xmos-ai-tools 1.3.2.dev180__py3-none-macosx_10_15_universal2.whl

xmos_ai_tools/xinterpreters/host_interpreter.py

@@ -0,0 +1,651 @@
+# Copyright 2022 XMOS LIMITED. This Software is subject to the terms of the
+# XMOS Public License: Version 1
+import sys
+import ctypes
+from typing import Optional, Dict, Any, List
+from tflite.Model import Model
+from tflite.TensorType import TensorType
+from tflite import opcode2name
+from enum import Enum
+
+import numpy as np
+from pathlib import Path
+
+from numpy import ndarray
+
+# DLL path for different platforms
+__PARENT_DIR = Path(__file__).parent.absolute()
+if sys.platform.startswith("linux"):
+    lib_path = str(Path.joinpath(__PARENT_DIR, "libs", "linux", "xtflm_python.so"))
+elif sys.platform == "darwin":
+    lib_path = str(Path.joinpath(__PARENT_DIR, "libs", "macos", "xtflm_python.dylib"))
+else:
+    lib_path = str(Path.joinpath(__PARENT_DIR, "libs", "windows", "xtflm_python.dll"))
+
+lib = ctypes.cdll.LoadLibrary(lib_path)
+
+from xmos_ai_tools.xinterpreters.exceptions import (
+    InterpreterError,
+    AllocateTensorsError,
+    InvokeError,
+    SetTensorError,
+    GetTensorError,
+    ModelSizeError,
+    ArenaSizeError,
+    DeviceTimeoutError,
+)
+
+MAX_TENSOR_ARENA_SIZE = 10000000
+
+
+class XTFLMInterpreterStatus(Enum):
+    OK = 0
+    ERROR = 1
+
+
+class TFLMHostInterpreter:
+    """! The xcore interpreters host class.
+    The interpreter to be used on a host.
+    """
+
+    def __init__(self, max_tensor_arena_size: int = MAX_TENSOR_ARENA_SIZE) -> None:
+        """! Host interpreter initializer.
+        Sets up functions from the cdll, and calls to cdll function to create a new interpreter.
+        """
+        self._error_msg = ctypes.create_string_buffer(4096)
+
+        lib.new_interpreter.restype = ctypes.c_void_p
+        lib.new_interpreter.argtypes = [
+            ctypes.c_size_t,
+        ]
+
+        lib.print_memory_plan.restype = None
+        lib.print_memory_plan.argtypes = [ctypes.c_void_p]
+
+        lib.delete_interpreter.restype = None
+        lib.delete_interpreter.argtypes = [ctypes.c_void_p]
+
+        lib.initialize.restype = ctypes.c_int
+        lib.initialize.argtypes = [
+            ctypes.c_void_p,
+            ctypes.c_char_p,
+            ctypes.c_size_t,
+            ctypes.c_char_p,
+        ]
+
+        lib.set_input_tensor.restype = ctypes.c_int
+        lib.set_input_tensor.argtypes = [
+            ctypes.c_void_p,
+            ctypes.c_size_t,
+            ctypes.c_void_p,
+            ctypes.c_int,
+        ]
+
+        lib.get_output_tensor.restype = ctypes.c_int
+        lib.get_output_tensor.argtypes = [
+            ctypes.c_void_p,
+            ctypes.c_size_t,
+            ctypes.c_void_p,
+            ctypes.c_int,
+        ]
+
+        lib.get_input_tensor.restype = ctypes.c_int
+        lib.get_input_tensor.argtypes = [
+            ctypes.c_void_p,
+            ctypes.c_size_t,
+            ctypes.c_void_p,
+            ctypes.c_int,
+        ]
+
+        lib.reset.restype = ctypes.c_int
+        lib.reset.argtypes = [ctypes.c_void_p]
+
+        lib.invoke.restype = ctypes.c_int
+        lib.invoke.argtypes = [ctypes.c_void_p]
+
+        lib.get_error.restype = ctypes.c_size_t
+        lib.get_error.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
+
+        lib.arena_used_bytes.restype = ctypes.c_size_t
+        lib.arena_used_bytes.argtypes = [
+            ctypes.c_void_p,
+        ]
+
+        self._max_tensor_arena_size = max_tensor_arena_size
+        self.models: List[TFLMHostInterpreter.modelData] = []
+
+    def __enter__(self) -> "TFLMHostInterpreter":
+        return self
+
+    def __exit__(self, exc_type, exc_value, exc_traceback) -> None:
+        """! Exit calls close function to delete interpreter"""
+        self.close()
+
+    def initialise_interpreter(self, model_index: int = 0) -> None:
+        """! Interpreter initialiser, initialised interpreter with model and parameters (optional)
+        @param model_index  The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        """
