customized_example

How to create and run a simple example

This model is a simple model that classifies images of clothes in 10 classes. The model is trained with the dataset FashionMNIST.

All things in this example are customized. There are some files:

• model.py: contains the arch's model
• handler.py: contains the handler that manage your data thus handles pre-processing-post of data
• parameters.pt: contains the state_dict of the model created with torch.save(model.state_dict(), 'parameters.pt')
• checkpoint.pt: contains the checkpoint of the model created with torch.save(model, 'checkpoint.pt')
• script_module.pt and trace_module.pt: are two files created respectively by torch.jit.script(model, 'script_module.pt') and torch.jit.trace(model, 'trace_module.pt') that are used in case of TorchScript mode

1. Create a new archiver in EagerMode

torch-model-archiver -f --model-name fashion --version 1.0 --handler ./customized_example/handler.py --model-file ./customized_example/model.py --serialized-file ./customized_example/parameters.pt --export-path ./model_store/ 

2. Run the Torchserve

torchserve --start --model-store model_store --ncs

3. Register the model

curl -X POST "http://localhost:8081/models?url=fashion.mar&model_name=fashion&model_version=1.0&handler=handler.py&batch_size=1&max_batch_delay=5000&initial_workers=1&synchronous=true"

4. Get an inference from Torchserve

In directory /img there are some images that you can use to test the model. You can use curl to get an inference from Torchserve:

curl "http://localhost:8080/predictions/fashion" -T ./customized_example/img/test1.jpg

If you want use file download.ipynb to download other images from the web. In first you have to install jupyter and ipykernel.