Sherif akoush/mlsv 42/add alibi explain runtime

docs/examples/alibi-explain/model-settings.json

@@ -0,0 +1,14 @@
+{
+  "name": "anchor-image-explain-model",
+  "implementation": "mlserver_alibi_explain.AlibiExplainRuntime",
+  "parallel_workers": 0,
+  "parameters": {
+    "uri": "./data/mnist_anchor_image",
+    "version": "v0.1.0",
+    "extra": {
+      "explainer_type": "anchor_image",
+      "infer_uri": "http://localhost:42315/v2/models/test-pytorch-mnist/infer"
+    }
+  }
+}
+

docs/examples/alibi-explain/settings.json

@@ -0,0 +1,3 @@
+{
+    "debug": "true"
+}

mlserver/codecs/numpy.py

@@ -1,8 +1,8 @@
 import numpy as np
 
-from typing import Any
+from typing import Any, Union
 
-from ..types import RequestInput, ResponseOutput
+from ..types import RequestInput, ResponseOutput, Parameters
 
 from .base import InputCodec, register_input_codec, register_request_codec
 from .utils import FirstInputRequestCodec
@@ -30,13 +30,13 @@
 _NumpyToDatatype["S"] = "BYTES"
 
 
-def _to_dtype(request_input: RequestInput) -> "np.dtype":
-    dtype = _DatatypeToNumpy[request_input.datatype]
+def _to_dtype(v2_data: Union[RequestInput, ResponseOutput]) -> "np.dtype":
+    dtype = _DatatypeToNumpy[v2_data.datatype]
 
-    if request_input.datatype == "BYTES":
+    if v2_data.datatype == "BYTES":
         # bytes have variable size, so need to specify as part of type
         # TODO: Make elem size variable (and not just the last dimension)
-        elem_size = request_input.shape[-1]
+        elem_size = v2_data.shape[-1]
         return np.dtype((dtype, elem_size))
 
     return np.dtype(dtype)
@@ -54,11 +54,11 @@ def to_datatype(dtype: np.dtype) -> str:
     return datatype
 
 
-def _to_ndarray(request_input: RequestInput) -> np.ndarray:
-    data = getattr(request_input.data, "__root__", request_input.data)
-    dtype = _to_dtype(request_input)
+def _to_ndarray(v2_data: Union[RequestInput, ResponseOutput]) -> np.ndarray:
+    data = getattr(v2_data.data, "__root__", v2_data.data)
+    dtype = _to_dtype(v2_data)
 
-    if request_input.datatype == "BYTES":
+    if v2_data.datatype == "BYTES":
         return np.frombuffer(data, dtype)
 
     return np.array(data, dtype)
@@ -99,6 +99,24 @@ def decode(cls, request_input: RequestInput) -> np.ndarray:
         # TODO: Check if reshape not valid
         return model_data.reshape(request_input.shape)
 
+    @classmethod
+    def decode_response_output(cls, response_output: ResponseOutput) -> np.ndarray:
+        # TODO: merge this logic with `decode`
+        return cls.decode(response_output)  # type: ignore
+
+    @classmethod
+    def encode_request_input(cls, name: str, payload: np.ndarray) -> RequestInput:
+        # TODO: merge this logic with `encode`
+        data = cls.encode(name=name, payload=payload)
+
+        return RequestInput(
+            name=data.name,
+            datatype=data.datatype,
+            shape=data.shape,
+            data=data.data,
+            parameters=Parameters(content_type=cls.ContentType),
+        )
+
 
 @register_request_codec
 class NumpyRequestCodec(FirstInputRequestCodec):

mlserver/codecs/string.py

@@ -1,6 +1,6 @@
 from typing import List
 
-from ..types import RequestInput, ResponseOutput
+from ..types import RequestInput, ResponseOutput, Parameters
 
 from .utils import FirstInputRequestCodec
 from .base import InputCodec, register_input_codec, register_request_codec
@@ -51,6 +51,19 @@ def decode(cls, request_input: RequestInput) -> List[str]:
         unpacked = map(_decode_str, unpack(packed, shape))
         return list(unpacked)
 