+        max_model_size = 50000000
+        self.obj = lib.new_interpreter(max_model_size)
+        currentModel = None
+
+        for model in self.models:
+            if model.tile == model_index:
+                currentModel = model
+
+        if currentModel is None:
+            print(f"No model at index {model_index} found.", sys.stderr)
+            raise IndexError
+
+        assert currentModel.model_content is not None
+
+        status = lib.initialize(
+            self.obj,
+            currentModel.model_content,
+            len(currentModel.model_content),
+            currentModel.params_content,
+        )
+        if XTFLMInterpreterStatus(status) is XTFLMInterpreterStatus.ERROR:
+            raise RuntimeError("Unable to initialize interpreter")
+
+    def set_tensor(self, tensor_index: int, value: ndarray, model_index=0) -> None:
+        """! Write the input tensor of a model.
+        @param value  The blob of data to set the tensor to.
+        @param tensor_index  The index of input tensor to target. Defaults to 0.
+        @param model_index  The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        """
+        val = value.tobytes()
+
+        length = len(val)
+        length2 = self.get_input_tensor_size(tensor_index)
+        if length != length2:
+            print(
+                "ERROR: mismatching size in set_input_tensor %d vs %d"
+                % (length, length2)
+            )
+
+        self._check_status(lib.set_input_tensor(self.obj, tensor_index, val, length))
+
+    def get_tensor(
+        self, tensor_index: int = 0, model_index: int = 0, tensor: ndarray = None
+    ) -> ndarray:
+        """! Read data from the output tensor of a model.
+        @param tensor_index  The index of output tensor to target.
+        @param model_index  The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @param tensor  Tensor of correct shape to write into (optional).
+        @return  The data that was stored in the output tensor.
+        """
+
+        count: Optional[int]
+        tensor_details: Optional[Dict[str, Any]]
+        count, tensor_details = next(
+            filter(
+                lambda x: x[1]["index"] == tensor_index,
+                enumerate(self.get_output_details()),
+            ),
+            (None, None),
+        )
+
+        if count is None or tensor_details is None:
+            print(f"No tensor at index {tensor_index} found.", sys.stderr)
+            raise IndexError
+
+        length = self.get_tensor_size(tensor_index)
+        if tensor is None:
+            tensor = np.zeros(tensor_details["shape"], dtype=tensor_details["dtype"])
+        else:
+            length = len(tensor.tobytes())
+            if length != length:
+                print(
+                    "ERROR: mismatching size in get_output_tensor %d vs %d"
+                    % (length, length)
+                )
+
+        data_ptr = tensor.ctypes.data_as(ctypes.c_void_p)
+        self._check_status(lib.get_output_tensor(self.obj, count, data_ptr, length))
+        return tensor
+
+    def get_input_tensor(self, input_index: int = 0, model_index: int = 0) -> ndarray:
+        """! Read the data in the input tensor of a model.
+        @param input_index  The index of input tensor to target.
+        @param model_index The engine to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @return The data that was stored in the output tensor.
+        """
+        tensor_details = self.get_input_details(model_index)[input_index]
+        tensor = np.zeros(tensor_details["shape"], dtype=tensor_details["dtype"])
+        data_ptr = tensor.ctypes.data_as(ctypes.c_void_p)
+
+        l = len(tensor.tobytes())
+        self._check_status(lib.get_input_tensor(self.obj, input_index, data_ptr, l))
+        return tensor
+
+    def reset(self, model_index: int = 0) -> None:
+        """! Resets the model."""
+        self._check_status(lib.reset(self.obj))
+
+    def invoke(self, model_index: int = 0) -> None:
+        """! Invoke the model and starting inference of the current
+        state of the tensors.
