Module fl_server_api.views.inference¶
View Source
# SPDX-FileCopyrightText: 2024 Benedikt Franke <benedikt.franke@dlr.de>
# SPDX-FileCopyrightText: 2024 Florian Heinrich <florian.heinrich@dlr.de>
#
# SPDX-License-Identifier: Apache-2.0
import base64
from django.http import HttpRequest, HttpResponse
from drf_spectacular.utils import inline_serializer, extend_schema, OpenApiExample
import json
from io import BytesIO
import pickle
from PIL import Image
from rest_framework import status
from rest_framework.exceptions import APIException, UnsupportedMediaType, ValidationError
from rest_framework.fields import CharField, ChoiceField, DictField, FloatField, ListField
import torch
from torchvision.transforms.functional import to_tensor
from typing import Any, Dict, List, Literal, Optional, Tuple, Type
from fl_server_ai.uncertainty import get_uncertainty_class, UncertaintyBase
from fl_server_core.exceptions import TorchDeserializationException
from fl_server_core.models import Model, GlobalModel, LocalModel
from fl_server_core.utils.logging import disable_logger
from fl_server_core.utils.torch_serialization import to_torch_tensor
from .base import ViewSet
from ..serializers.generic import ErrorSerializer
from ..utils import get_entity
class Inference(ViewSet):
"""
Inference ViewSet for performing inference on a model.
"""
serializer_class = inline_serializer("InferenceSerializer", fields={
"inference": ListField(child=ListField(child=FloatField())),
"uncertainty": DictField(child=FloatField())
})
"""The serializer for the ViewSet."""
@extend_schema(
request=inline_serializer(
"InferenceJsonSerializer",
fields={
"model_id": CharField(),
"model_input": ListField(child=ListField(child=FloatField())),
"return_format": ChoiceField(["binary", "json"])
}
),
responses={
status.HTTP_200_OK: serializer_class,
status.HTTP_400_BAD_REQUEST: ErrorSerializer,
},
examples=[
OpenApiExample("JSON Example", value={
"model_id": "mymodel",
"model_input": [
[1.0, 2.3, -0.4, 3],
[0.01, 9.7, 5.6, 7]
],
"return_format": "json"
}, request_only=True),
]
)
def inference(self, request: HttpRequest) -> HttpResponse:
"""
Performs inference on the provided model and input data.
This method takes in an HTTP request containing the necessary metadata and input data,
performs any required preprocessing on the input data, runs the inference using the specified model,
and returns a response in the format specified by the `return_format` parameter including
possible uncertainty measurements if defined.
Args:
request (HttpRequest): The current HTTP request.
Returns:
HttpResponse: A HttpResponse containing the result of the inference as well as its uncertainty.
"""
request_body, is_json = self._get_handle_content_type(request)
model, preprocessing, input_shape, return_format = self._get_inference_metadata(
request_body,
"json" if is_json else "binary"
)
model_input = self._get_model_input(request, request_body)
if preprocessing:
model_input = preprocessing(model_input)
else:
# if no preprocessing is defined, at least try to convert/interpret the model_input as
# PyTorch tensor, before raising an exception
model_input = self._try_cast_model_input_to_tensor(model_input)
self._validate_model_input_after_preprocessing(model_input, input_shape, bool(preprocessing))
uncertainty_cls, inference, uncertainty = self._do_inference(model, model_input)
return self._make_response(uncertainty_cls, inference, uncertainty, return_format)
def _get_handle_content_type(self, request: HttpRequest) -> Tuple[dict, bool]:
"""
Handles HTTP request body based on their content type.
This function checks if the request content type is either `application/json`
or `multipart/form-data`. If it matches, it returns the corresponding data and
a boolean indicating whether it's JSON (True) or multipart/form-data (False).
Args:
request (HttpRequest): The request.
Returns:
tuple: A tuple containing the parsed data and a boolean indicating the content type.
* If content type is `application/json`, returns the JSON payload as a Python object (dict)
and True to indicate it's JSON.
* If content type is `multipart/form-data`, returns the request POST data and False.
Raises:
UnsupportedMediaType: If an unknown content type is specified, raising an error with
details on supported types (`application/json` and `multipart/form-data`).
"""
match request.content_type.lower():
case s if s.startswith("multipart/form-data"):
return request.POST, False
case s if s.startswith("application/json"):
return json.loads(request.body), True
# if the content type is specified, but not supported, return 415
self._logger.error(f"Unknown Content-Type '{request.content_type}'")
raise UnsupportedMediaType(
"Only Content-Type 'application/json' and 'multipart/form-data' is supported."
)
def _get_inference_metadata(
self,
request_body: dict,
return_format_default: Literal["binary", "json"]
) -> Tuple[Model, Optional[torch.nn.Module], Optional[List[Optional[int]]], str]:
"""
Retrieves inference metadata based on the content of the provided request body.
This method checks if a `model_id` is present in the request body and retrieves
the corresponding model entity. It then determines the return format based on the
request body or default to one of the two supported formats (`binary` or `json`).
Args:
request_body (dict): The data sent with the request, containing at least `model_id`.
return_format_default (Literal["binary", "json"]): The default return format to use if not specified in
the request body.
Returns:
Tuple[Model, Optional[torch.nn.Module], Optional[List[Optional[int]]], str]: A tuple containing:
* The retrieved model entity.
* The global model's preprocessing torch module (if applicable).
* The input shape of the global model (if applicable).
* The return format (`binary` or `json`).
Raises:
ValidationError: If no valid `model_id` is provided in the request body, or if an unknown return format
is specified.
