Skip to content

Latest commit

 

History

History
163 lines (120 loc) · 5.03 KB

register.md

File metadata and controls

163 lines (120 loc) · 5.03 KB
Raw

1. Register a Model to MLModelCI

In the first tutorial, we demonstrate how to start to use MLModelCI. The first step is to publish a model in the system.

Firstly, make sure that you have started a MongoDB service and configured the MongoDB environment. See installation.

We have a quick script including automatic model registration, conversion, and profiling, for you. Just run this command in the package root.

python modelci/hub/init_data.py export --model {MODEL_NAME} --framework {FRAMEWORK}

Now let us finish this step by step.

Register a Pre-trained Model

MLModelCI provides two methods to publish models, using .yaml file and Python code.

With a simple configuration file [template]

from modelci.hub.manager import register_model_from_yaml

model_yaml = ...

register_model_from_yaml(model_yaml)

With Python API

We can register a pre-trained model using modelci.hub.manager.register_model(...). This API can trigger two functions with two parameters, respectively:

  • convert:

    Enabled by setting convert=True (default), which converts your model (a PyTorch nn.Module or a TensorFlow keras.Model object) into all possible optimized model formats.

  • profile:

    Enabled by setting profile=True. This will profile converted models automatically.

If you use convert=True, ModelCI will help you to convert your model automatically after adding to the database. Similarly, if you use profile=True, ModelCI will find a free device to start model benchmarking and profiling automatically.

Understanding convert=True mode

Here is an example of the registration using convert mode.

import torch.hub

from modelci.hub.manager import register_model
from modelci.types.bo.model_objects import IOShape, Framework, ModelVersion, Task, Metric
from modelci.types.trtis_objects import ModelInputFormat

model = torch.hub.load('pytorch/torchvision:v0.5.0', model='resnet50', pretrained=True)
inputs = [IOShape(shape=[-1, 3, 224, 224], dtype=float, format=ModelInputFormat.FORMAT_NCHW)]
outputs = [IOShape(shape=[-1, 1000], dtype=float)]

register_model(
    model,
    dataset='ImageNet',
    metric={Metric.ACC: 0.76},
    task=Task.IMAGE_CLASSIFICATION,
    inputs=inputs,
    outputs=outputs,
    architecture='ResNet50',
    framework=Framework.PYTORCH,
    version=ModelVersion(1),
    convert=True,
    profile=True
)

Understanding convert=False mode

Assume we have a saved pre-trained ResNet50 model at the current working directory named 1.zip. It was trained on ImageNet and converted by TorchScript. You should register it using using convert=False mode and set the engine=Engine.TorchScript

from modelci.hub.manager import register_model
from modelci.types.bo import Framework, IOShape, Engine, ModelVersion, Task, Metric

register_model(
    'path/to/model/1.zip',
    dataset='ImageNet',
    metric={Metric.ACC: 0.76},
    task=Task.IMAGE_CLASSIFICATION,
    inputs=[IOShape([-1, 3, 224, 224], float)],
    outputs=[IOShape([1], int)],
    architecture='ResNet50',
    framework=Framework.PYTORCH,
    engine=Engine.TORCHSCRIPT,
    version=ModelVersion(1),
    convert=False,
    profile=True
)

Trick: to save your time, you can follow the following rule to name your model: (See Tricks with Model Saved Path)
~/.modelci/<model name>/<framework>-<engine>/<task>/<version>.<extension> In this case, we are only able to specify the path, without architecture, framework, engine and version.

from modelci.hub.manager import register_model
from modelci.types.bo.model_objects import IOShape, Task, Metric

register_model(
    '~/.modelci/ResNet50/pytorch-torchscript/image_classification/1.zip',
    dataset='ImageNet',
    metric={Metric.ACC: 0.76},
    task=Task.IMAGE_CLASSIFICATION,
    inputs=[IOShape([-1, 3, 224, 224], float)],
    outputs=[IOShape([-1, 1000], float)],
    convert=False,
    profile=True,
)

Tricks with Model Saved Path

If you use modelci/hub/init_data.py to download model, the default model local cache path is as the following form:
~/.modelci/<model name>/<framework>-<engine>/<task>/<version>.<extension>

We can use modelci.hub.utils.parse_path(...) to extract model identification.

from modelci.hub.utils import parse_path

# from return value of model retrieval
model_bo = ...

info = parse_path(model_bo.saved_path)

The extracted information is a dictionary containing:

{
    "architecture": architecture,
    "framework": framework,
    "engine": engine,
    "task": task,
    "version": version,
    "filename": filename,
}

Vice versa, we can generate a default path by modelci.hub.utils.generate_path(...):

from modelci.hub.utils import generate_path
from modelci.types.bo import Framework, Engine, Task

saved_path = generate_path(
    model_name='ResNet50', 
    framework=Framework.PYTORCH, 
    task=Task.IMAGE_CLASSIFICATION,
    engine=Engine.TORCHSCRIPT, 
    version=1
)