+        """
+        INVOKE_CALLBACK_FUNC = ctypes.CFUNCTYPE(ctypes.c_void_p, ctypes.c_int)
+
+        self._check_status(lib.invoke(self.obj))
+
+    def close(self, model_index: int = 0) -> None:
+        """! Delete the interpreter.
+        @params model_index Defines which interpreter to target in systems with multiple.
+        """
+        if self.obj:
+            lib.delete_interpreter(self.obj)
+            self.obj = None
+
+    def tensor_arena_size(self) -> int:
+        """! Read the size of the tensor arena required.
+        @return size of the tensor arena as an integer.
+        """
+        return lib.arena_used_bytes(self.obj)
+
+    def _check_status(self, status) -> None:
+        """! Read a status code and raise an exception.
+        @param status Status code.
+        """
+        if XTFLMInterpreterStatus(status) is XTFLMInterpreterStatus.ERROR:
+            lib.get_error(self.obj, self._error_msg)
+            raise RuntimeError(self._error_msg.value.decode("utf-8"))
+
+    def print_memory_plan(self) -> None:
+        """! Print a plan of memory allocation"""
+        lib.print_memory_plan(self.obj)
+
+    def allocate_tensors(self):
+        """! Dummy function to match tf.lite.Interpreter() API"""
+        return
+
+    def get_input_tensor_size(self, input_index: int = 0, model_index: int = 0) -> int:
+        """! Read the size of the input tensor from the model.
+        @param input_index  The index of input tensor to target.
+        @param model_index The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @return The size of the input tensor as an integer.
+        """
+
+        # Select correct model from model list
+        model = self.get_model(model_index)
+        modelBuf = Model.GetRootAsModel(model.model_content, 0)
+
+        # Get index of specific input tensor
+        tensorIndex = modelBuf.Subgraphs(0).Inputs(input_index)
+
+        tensorType = modelBuf.Subgraphs(0).Tensors(tensorIndex).Type()
+
+        tensorSize: int
+        if tensorType == TensorType.INT8:
+            tensorSize = 1  # int8 is 1 byte
+        elif tensorType == TensorType.INT16:
+            tensorSize = 2  # int16 is 2 bytes
+        elif tensorType == TensorType.INT32:
+            tensorSize = 4  # int32 is 4 bytes
+        elif tensorType == TensorType.FLOAT32:
+            tensorSize = 4  # float32 is 4 bytes
+        else:
+            print(tensorType)
+            self._check_status(XTFLMInterpreterStatus.ERROR)
+            tensorSize = 0
+
+        # Calculate tensor size by multiplying shape elements
+        for i in range(0, modelBuf.Subgraphs(0).Tensors(tensorIndex).ShapeLength()):
+            tensorSize = tensorSize * modelBuf.Subgraphs(0).Tensors(tensorIndex).Shape(
+                i
+            )
+        return tensorSize
+
+    def get_output_tensor_size(
+        self, output_index: int = 0, model_index: int = 0
+    ) -> int:
+        """! Read the size of the output tensor from the model.
+        @param output_index  The index of output tensor to target.
+        @param model_index The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @return The size of the output tensor as an integer.
+        """
+
+        # Select correct model from model list
+        modelBuf = None
+        model = self.get_model(model_index)
+        modelBuf = Model.GetRootAsModel(model.model_content, 0)
+
+        # Get index of specific output tensor
+        tensorIndex = modelBuf.Subgraphs(0).Outputs(output_index)
+
+        tensorType = modelBuf.Subgraphs(0).Tensors(tensorIndex).Type()
+
+        tensorSize: int
+        if tensorType == TensorType.INT8:
+            tensorSize = 1  # int8 is 1 byte
+        elif tensorType == TensorType.INT16:
+            tensorSize = 2  # int16 is 2 bytes
+        elif tensorType == TensorType.INT32:
+            tensorSize = 4  # int32 is 4 bytes
+        elif tensorType == TensorType.FLOAT32:
+            tensorSize = 4  # float32 is 4 bytes
+        else:
+            print(tensorType)
+            self._check_status(XTFLMInterpreterStatus.ERROR)
+            tensorSize = 0
+
+        # Calculate tensor size by multiplying shape elements
+        for i in range(0, modelBuf.Subgraphs(0).Tensors(tensorIndex).ShapeLength()):
+            tensorSize = tensorSize * modelBuf.Subgraphs(0).Tensors(tensorIndex).Shape(
+                i
+            )
+        return tensorSize
+
+    def get_tensor_size(self, tensor_index: int = 0, model_index: int = 0) -> int:
+        """! Read the size of the input tensor from the model.