"""
if "model_id" not in request_body:
self._logger.error("No 'model_id' provided in request.")
raise ValidationError("No 'model_id' provided in request.")
model_id = request_body["model_id"]
model = get_entity(Model, pk=model_id)
return_format = request_body.get("return_format", return_format_default)
if return_format not in ["binary", "json"]:
self._logger.error(f"Unknown return format '{return_format}'. Supported are binary and json.")
raise ValidationError(f"Unknown return format '{return_format}'. Supported are binary and json.")
global_model: Optional[GlobalModel] = None
if isinstance(model, GlobalModel):
global_model = model
elif isinstance(model, LocalModel):
global_model = model.base_model
else:
self._logger.error("Unknown model type. Not a GlobalModel and not a LocalModel. Skip preprocessing.")
preprocessing: Optional[torch.nn.Module] = None
input_shape: Optional[List[Optional[int]]] = None
if global_model:
if global_model.preprocessing is not None:
preprocessing = global_model.get_preprocessing_torch_model()
if global_model.input_shape is not None:
input_shape = global_model.input_shape
return model, preprocessing, input_shape, return_format
def _get_model_input(self, request: HttpRequest, request_body: dict) -> Any:
"""
Retrieves and decodes the model input from either an uploaded file or the request body.
Args:
request (HttpRequest): The current HTTP request.
request_body (dict): The parsed request body as a dictionary.
Returns:
Any: The decoded model input data.
Raises:
ValidationError: If no `model_input` is found in the uploaded file or the request body.
"""
uploaded_file = request.FILES.get("model_input", None)
if uploaded_file and uploaded_file.file:
model_input = uploaded_file.file.read()
else:
model_input = request_body.get("model_input", None)
if not model_input:
raise ValidationError("No uploaded file 'model_input' found.")
return self._try_decode_model_input(model_input)
def _try_decode_model_input(self, model_input: Any) -> Any:
"""
Attempts to decode the input `model_input` from various formats and returns it in a usable form.
This function first tries to deserialize the input as a PyTorch tensor. If that fails, it attempts to
decode the input as a base64-encoded string. If neither attempt is successful, the original input is returned.
Args:
model_input (Any): The input to be decoded, which can be in any format.
Returns:
Any: The decoded input, which may still be in an unknown format if decoding attempts fail.
"""
# 1. try to deserialize model_input as PyTorch tensor
try:
with disable_logger(self._logger):
model_input = to_torch_tensor(model_input)
except Exception:
pass
# 2. try to decode model_input as base64
try:
is_base64, tmp_model_input = self._is_base64(model_input)
if is_base64:
model_input = tmp_model_input
except Exception:
pass
# result
return model_input
def _try_cast_model_input_to_tensor(self, model_input: Any) -> Any:
"""
Attempt to cast the given model input to a PyTorch tensor.
This function tries to interpret the input in several formats:
1. PIL Image (and later convert it to a PyTorch tensor, see 3.)
2. PyTorch tensor via `torch.as_tensor`
3. PyTorch tensor via torchvision `ToTensor` (supports e.g. PIL images)
If none of these attempts are successful, the original input is returned.
Args:
model_input: The input data to be cast to a PyTorch tensor.
Can be any type that can be converted to a tensor.
Returns:
A PyTorch tensor representation of the input data, or the original
input if it cannot be converted.
"""
def _try_to_pil_image(model_input: Any) -> Any:
stream = BytesIO(model_input)
return Image.open(stream)
if isinstance(model_input, torch.Tensor):
return model_input
# In the following order, try to:
# 1. interpret model_input as PIL image (and later to PyTorch tensor, see step 3),
# 2. interpret model_input as PyTorch tensor,
# 3. interpret model_input as PyTorch tensor via torchvision ToTensor (supports e.g. PIL images).
for fn in [_try_to_pil_image, torch.as_tensor, to_tensor]:
try:
model_input = fn(model_input) # type: ignore
except Exception:
pass
return model_input
def _is_base64(self, sb: str | bytes) -> Tuple[bool, bytes]:
"""
Check if a string or bytes object is a valid Base64 encoded string.
This function checks if the input can be decoded and re-encoded without any changes.
If decoding and encoding returns the same result as the original input, it's likely
that the input was indeed a valid Base64 encoded string.
Note: This code is based on the reference implementation from the linked Stack Overflow answer.
Args:
sb (str | bytes): The input string or bytes object to check.
Returns:
Tuple[bool, bytes]: A tuple containing a boolean indicating whether the input is
a valid Base64 encoded string and the decoded bytes if it is.
References:
https://stackoverflow.com/a/45928164
"""
try:
if isinstance(sb, str):
# If there's any unicode here, an exception will be thrown and the function will return false
sb_bytes = bytes(sb, "ascii")
elif isinstance(sb, bytes):
sb_bytes = sb
else:
raise ValueError("Argument must be string or bytes")
decoded = base64.b64decode(sb_bytes)
return base64.b64encode(decoded) == sb_bytes, decoded
except Exception:
return False, b""
def _validate_model_input_after_preprocessing(
self,
model_input: Any,
model_input_shape: Optional[List[Optional[int]]],
preprocessing: bool
) -> None:
"""
Validates the model input after preprocessing.
Ensures that the provided `model_input` is a valid PyTorch tensor and its shape matches
the expected`model_input_shape`.
Args:
model_input (Any): The model input to be validated.
model_input_shape (Optional[List[Optional[int]]]): The expected shape of the model input.
