki. If the local mode is enabled, it should accepts command line input. It is optional if the model is shared only for training tasks.
- model.py. It has the functions for creating the neural net. It could be merged into train.py and serve.py, hence are optional.
- index.html. This file is used for the serving task, which provides a web page for users to submit queries and get responses for the results. If is required for running the serving task in the cloud mode. If the model is shared only for training or running in the local mode, it is optional.
- requirements.txt. For specifying the python libraries used by users' code. It is optional if no third-party libs are used.
Some models may have other files and scripts. Typically, it is not recommended to put large files (e.g. >10MB) into this folder as it would be slow to clone the gist repo.
To run the scripts on your local computer, you need to install SINGA. Please refer to the installation page for detailed instructions.
The training data should be downloaded and decompressed into the current folder
wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
tar xvf cifar-10-python.tar.gz
The training program could be started by
python train.py
By default, the training is conducted on a GPU card, to use CPU for training (very slow), run
python train.py --use_cpu
The model parameters would be dumped periodically, into model-<epoch ID> and the last one is model.
The pre-trained model parameters should be downloaded and decompressed into the current folder.
This example does not have the serving script for local mode. To simulate the local mode, you can start the prediction script and use curl to pass the query image.
python serve.py &
curl -i -F image=@image1.jpg http://localhost:9999/api
curl -i -F image=@image2.jpg http://localhost:9999/api
curl -i -F image=@image3.jpg http://localhost:9999/api
The above commands start the serving program using the model trained for VGG as a daemon,
and then submit three queries (image1.jpg, image2.jpg, image3.jpg) to the port (the default port is 9999).
To use other port, please add -p PORT_NUMBER to the running command.
If you run the serving task after finishing the training task,
then the model parameters from model would be used.
Otherwise, it would use the one downloaded using data.py.
To run the scripts on the Rafiki platform, you don't need to install SINGA. But you need to add the dependent libs introduced by your code into the requirement.txt file.
The Rafiki front-end provides a web page for users to import gist repos directly.
Users just specify the HTTPS (NOT the git web URL) clone link and click load to import a repo.
The Rafiki font-end has a Job view for adding a new training job. Users need to configure the job type as 'training',
select the model (i.e. the repo added in the above step) and the training dataset. If the training dataset is not there, please upload it to Rafiki.
With these fields configured, the job could be started by clicking the start button.
Afterwards, the users would be redirected to the monitoring view.
Note that it may take sometime to download the data for the first time.
The Rafiki backend would run python train.py in the backend.
The serving job is similar to the training job except the job type is 'serving' and the pre-trained parameters are used.
The Rafiki backend would run python serve.py. Users can jump to the serving view rendered using the index.html from the gist repo.