+        @param tensor_index  The index of input tensor to target.
+        @param model_index The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @return The size of the input tensor as an integer.
+        """
+
+        # Select correct model from model list
+        modelBuf = None
+        model = self.get_model(model_index)
+        modelBuf = Model.GetRootAsModel(model.model_content, 0)
+
+        tensorType = modelBuf.Subgraphs(0).Tensors(tensor_index).Type()
+        if tensorType == TensorType.INT8:
+            tensorSize = 1  # int8 is 1 byte
+        elif tensorType == TensorType.INT16:
+            tensorSize = 2  # int16 is 2 bytes
+        elif tensorType == TensorType.INT32:
+            tensorSize = 4  # int32 is 4 bytes
+        elif tensorType == TensorType.FLOAT32:
+            tensorSize = 4  # float32 is 4 bytes
+        else:
+            print(tensorType)
+            self._check_status(XTFLMInterpreterStatus.ERROR)
+
+        # Calculate tensor size by multiplying shape elements
+        for i in range(0, modelBuf.Subgraphs(0).Tensors(tensor_index).ShapeLength()):
+            tensorSize = tensorSize * modelBuf.Subgraphs(0).Tensors(tensor_index).Shape(
+                i
+            )
+        return tensorSize
+
+    def get_input_details(self, model_index: int = 0) -> List[Dict[str, Any]]:
+        """! Reads the input tensor details from the model.
+        @param model_index The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @return Tensor details, including the index, name, shape, data type, and quantization
+        parameters.
+        """
+
+        # Select correct model from model list
+        modelBuf = None
+        model = self.get_model(model_index)
+        modelBuf = Model.GetRootAsModel(model.model_content, 0)
+
+        inputsList = []
+        for input_ in range(0, modelBuf.Subgraphs(0).InputsLength()):
+            tensorIndex = modelBuf.Subgraphs(0).Inputs(input_)
+
+            # Generate dictioary of tensor details
+            dtype: Union[Type[Any]]
+            if modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.INT8:
+                dtype = np.int8
+            elif modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.INT16:
+                dtype = np.int16
+            elif modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.INT32:
+                dtype = np.int32
+            elif (
+                modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.FLOAT32
+            ):
+                dtype = np.float32
+            else:
+                raise TypeError
+
+            details = {
+                "name": str(modelBuf.Subgraphs(0).Tensors(tensorIndex).Name())[
+                    1:
+                ].strip("'"),
+                "index": tensorIndex,
+                "shape": modelBuf.Subgraphs(0).Tensors(tensorIndex).ShapeAsNumpy(),
+                "shape_signature": modelBuf.Subgraphs(0)
+                .Tensors(tensorIndex)
+                .ShapeSignatureAsNumpy(),
+                "dtype": dtype,
+                "quantization": (
+                    modelBuf.Subgraphs(0).Tensors(tensorIndex).Quantization().Scale(0),
+                    modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .ZeroPoint(0),
+                ),
+                "quantization_parameters": {
+                    "scales": modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .ScaleAsNumpy(),
+                    "zero_points": modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .ZeroPointAsNumpy(),
+                    "quantized_dimension": modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .QuantizedDimension(),
+                },
+                "sparsity_parameters": {
+                    modelBuf.Subgraphs(0).Tensors(tensorIndex).Sparsity()
+                },
+            }
+            inputsList.append(details)
+
+        return inputsList
+
+    def get_output_details(self, model_index: int = 0) -> List[Dict[str, Any]]:
+        """! Reads the output tensor details from the model.
+        @param output_index  The index of output tensor to target.
+        @param model_index The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        @return Tensor details, including the index, name, shape, data type, and quantization
+        parameters.