Can contain None values if not all dimensions are fixed (e.g. first dimension as batch size).
preprocessing (bool): Whether a preprocessing model was defined or not. (Only for a better error message.)
Raises:
ValidationError: If the `model_input` is not a valid PyTorch tensor or
its shape does not match the expected `model_input_shape`.
"""
if not isinstance(model_input, torch.Tensor):
msg = "Model input could not be casted or interpreted as a PyTorch tensor object"
if preprocessing:
msg += " and is still not a PyTorch tensor after preprecessing."
else:
msg += " and no preprecessing is defined."
raise ValidationError(msg)
if model_input_shape and not all(
dim_input == dim_model
for (dim_input, dim_model) in zip(model_input.shape, model_input_shape)
if dim_model is not None
):
raise ValidationError("Input shape does not match model input shape.")
def _make_response(
self,
uncertainty_cls: Type[UncertaintyBase],
inference: torch.Tensor,
uncertainty: Any,
return_type: str
) -> HttpResponse:
"""
Build the response object with the result data.
This method checks the return type and makes a response with the appropriate content type.
If return_type is "binary", a binary-encoded response will be generated using pickle.
Otherwise, a JSON response will be generated by serializing the uncertainty object using its to_json method.
Args:
uncertainty_cls (Type[UncertaintyBase]): The uncertainty class.
inference (torch.Tensor): The inference.
uncertainty (Any): The uncertainty.
return_type (str): The return type.
Returns:
HttpResponse: The inference result response.
"""
if return_type == "binary":
response_bytes = pickle.dumps(dict(inference=inference, uncertainty=uncertainty))
return HttpResponse(response_bytes, content_type="application/octet-stream")
return HttpResponse(uncertainty_cls.to_json(inference, uncertainty), content_type="application/json")
def _do_inference(
self, model: Model, input_tensor: torch.Tensor
) -> Tuple[Type[UncertaintyBase], torch.Tensor, Dict[str, Any]]:
"""
Perform inference on a given input tensor using the provided model.
This methods retrieves the uncertainty class, performs the prediction.
The output of this method consists of:
* The uncertainty class used for inference
* The result of the model's prediction on the input tensor
* Any associated uncertainty for the prediction
Args:
model (Model): The model to perform inference with.
input_tensor (torch.Tensor): Input tensor to pass through the model.
Returns:
Tuple[Type[UncertaintyBase], torch.Tensor, Dict[str, Any]]:
A tuple containing the uncertainty class, prediction result, and any associated uncertainty.
Raises:
APIException: If an error occurs during inference
"""
try:
uncertainty_cls = get_uncertainty_class(model)
inference, uncertainty = uncertainty_cls.prediction(input_tensor, model)
return uncertainty_cls, inference, uncertainty
except TorchDeserializationException as e:
raise APIException(e) from e
except Exception as e:
self._logger.error(e)
raise APIException("Internal Server Error occurred during inference!") from e
Classes¶
Inference¶
Inference ViewSet for performing inference on a model.
View Source
class Inference(ViewSet):
"""
Inference ViewSet for performing inference on a model.
"""
serializer_class = inline_serializer("InferenceSerializer", fields={
"inference": ListField(child=ListField(child=FloatField())),
"uncertainty": DictField(child=FloatField())
})
"""The serializer for the ViewSet."""
@extend_schema(
request=inline_serializer(
"InferenceJsonSerializer",
fields={
"model_id": CharField(),
"model_input": ListField(child=ListField(child=FloatField())),
"return_format": ChoiceField(["binary", "json"])
}
),
responses={
status.HTTP_200_OK: serializer_class,
status.HTTP_400_BAD_REQUEST: ErrorSerializer,
},
examples=[
OpenApiExample("JSON Example", value={
"model_id": "mymodel",
"model_input": [
[1.0, 2.3, -0.4, 3],
[0.01, 9.7, 5.6, 7]
],
"return_format": "json"
}, request_only=True),
]
)
def inference(self, request: HttpRequest) -> HttpResponse:
"""
Performs inference on the provided model and input data.
This method takes in an HTTP request containing the necessary metadata and input data,
performs any required preprocessing on the input data, runs the inference using the specified model,
and returns a response in the format specified by the `return_format` parameter including
possible uncertainty measurements if defined.
Args:
request (HttpRequest): The current HTTP request.
Returns:
HttpResponse: A HttpResponse containing the result of the inference as well as its uncertainty.
"""
request_body, is_json = self._get_handle_content_type(request)
model, preprocessing, input_shape, return_format = self._get_inference_metadata(
request_body,
"json" if is_json else "binary"
)
model_input = self._get_model_input(request, request_body)
if preprocessing:
model_input = preprocessing(model_input)
else:
# if no preprocessing is defined, at least try to convert/interpret the model_input as
# PyTorch tensor, before raising an exception
model_input = self._try_cast_model_input_to_tensor(model_input)
self._validate_model_input_after_preprocessing(model_input, input_shape, bool(preprocessing))
uncertainty_cls, inference, uncertainty = self._do_inference(model, model_input)
return self._make_response(uncertainty_cls, inference, uncertainty, return_format)
def _get_handle_content_type(self, request: HttpRequest) -> Tuple[dict, bool]:
"""
Handles HTTP request body based on their content type.
This function checks if the request content type is either `application/json`
or `multipart/form-data`. If it matches, it returns the corresponding data and
a boolean indicating whether it's JSON (True) or multipart/form-data (False).
Args:
request (HttpRequest): The request.
Returns:
tuple: A tuple containing the parsed data and a boolean indicating the content type.