+    @classmethod
+    def encode_request_input(cls, name: str, payload: List[str]) -> RequestInput:
+        # TODO: merge this logic with `encode`
+        # note: this will only work with REST and not grpc as we might have
+        # variable length strings
+        return RequestInput(
+            name=name,
+            datatype="BYTES",
+            shape=[len(payload), -1],  # this is discarded downstream?
+            data=payload,
+            parameters=Parameters(content_type=cls.ContentType),
+        )
+
 
 @register_request_codec
 class StringRequestCodec(FirstInputRequestCodec):

mlserver/errors.py

@@ -24,3 +24,13 @@ def __init__(self, name: str, version: str = None):
 class InferenceError(MLServerError):
     def __init__(self, msg: str):
         super().__init__(msg)
+
+
+class RemoteInferenceError(MLServerError):
+    def __init__(self, code: int, reason: str):
+        super().__init__(f"Remote inference call failed with {code}, {reason}")
+
+
+class ModelParametersMissing(MLServerError):
+    def __init__(self, model_name: str):
+        super().__init__(f"Parameters missing for model {model_name}")

mlserver/handlers/dataplane.py

@@ -54,7 +54,7 @@ async def infer(
         model = await self._model_registry.get_model(name, version)
 
         # Run middlewares
-        inference_middlewares(payload, model._settings)
+        inference_middlewares(payload, model.settings)
 
         # TODO: Make await optional for sync methods
         prediction = await model.predict(payload)

mlserver/model.py

@@ -1,20 +1,22 @@
 from typing import Any, Dict, Optional, List
 
+from .codecs import (
+    decode_request_input,
+    decode_inference_request,
+    InputCodec,
+    has_decoded,
+    get_decoded,
+)
+from .settings import ModelSettings
 from .types import (
     InferenceRequest,
     InferenceResponse,
     RequestInput,
-    Parameters,
     MetadataModelResponse,
     MetadataTensor,
 )
-from .settings import ModelSettings
-from .codecs import (
-    decode_request_input,
-    decode_inference_request,
-    InputCodec,
-    has_decoded,
-    get_decoded,
+from .types import (
+    Parameters,
 )
 
 

runtimes/alibi-detect/mlserver_alibi_detect/cd/runtime.py

@@ -33,7 +33,7 @@ async def load(self) -> bool:
 
             parameters = self.alibi_detect_settings.init_parameters
 
-            self._model = drift_detector(**detector_data, **parameters)
+            self._model = drift_detector(**detector_data, **parameters)  # type: ignore
 
         self.ready = True
         return self.ready

runtimes/alibi-detect/mlserver_alibi_detect/runtime.py

@@ -26,7 +26,7 @@ class Config:
         env_prefix = ENV_PREFIX_ALIBI_DETECT_SETTINGS
 
     init_detector: bool = False
-    detector_type: PyObject = ""
+    detector_type: PyObject = ""  # type: ignore
     protocol: Optional[str] = "seldon.http"
     init_parameters: Optional[dict] = {}
     predict_parameters: Optional[dict] = {}
@@ -46,8 +46,11 @@ class AlibiDetectRuntime(MLModel):
     """
 
     def __init__(self, settings: ModelSettings):
-
-        self.alibi_detect_settings = AlibiDetectSettings(**settings.parameters.extra)
+        if settings.parameters is None:
+            self.alibi_detect_settings = AlibiDetectSettings()
+        else:
+            extra = settings.parameters.extra
+            self.alibi_detect_settings = AlibiDetectSettings(**extra)  # type: ignore
         super().__init__(settings)
 
     @custom_handler(rest_path="/")
@@ -93,7 +96,7 @@ async def predict(self, payload: types.InferenceRequest) -> types.InferenceRespo
 
     async def predict_fn(self, input_data: Any) -> dict:
         parameters = self.alibi_detect_settings.predict_parameters
-        return self._model.predict(input_data, **parameters)
+        return self._model.predict(input_data, **parameters)  # type: ignore
 
     def _check_request(self, payload: types.InferenceRequest) -> types.InferenceRequest:
         if len(payload.inputs) != 1: