Release 0.0.71 (#78)

* Bump version and add mxnet to changelog

* Fix variable in test

* Try 0.0.7

* 0.0.7 already exists, boo

* Bump makefile to 0.0.71

* Skip downloads for hosted storage

* Fix bug in pickle version

* Add missing data load

* Fix pytorch example

* Fix pytorch-lightning example

* Capture model type for pytorch lightning models

* Fix error message

* Update pytorch-lightning example with right kwargs

* Freeze libomp version, version 12 breaks xgboost

CHANGELOG.md

@@ -1,14 +1,14 @@
 # Change log
 
-## Unreleased
+## modelstore 0.0.71
 
 🆕  Load models straight into memory! Model Store previously had `modelstore.download()` to download an artifact archive to a local path, it now also has `modelstore.load()` to load a model straight into memory. Currently works with `scikit-learn` models only.
 
 🆕  Upload models from frameworks that are not (yet) supported by modelstore! The `modelstore.upload()` function now works if you give it a `model=` kwarg that is a path to a file.
 
 🆕  Read a specific model's metadata with `modelstore.get_model_info()`
 
-🆕  Added [Annoy](https://github.com/spotify/annoy) and [ONNX](https://github.com/onnx/onnx) support.
+🆕  Added [Annoy](https://github.com/spotify/annoy), [ONNX](https://github.com/onnx/onnx), and [MXNet](https://mxnet.apache.org) (hybrid models) support.
 
 ## modelstore 0.0.65
 

bin/_setup_brew

@@ -5,6 +5,10 @@ brew update
 
 # To use xgboost
 # https://xgboost.readthedocs.io/en/latest/build.html#building-on-osx
-brew install libomp
+
+# Note: it looks like there's trouble with libomp 12
+# https://github.com/dmlc/xgboost/issues/7039
+
+brew install libomp@11.1.0
 
 echo "\n ✅  Done."

examples/Makefile-example

@@ -15,7 +15,7 @@ pyenv-local: pyenv-uninstall pyenv
 
 pyenv-test: pyenv-uninstall pyenv
 	pip uninstall -y modelstore
-	pip install -i https://test.pypi.org/simple/ modelstore==0.0.7
+	pip install -i https://test.pypi.org/simple/ modelstore==0.0.71
 
 pyenv-prod: pyenv-uninstall pyenv
 	pip uninstall -y modelstore

examples/examples-by-ml-library/libraries/keras_example.py

@@ -39,5 +39,6 @@ def load_and_test(modelstore: ModelStore, model_id: str):
     model = modelstore.load(DIABETES_DOMAIN, model_id)
 
     # Run some test predictions
+    _, X_test, _, y_test = load_diabetes_dataset()
     results = mean_squared_error(y_test, model.predict(X_test))
     print(f"🔍  Loaded model MSE={results}.")

examples/examples-by-ml-library/libraries/pytorch_example.py

@@ -36,16 +36,18 @@ def _train_example_model() -> ExampleNet:
 
     results = mean_squared_error(y_test, model(X_test).detach().numpy())
     print(f"🔍  Fit model MSE={results}.")
-    return model
+    return model, optimizer
 
 
 def train_and_upload(modelstore: ModelStore) -> dict:
     # Train a PyTorch model
-    model = _train_example_model()
+    model, optimizer = _train_example_model()
 
     # Upload the model to the model store
     print(f'⤴️  Uploading the pytorch model to the "{DIABETES_DOMAIN}" domain.')
-    meta_data = modelstore.upload(DIABETES_DOMAIN, model=model)
+    meta_data = modelstore.upload(
+        DIABETES_DOMAIN, model=model, optimizer=optimizer
+    )
     return meta_data
 
 

examples/examples-by-ml-library/libraries/pytorch_lightning_example.py

@@ -59,18 +59,18 @@ def _train_example_model() -> ExampleLightningNet:
 
     results = mean_squared_error(y_test, model(X_test).detach().numpy())
     print(f"🔍  Fit model MSE={results}.")
-    return model
+    return model, trainer
 
 
 def train_and_upload(modelstore: ModelStore) -> dict:
     # Train a PyTorch model
-    model = _train_example_model()
+    model, trainer = _train_example_model()
 
     # Upload the model to the model store
     print(
         f'⤴️  Uploading the pytorch lightning model to the "{DIABETES_DOMAIN}" domain.'
     )
-    meta_data = modelstore.upload(DIABETES_DOMAIN, model=model)
+    meta_data = modelstore.upload(DIABETES_DOMAIN, model=model, trainer=trainer)
     return meta_data
 
 

examples/examples-by-ml-library/main.py

@@ -66,9 +66,11 @@ def main(modelstore_in, ml_framework):
     example = EXAMPLES[ml_framework]
     meta_data = example.train_and_upload(modelstore)
 
-    # Run the example: download and load a model
-    model_id = meta_data["model"]["model_id"]
-    example.load_and_test(modelstore, model_id)
+    if modelstore_in != "hosted":
+        # Run the example: download and load a model
+        # Currently unimplemented in the hosted storage
+        model_id = meta_data["model"]["model_id"]
+        example.load_and_test(modelstore, model_id)
 
     # The upload returns meta-data about the model that was uploaded
     # In this example, we just print it out to the terminal

modelstore/meta/dependencies.py

@@ -12,14 +12,23 @@
 
 def _get_version(modname: str) -> str:
     try:
+        if modname == "pickle":
+            # pylint: disable=import-outside-toplevel
+            import pickle
+
+            return pickle.format_version
         if modname in sys.modules:
             mod = sys.modules[modname]
         else:
             mod = importlib.import_module(modname)
         return mod.__version__
     except AttributeError:
-        #  Annoy does not have a __version__
-        return pkg_resources.get_distribution(modname).version
+        try:
+            #  Annoy does not have a __version__
+            return pkg_resources.get_distribution(modname).version
+        except:
+            logger.debug("Unable to get %s's version", modname)
+            return None
     except ImportError:
         logger.debug("%s is not installed.", modname)
         return None

modelstore/model_store.py

@@ -123,7 +123,9 @@ def upload(self, domain: str, **kwargs) -> dict:
         for manager in self._managers:
             if manager.matches_with(**kwargs):
                 return manager.upload(domain, **kwargs)
-        raise ValueError("unable to upload: could not find matching manager")
+        raise ValueError(
+            "unable to upload: could not find matching manager (did you add all of the required kwargs?)"
+        )
 
     def load(self, domain: str, model_id: str):
         meta_data = self.get_model_info(domain, model_id)

modelstore/models/pytorch_lightning.py

@@ -28,7 +28,7 @@ class PyTorchLightningManager(ModelManager):
 
     """
     Model persistence for PyTorch Lightning models:
-    https://pytorch-lightning.readthedocs.io/en/stable/weights_loading.html#checkpoint-saving
+    https://pytorch-lightning.readthedocs.io/en/latest/common/weights_loading.html#manual-saving
 
     // @TODO: export as for onnx & torchscript
     https://pytorch-lightning.readthedocs.io/en/latest/new-project.html#predict-or-deploy
@@ -49,13 +49,15 @@ def optional_dependencies(cls) -> list:
         return deps + ["torch", "torchvision"]
 
     def _required_kwargs(self):
-        return ["trainer"]
+        return ["trainer", "model"]
 
     def matches_with(self, **kwargs) -> bool:
         # pylint: disable=import-outside-toplevel
-        from pytorch_lightning import Trainer
+        from pytorch_lightning import LightningModule, Trainer
 
-        return isinstance(kwargs.get("trainer"), Trainer)
+        return isinstance(kwargs.get("trainer"), Trainer) and isinstance(
+            kwargs.get("model"), LightningModule
+        )
 
     def _get_functions(self, **kwargs) -> list:
         if not self.matches_with(**kwargs):
@@ -105,7 +107,9 @@ def _model_file_path(parent_dir: str) -> str:
     return os.path.join(parent_dir, MODEL_CHECKPOINT)
 
 
-def _save_lightning_model(tmp_dir: str, trainer: "Trainer") -> str:
+def _save_lightning_model(
+    tmp_dir: str, trainer: "pytorch_lightning.Trainer"
+) -> str:
     file_path = _model_file_path(tmp_dir)
     trainer.save_checkpoint(file_path)
     return file_path

setup.py

@@ -8,10 +8,10 @@
 
 setup(
     name="modelstore",
-    version="0.0.7",
+    version="0.0.71",
     packages=find_packages(exclude=["tests", "examples", "docs"]),
     include_package_data=True,
-    description="modelstore is a library for versioning, exporting, and storing machine learning models",
+    description="modelstore is a library for versioning, exporting, storing, and loading machine learning models",
     long_description=long_description,
     long_description_content_type="text/markdown",
     url="https://github.com/operatorai/modelstore",

tests/meta/test_dependencies.py

@@ -39,6 +39,7 @@ def test_get_dependency_versions():
         "flake8",
         "isort",
         "a-missing-dependency",
+        "pickle",
     ]
     expected = {
         "annoy": "1.17.0",
@@ -48,6 +49,7 @@ def test_get_dependency_versions():
         "flake8": "3.8.4",
         "isort": "5.6.4",
         "a-missing-dependency": None,
+        "pickle": "4.0",
     }
     result = dependencies.get_dependency_versions(test_deps)
     assert result == expected

tests/models/test_mxnet.py

@@ -124,7 +124,7 @@ def test_load_model(tmp_path, mxnet_manager, mxnet_model):
             }
         },
     )
-    y_loaded_pred = loaded_model(y).asnumpy()
+    y_loaded_pred = loaded_model(x).asnumpy()
 
     # Expect the two to be the same
     assert isinstance(loaded_model, nn.SymbolBlock)

tests/models/test_pytorch_lightning.py

@@ -122,11 +122,13 @@ def test_model_data(lightning_manager, lightning_model):
 
 
 def test_required_kwargs(lightning_manager):
-    assert lightning_manager._required_kwargs() == ["trainer"]
+    assert lightning_manager._required_kwargs() == ["trainer", "model"]
 
 
-def test_matches_with(lightning_manager, lightning_trainer):
-    assert lightning_manager.matches_with(trainer=lightning_trainer)
+def test_matches_with(lightning_manager, lightning_trainer, lightning_model):
+    assert lightning_manager.matches_with(
+        trainer=lightning_trainer, model=lightning_model
+    )
     assert not lightning_manager.matches_with(model="a-string-value")
     assert not lightning_manager.matches_with(classifier=lightning_trainer)