* If content type is `application/json`, returns the JSON payload as a Python object (dict)
and True to indicate it's JSON.
* If content type is `multipart/form-data`, returns the request POST data and False.
Raises:
UnsupportedMediaType: If an unknown content type is specified, raising an error with
details on supported types (`application/json` and `multipart/form-data`).
"""
match request.content_type.lower():
case s if s.startswith("multipart/form-data"):
return request.POST, False
case s if s.startswith("application/json"):
return json.loads(request.body), True
# if the content type is specified, but not supported, return 415
self._logger.error(f"Unknown Content-Type '{request.content_type}'")
raise UnsupportedMediaType(
"Only Content-Type 'application/json' and 'multipart/form-data' is supported."
)
def _get_inference_metadata(
self,
request_body: dict,
return_format_default: Literal["binary", "json"]
) -> Tuple[Model, Optional[torch.nn.Module], Optional[List[Optional[int]]], str]:
"""
Retrieves inference metadata based on the content of the provided request body.
This method checks if a `model_id` is present in the request body and retrieves
the corresponding model entity. It then determines the return format based on the
request body or default to one of the two supported formats (`binary` or `json`).
Args:
request_body (dict): The data sent with the request, containing at least `model_id`.
return_format_default (Literal["binary", "json"]): The default return format to use if not specified in
the request body.
Returns:
Tuple[Model, Optional[torch.nn.Module], Optional[List[Optional[int]]], str]: A tuple containing:
* The retrieved model entity.
* The global model's preprocessing torch module (if applicable).
* The input shape of the global model (if applicable).
* The return format (`binary` or `json`).
Raises:
ValidationError: If no valid `model_id` is provided in the request body, or if an unknown return format
is specified.
"""
if "model_id" not in request_body:
self._logger.error("No 'model_id' provided in request.")
raise ValidationError("No 'model_id' provided in request.")
model_id = request_body["model_id"]
model = get_entity(Model, pk=model_id)
return_format = request_body.get("return_format", return_format_default)
if return_format not in ["binary", "json"]:
self._logger.error(f"Unknown return format '{return_format}'. Supported are binary and json.")
raise ValidationError(f"Unknown return format '{return_format}'. Supported are binary and json.")
global_model: Optional[GlobalModel] = None
if isinstance(model, GlobalModel):
global_model = model
elif isinstance(model, LocalModel):
global_model = model.base_model
else:
self._logger.error("Unknown model type. Not a GlobalModel and not a LocalModel. Skip preprocessing.")
preprocessing: Optional[torch.nn.Module] = None
input_shape: Optional[List[Optional[int]]] = None
if global_model:
if global_model.preprocessing is not None:
preprocessing = global_model.get_preprocessing_torch_model()
if global_model.input_shape is not None:
input_shape = global_model.input_shape
return model, preprocessing, input_shape, return_format
def _get_model_input(self, request: HttpRequest, request_body: dict) -> Any:
"""
Retrieves and decodes the model input from either an uploaded file or the request body.
Args:
request (HttpRequest): The current HTTP request.
request_body (dict): The parsed request body as a dictionary.
Returns:
Any: The decoded model input data.
Raises:
ValidationError: If no `model_input` is found in the uploaded file or the request body.
"""
uploaded_file = request.FILES.get("model_input", None)
if uploaded_file and uploaded_file.file:
model_input = uploaded_file.file.read()
else:
model_input = request_body.get("model_input", None)
if not model_input:
raise ValidationError("No uploaded file 'model_input' found.")
return self._try_decode_model_input(model_input)
def _try_decode_model_input(self, model_input: Any) -> Any:
"""
Attempts to decode the input `model_input` from various formats and returns it in a usable form.
This function first tries to deserialize the input as a PyTorch tensor. If that fails, it attempts to
decode the input as a base64-encoded string. If neither attempt is successful, the original input is returned.
Args:
model_input (Any): The input to be decoded, which can be in any format.
Returns:
Any: The decoded input, which may still be in an unknown format if decoding attempts fail.
"""
# 1. try to deserialize model_input as PyTorch tensor
try:
with disable_logger(self._logger):
model_input = to_torch_tensor(model_input)
except Exception:
pass
# 2. try to decode model_input as base64
try:
is_base64, tmp_model_input = self._is_base64(model_input)
if is_base64:
model_input = tmp_model_input
except Exception:
pass
# result
return model_input
def _try_cast_model_input_to_tensor(self, model_input: Any) -> Any:
"""
Attempt to cast the given model input to a PyTorch tensor.
This function tries to interpret the input in several formats:
1. PIL Image (and later convert it to a PyTorch tensor, see 3.)
2. PyTorch tensor via `torch.as_tensor`
3. PyTorch tensor via torchvision `ToTensor` (supports e.g. PIL images)
If none of these attempts are successful, the original input is returned.
Args:
model_input: The input data to be cast to a PyTorch tensor.
Can be any type that can be converted to a tensor.
Returns:
A PyTorch tensor representation of the input data, or the original
input if it cannot be converted.
"""
def _try_to_pil_image(model_input: Any) -> Any:
stream = BytesIO(model_input)
return Image.open(stream)
if isinstance(model_input, torch.Tensor):
return model_input
# In the following order, try to:
# 1. interpret model_input as PIL image (and later to PyTorch tensor, see step 3),
# 2. interpret model_input as PyTorch tensor,
# 3. interpret model_input as PyTorch tensor via torchvision ToTensor (supports e.g. PIL images).
for fn in [_try_to_pil_image, torch.as_tensor, to_tensor]:
try:
model_input = fn(model_input) # type: ignore
except Exception:
pass
return model_input
def _is_base64(self, sb: str | bytes) -> Tuple[bool, bytes]:
"""
Check if a string or bytes object is a valid Base64 encoded string.
This function checks if the input can be decoded and re-encoded without any changes.
If decoding and encoding returns the same result as the original input, it's likely
that the input was indeed a valid Base64 encoded string.
Note: This code is based on the reference implementation from the linked Stack Overflow answer.
Args:
sb (str | bytes): The input string or bytes object to check.
Returns:
Tuple[bool, bytes]: A tuple containing a boolean indicating whether the input is
a valid Base64 encoded string and the decoded bytes if it is.
References:
https://stackoverflow.com/a/45928164
"""
try:
if isinstance(sb, str):
# If there's any unicode here, an exception will be thrown and the function will return false
sb_bytes = bytes(sb, "ascii")
elif isinstance(sb, bytes):
sb_bytes = sb
else:
raise ValueError("Argument must be string or bytes")
decoded = base64.b64decode(sb_bytes)
return base64.b64encode(decoded) == sb_bytes, decoded
except Exception:
return False, b""
def _validate_model_input_after_preprocessing(
self,
model_input: Any,
model_input_shape: Optional[List[Optional[int]]],
preprocessing: bool
) -> None:
"""
Validates the model input after preprocessing.
Ensures that the provided `model_input` is a valid PyTorch tensor and its shape matches
the expected`model_input_shape`.
Args:
model_input (Any): The model input to be validated.
model_input_shape (Optional[List[Optional[int]]]): The expected shape of the model input.
Can contain None values if not all dimensions are fixed (e.g. first dimension as batch size).
preprocessing (bool): Whether a preprocessing model was defined or not. (Only for a better error message.)
Raises:
ValidationError: If the `model_input` is not a valid PyTorch tensor or
its shape does not match the expected `model_input_shape`.
"""
if not isinstance(model_input, torch.Tensor):
msg = "Model input could not be casted or interpreted as a PyTorch tensor object"
if preprocessing:
msg += " and is still not a PyTorch tensor after preprecessing."
else:
msg += " and no preprecessing is defined."
raise ValidationError(msg)
if model_input_shape and not all(
dim_input == dim_model
for (dim_input, dim_model) in zip(model_input.shape, model_input_shape)
if dim_model is not None
):
raise ValidationError("Input shape does not match model input shape.")
def _make_response(
self,
uncertainty_cls: Type[UncertaintyBase],
inference: torch.Tensor,
uncertainty: Any,
return_type: str
) -> HttpResponse:
"""
Build the response object with the result data.
This method checks the return type and makes a response with the appropriate content type.
If return_type is "binary", a binary-encoded response will be generated using pickle.
Otherwise, a JSON response will be generated by serializing the uncertainty object using its to_json method.
Args:
uncertainty_cls (Type[UncertaintyBase]): The uncertainty class.
inference (torch.Tensor): The inference.
uncertainty (Any): The uncertainty.
return_type (str): The return type.
Returns:
HttpResponse: The inference result response.
"""
if return_type == "binary":
response_bytes = pickle.dumps(dict(inference=inference, uncertainty=uncertainty))
return HttpResponse(response_bytes, content_type="application/octet-stream")
return HttpResponse(uncertainty_cls.to_json(inference, uncertainty), content_type="application/json")
def _do_inference(
self, model: Model, input_tensor: torch.Tensor
) -> Tuple[Type[UncertaintyBase], torch.Tensor, Dict[str, Any]]:
"""
Perform inference on a given input tensor using the provided model.
This methods retrieves the uncertainty class, performs the prediction.
The output of this method consists of:
* The uncertainty class used for inference
* The result of the model's prediction on the input tensor
* Any associated uncertainty for the prediction
Args:
model (Model): The model to perform inference with.
input_tensor (torch.Tensor): Input tensor to pass through the model.
Returns:
Tuple[Type[UncertaintyBase], torch.Tensor, Dict[str, Any]]:
A tuple containing the uncertainty class, prediction result, and any associated uncertainty.
Raises:
APIException: If an error occurs during inference
"""
try:
uncertainty_cls = get_uncertainty_class(model)
inference, uncertainty = uncertainty_cls.prediction(input_tensor, model)
return uncertainty_cls, inference, uncertainty
except TorchDeserializationException as e:
raise APIException(e) from e
except Exception as e:
self._logger.error(e)
raise APIException("Internal Server Error occurred during inference!") from e
Ancestors (in MRO)¶
- fl_server_api.views.base.ViewSet
- rest_framework.viewsets.ViewSet
- rest_framework.viewsets.ViewSetMixin
- rest_framework.views.APIView
- django.views.generic.base.View
Class variables¶
The serializer for the ViewSet.
Static methods¶
as_view¶
Because of the way class based views create a closure around the
instantiated view, we need to totally reimplement .as_view,
and slightly modify the view function that is created and returned.
View Source
@classonlymethod
def as_view(cls, actions=None, **initkwargs):
"""
Because of the way class based views create a closure around the
instantiated view, we need to totally reimplement `.as_view`,
and slightly modify the view function that is created and returned.
"""
# The name and description initkwargs may be explicitly overridden for
# certain route configurations. eg, names of extra actions.
cls.name = None
cls.description = None
# The suffix initkwarg is reserved for displaying the viewset type.
# This initkwarg should have no effect if the name is provided.
# eg. 'List' or 'Instance'.
cls.suffix = None
# The detail initkwarg is reserved for introspecting the viewset type.
cls.detail = None
# Setting a basename allows a view to reverse its action urls. This
# value is provided by the router through the initkwargs.
cls.basename = None
# actions must not be empty
if not actions:
raise TypeError("The `actions` argument must be provided when "
"calling `.as_view()` on a ViewSet. For example "
"`.as_view({'get': 'list'})`")
# sanitize keyword arguments
for key in initkwargs:
if key in cls.http_method_names:
raise TypeError("You tried to pass in the %s method name as a "
"keyword argument to %s(). Don't do that."
% (key, cls.__name__))
if not hasattr(cls, key):
raise TypeError("%s() received an invalid keyword %r" % (
cls.__name__, key))
# name and suffix are mutually exclusive
if 'name' in initkwargs and 'suffix' in initkwargs:
raise TypeError("%s() received both `name` and `suffix`, which are "
"mutually exclusive arguments." % (cls.__name__))
def view(request, *args, **kwargs):
self = cls(**initkwargs)
if 'get' in actions and 'head' not in actions:
actions['head'] = actions['get']
# We also store the mapping of request methods to actions,
# so that we can later set the action attribute.
# eg. `self.action = 'list'` on an incoming GET request.
self.action_map = actions
# Bind methods to actions
# This is the bit that's different to a standard view
for method, action in actions.items():
handler = getattr(self, action)
setattr(self, method, handler)
self.request = request
self.args = args
self.kwargs = kwargs
# And continue as usual
return self.dispatch(request, *args, **kwargs)
# take name and docstring from class
update_wrapper(view, cls, updated=())
# and possible attributes set by decorators
# like csrf_exempt from dispatch
update_wrapper(view, cls.dispatch, assigned=())
# We need to set these on the view function, so that breadcrumb
# generation can pick out these bits of information from a
# resolved URL.
view.cls = cls
view.initkwargs = initkwargs
view.actions = actions
return csrf_exempt(view)
get_extra_actions¶
Get the methods that are marked as an extra ViewSet @action.
View Source
Instance variables¶
Wrap Django's private _allowed_methods interface in a public property.
Methods¶
check_object_permissions¶
Check if the request should be permitted for a given object.
Raises an appropriate exception if the request is not permitted.
View Source
def check_object_permissions(self, request, obj):
"""
Check if the request should be permitted for a given object.
Raises an appropriate exception if the request is not permitted.
"""
for permission in self.get_permissions():
if not permission.has_object_permission(request, self, obj):
self.permission_denied(
request,
message=getattr(permission, 'message', None),
code=getattr(permission, 'code', None)
)
check_permissions¶
Check if the request should be permitted.
Raises an appropriate exception if the request is not permitted.
View Source
def check_permissions(self, request):
"""
Check if the request should be permitted.
Raises an appropriate exception if the request is not permitted.
"""
for permission in self.get_permissions():
if not permission.has_permission(request, self):
self.permission_denied(
request,
message=getattr(permission, 'message', None),
code=getattr(permission, 'code', None)
)
check_throttles¶
Check if request should be throttled.
Raises an appropriate exception if the request is throttled.
View Source
def check_throttles(self, request):
"""
Check if request should be throttled.
Raises an appropriate exception if the request is throttled.
"""
throttle_durations = []
for throttle in self.get_throttles():
if not throttle.allow_request(request, self):
throttle_durations.append(throttle.wait())
if throttle_durations:
# Filter out `None` values which may happen in case of config / rate
# changes, see #1438
durations = [
duration for duration in throttle_durations
if duration is not None
]
duration = max(durations, default=None)
self.throttled(request, duration)
determine_version¶
If versioning is being used, then determine any API version for the
incoming request. Returns a two-tuple of (version, versioning_scheme)
View Source
def determine_version(self, request, *args, **kwargs):
"""
If versioning is being used, then determine any API version for the
incoming request. Returns a two-tuple of (version, versioning_scheme)
"""
if self.versioning_class is None:
return (None, None)
scheme = self.versioning_class()
return (scheme.determine_version(request, *args, **kwargs), scheme)
dispatch¶
.dispatch() is pretty much the same as Django's regular dispatch,
but with extra hooks for startup, finalize, and exception handling.
View Source
def dispatch(self, request, *args, **kwargs):
"""
`.dispatch()` is pretty much the same as Django's regular dispatch,
but with extra hooks for startup, finalize, and exception handling.
"""
self.args = args
self.kwargs = kwargs
request = self.initialize_request(request, *args, **kwargs)
self.request = request
self.headers = self.default_response_headers # deprecate?
try:
self.initial(request, *args, **kwargs)
# Get the appropriate handler method
if request.method.lower() in self.http_method_names:
handler = getattr(self, request.method.lower(),
self.http_method_not_allowed)
else:
handler = self.http_method_not_allowed
response = handler(request, *args, **kwargs)
except Exception as exc:
response = self.handle_exception(exc)
self.response = self.finalize_response(request, response, *args, **kwargs)
return self.response
finalize_response¶
Returns the final response object.
View Source
def finalize_response(self, request, response, *args, **kwargs):
"""
Returns the final response object.
"""
# Make the error obvious if a proper response is not returned
assert isinstance(response, HttpResponseBase), (
'Expected a `Response`, `HttpResponse` or `HttpStreamingResponse` '
'to be returned from the view, but received a `%s`'
% type(response)
)
if isinstance(response, Response):
if not getattr(request, 'accepted_renderer', None):
neg = self.perform_content_negotiation(request, force=True)
request.accepted_renderer, request.accepted_media_type = neg
response.accepted_renderer = request.accepted_renderer
response.accepted_media_type = request.accepted_media_type
response.renderer_context = self.get_renderer_context()
# Add new vary headers to the response instead of overwriting.
vary_headers = self.headers.pop('Vary', None)
if vary_headers is not None:
patch_vary_headers(response, cc_delim_re.split(vary_headers))
for key, value in self.headers.items():
response[key] = value
return response
get_authenticate_header¶
If a request is unauthenticated, determine the WWW-Authenticate
header to use for 401 responses, if any.
View Source
get_authenticators¶
Get the authenticators for the ViewSet.
This method gets the view method and, if it has authentication classes defined via the decorator, returns them. Otherwise, it falls back to the default authenticators.
Returns:
| Type | Description |
|---|---|
| list | The authenticators for the ViewSet. |
View Source
def get_authenticators(self):
"""
Get the authenticators for the ViewSet.
This method gets the view method and, if it has authentication classes defined via the decorator, returns them.
Otherwise, it falls back to the default authenticators.
Returns:
list: The authenticators for the ViewSet.
"""
if method := self._get_view_method():
if hasattr(method, "authentication_classes"):
return method.authentication_classes
return super().get_authenticators()
get_content_negotiator¶
Instantiate and return the content negotiation class to use.
View Source
get_exception_handler¶
Returns the exception handler that this view uses.
View Source
get_exception_handler_context¶
Returns a dict that is passed through to EXCEPTION_HANDLER,
as the context argument.
View Source
get_extra_action_url_map¶
Build a map of {names: urls} for the extra actions.
This method will noop if detail was not provided as a view initkwarg.
View Source
def get_extra_action_url_map(self):
"""
Build a map of {names: urls} for the extra actions.
This method will noop if `detail` was not provided as a view initkwarg.
"""
action_urls = OrderedDict()
# exit early if `detail` has not been provided
if self.detail is None:
return action_urls
# filter for the relevant extra actions
actions = [
action for action in self.get_extra_actions()
if action.detail == self.detail
]
for action in actions:
try:
url_name = '%s-%s' % (self.basename, action.url_name)
namespace = self.request.resolver_match.namespace
if namespace:
url_name = '%s:%s' % (namespace, url_name)
url = reverse(url_name, self.args, self.kwargs, request=self.request)
view = self.__class__(**action.kwargs)
action_urls[view.get_view_name()] = url
except NoReverseMatch:
pass # URL requires additional arguments, ignore
return action_urls
get_format_suffix¶
Determine if the request includes a '.json' style format suffix
View Source
get_parser_context¶
Returns a dict that is passed through to Parser.parse(),
as the parser_context keyword argument.
View Source
def get_parser_context(self, http_request):
"""
Returns a dict that is passed through to Parser.parse(),
as the `parser_context` keyword argument.
"""
# Note: Additionally `request` and `encoding` will also be added
# to the context by the Request object.
return {
'view': self,
'args': getattr(self, 'args', ()),
'kwargs': getattr(self, 'kwargs', {})
}
get_parsers¶
Instantiates and returns the list of parsers that this view can use.
View Source
get_permissions¶
Get the permissions for the ViewSet.
This method gets the view method and, if it has permission classes defined via the decorator, returns them. Otherwise, it falls back to the default permissions.
Returns:
| Type | Description |
|---|---|
| list | The permissions for the ViewSet. |
View Source
def get_permissions(self):
"""
Get the permissions for the ViewSet.
This method gets the view method and, if it has permission classes defined via the decorator, returns them.
Otherwise, it falls back to the default permissions.
Returns:
list: The permissions for the ViewSet.
"""
if method := self._get_view_method():
if hasattr(method, "permission_classes"):
return method.permission_classes
return super().get_permissions()
get_renderer_context¶
Returns a dict that is passed through to Renderer.render(),
as the renderer_context keyword argument.
View Source
def get_renderer_context(self):
"""
Returns a dict that is passed through to Renderer.render(),
as the `renderer_context` keyword argument.
"""
# Note: Additionally 'response' will also be added to the context,
# by the Response object.
return {
'view': self,
'args': getattr(self, 'args', ()),
'kwargs': getattr(self, 'kwargs', {}),
'request': getattr(self, 'request', None)
}
get_renderers¶
Instantiates and returns the list of renderers that this view can use.
View Source
get_throttles¶
Instantiates and returns the list of throttles that this view uses.
View Source
get_view_description¶
Return some descriptive text for the view, as used in OPTIONS responses
and in the browsable API.
View Source
get_view_name¶
Return the view name, as used in OPTIONS responses and in the
browsable API.
View Source
handle_exception¶
Handle any exception that occurs, by returning an appropriate response,
or re-raising the error.
View Source
def handle_exception(self, exc):
"""
Handle any exception that occurs, by returning an appropriate response,
or re-raising the error.
"""
if isinstance(exc, (exceptions.NotAuthenticated,
exceptions.AuthenticationFailed)):
# WWW-Authenticate header for 401 responses, else coerce to 403
auth_header = self.get_authenticate_header(self.request)
if auth_header:
exc.auth_header = auth_header
else:
exc.status_code = status.HTTP_403_FORBIDDEN
exception_handler = self.get_exception_handler()
context = self.get_exception_handler_context()
response = exception_handler(exc, context)
if response is None:
self.raise_uncaught_exception(exc)
response.exception = True
return response
http_method_not_allowed¶
If request.method does not correspond to a handler method,
determine what kind of exception to raise.
View Source
inference¶
def inference(
self,
request: django.http.request.HttpRequest
) -> django.http.response.HttpResponse
Performs inference on the provided model and input data.
This method takes in an HTTP request containing the necessary metadata and input data,
performs any required preprocessing on the input data, runs the inference using the specified model,
and returns a response in the format specified by the return_format parameter including
possible uncertainty measurements if defined.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
| request | HttpRequest | The current HTTP request. | None |
Returns:
| Type | Description |
|---|---|
| HttpResponse | A HttpResponse containing the result of the inference as well as its uncertainty. |
View Source
@extend_schema(
request=inline_serializer(
"InferenceJsonSerializer",
fields={
"model_id": CharField(),
"model_input": ListField(child=ListField(child=FloatField())),
"return_format": ChoiceField(["binary", "json"])
}
),
responses={
status.HTTP_200_OK: serializer_class,
status.HTTP_400_BAD_REQUEST: ErrorSerializer,
},
examples=[
OpenApiExample("JSON Example", value={
"model_id": "mymodel",
"model_input": [
[1.0, 2.3, -0.4, 3],
[0.01, 9.7, 5.6, 7]
],
"return_format": "json"
}, request_only=True),
]
)
def inference(self, request: HttpRequest) -> HttpResponse:
"""
Performs inference on the provided model and input data.
This method takes in an HTTP request containing the necessary metadata and input data,
performs any required preprocessing on the input data, runs the inference using the specified model,
and returns a response in the format specified by the `return_format` parameter including
possible uncertainty measurements if defined.
Args:
request (HttpRequest): The current HTTP request.
Returns:
HttpResponse: A HttpResponse containing the result of the inference as well as its uncertainty.
"""
request_body, is_json = self._get_handle_content_type(request)
model, preprocessing, input_shape, return_format = self._get_inference_metadata(
request_body,
"json" if is_json else "binary"
)
model_input = self._get_model_input(request, request_body)
if preprocessing:
model_input = preprocessing(model_input)
else:
# if no preprocessing is defined, at least try to convert/interpret the model_input as
# PyTorch tensor, before raising an exception
model_input = self._try_cast_model_input_to_tensor(model_input)
self._validate_model_input_after_preprocessing(model_input, input_shape, bool(preprocessing))
uncertainty_cls, inference, uncertainty = self._do_inference(model, model_input)
return self._make_response(uncertainty_cls, inference, uncertainty, return_format)
initial¶
Runs anything that needs to occur prior to calling the method handler.
View Source
def initial(self, request, *args, **kwargs):
"""
Runs anything that needs to occur prior to calling the method handler.
"""
self.format_kwarg = self.get_format_suffix(**kwargs)
# Perform content negotiation and store the accepted info on the request
neg = self.perform_content_negotiation(request)
request.accepted_renderer, request.accepted_media_type = neg
# Determine the API version, if versioning is in use.
version, scheme = self.determine_version(request, *args, **kwargs)
request.version, request.versioning_scheme = version, scheme
# Ensure that the incoming request is permitted
self.perform_authentication(request)
self.check_permissions(request)
self.check_throttles(request)
initialize_request¶
Set the .action attribute on the view, depending on the request method.
View Source
def initialize_request(self, request, *args, **kwargs):
"""
Set the `.action` attribute on the view, depending on the request method.
"""
request = super().initialize_request(request, *args, **kwargs)
method = request.method.lower()
if method == 'options':
# This is a special case as we always provide handling for the
# options method in the base `View` class.
# Unlike the other explicitly defined actions, 'metadata' is implicit.
self.action = 'metadata'
else:
self.action = self.action_map.get(method)
return request
options¶
Handler method for HTTP 'OPTIONS' request.
View Source
def options(self, request, *args, **kwargs):
"""
Handler method for HTTP 'OPTIONS' request.
"""
if self.metadata_class is None:
return self.http_method_not_allowed(request, *args, **kwargs)
data = self.metadata_class().determine_metadata(request, self)
return Response(data, status=status.HTTP_200_OK)
perform_authentication¶
Perform authentication on the incoming request.
Note that if you override this and simply 'pass', then authentication
will instead be performed lazily, the first time either
request.user or request.auth is accessed.
View Source
perform_content_negotiation¶
Determine which renderer and media type to use render the response.
View Source
def perform_content_negotiation(self, request, force=False):
"""
Determine which renderer and media type to use render the response.
"""
renderers = self.get_renderers()
conneg = self.get_content_negotiator()
try:
return conneg.select_renderer(request, renderers, self.format_kwarg)
except Exception:
if force:
return (renderers[0], renderers[0].media_type)
raise
permission_denied¶
If request is not permitted, determine what kind of exception to raise.
View Source
def permission_denied(self, request, message=None, code=None):
"""
If request is not permitted, determine what kind of exception to raise.
"""
if request.authenticators and not request.successful_authenticator:
raise exceptions.NotAuthenticated()
raise exceptions.PermissionDenied(detail=message, code=code)
raise_uncaught_exception¶
View Source
reverse_action¶
Reverse the action for the given url_name.
View Source
def reverse_action(self, url_name, *args, **kwargs):
"""
Reverse the action for the given `url_name`.
"""
url_name = '%s-%s' % (self.basename, url_name)
namespace = None
if self.request and self.request.resolver_match:
namespace = self.request.resolver_match.namespace
if namespace:
url_name = namespace + ':' + url_name
kwargs.setdefault('request', self.request)
return reverse(url_name, *args, **kwargs)
setup¶
Initialize attributes shared by all view methods.
View Source
throttled¶
If request is throttled, determine what kind of exception to raise.