+        """
+
+        # Select correct model from models list
+        model = self.get_model(model_index)
+        modelBuf = Model.GetRootAsModel(model.model_content, 0)
+
+        outputsList = []
+        for output_ in range(0, modelBuf.Subgraphs(0).OutputsLength()):
+            # Output tensor is last tensor
+            tensorIndex = modelBuf.Subgraphs(0).Outputs(output_)
+
+            dtype: Union[Type[Any]]
+            # Generate dictionary of tensor details
+            if modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.INT8:
+                dtype = np.int8
+            elif modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.INT16:
+                dtype = np.int16
+            elif modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.INT32:
+                dtype = np.int32
+            elif (
+                modelBuf.Subgraphs(0).Tensors(tensorIndex).Type() == TensorType.FLOAT32
+            ):
+                dtype = np.float32
+
+            details = {
+                "name": str(modelBuf.Subgraphs(0).Tensors(tensorIndex).Name())[
+                    1:
+                ].strip("'"),
+                "index": tensorIndex,
+                "shape": modelBuf.Subgraphs(0).Tensors(tensorIndex).ShapeAsNumpy(),
+                "shape_signature": modelBuf.Subgraphs(0)
+                .Tensors(tensorIndex)
+                .ShapeSignatureAsNumpy(),
+                "dtype": dtype,
+                "quantization": (
+                    modelBuf.Subgraphs(0).Tensors(tensorIndex).Quantization().Scale(0),
+                    modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .ZeroPoint(0),
+                ),
+                "quantization_parameters": {
+                    "scales": modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .ScaleAsNumpy(),
+                    "zero_points": modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .ZeroPointAsNumpy(),
+                    "quantized_dimension": modelBuf.Subgraphs(0)
+                    .Tensors(tensorIndex)
+                    .Quantization()
+                    .QuantizedDimension(),
+                },
+                "sparsity_parameters": {
+                    modelBuf.Subgraphs(0).Tensors(tensorIndex).Sparsity()
+                },
+            }
+            outputsList.append(details)
+
+        return outputsList
+
+    def set_model(
+        self,
+        model_path: Optional[str] = None,
+        model_content: Optional[bytes] = None,
+        params_path: Optional[str] = None,
+        params_content: Optional[bytes] = None,
+        model_index: int = 0,
+        secondary_memory: bool = False,
+        flash: bool = False,
+    ) -> None:
+        """! Adds a model to the interpreter's list of models.
+        @param model_path The path to the model file (.tflite), alternative to model_content.
+        @param model_content The byte array representing a model, alternative to model_path.
+        @param params_path The path to the params file for the model,
+        alternative to params_content (optional).
+        @param params_content The byte array representing the model parameters,
+        alternative to params_path (optional).
+        @param model_index The model to target, for interpreters that support multiple models
+        running concurrently. Defaults to 0 for use with a single model.
+        """
+
+        # Check model_path or model_content is valid
+        if not model_path and not model_content:
+            raise ValueError("model_path or model_content must be provided")
+        tile_found = False
+        # Find correct model and replace
+        for model in self.models:
+            if model.tile == model_index:
+                model = self.modelData(
+                    model_path,
+                    model_content,
+                    params_path,
+                    params_content,
+                    model_index,
+                    secondary_memory,
+                    flash,
+                )
+                tile_found = True
+                break
+        # If model wasn't previously set, add it to list
+        if not tile_found:
+            self.models.append(
+                self.modelData(
+                    model_path,
+                    model_content,
+                    params_path,
+                    params_content,
+                    model_index,
+                    secondary_memory,
+                    flash,
+                )
+            )
+        self.initialise_interpreter(model_index)
+
+    def get_model(self, model_index: int = 0):
+        for model in self.models:
+            if model.tile == model_index:
+                return model
+
+    class modelData:
+        """! The model data class
+        A class that holds a model and data associated with a single model.
+        """
+
+        def __init__(
+            self,
+            model_path: Optional[str],
+            model_content: Optional[bytes],
+            params_path: Optional[str],
+            params_content: Optional[bytes],
+            model_index: int,
+            secondary_memory: bool,
+            flash: bool,
+        ):
+            """! Model data initializer.
+            Sets up variables, generates a list of operators used in the model,
+            and reads model and params paths into byte arrays (content).
+            @param model_path Path to the model file (.tflite).
+            @param model_content Model model_content (byte array).
+            @param params_path Path to model parameters file.
+            @param params_content Model parameters content (byte array)
+            @param model_index The model to target, for interpreters that support multiple models
+            running concurrently. Defaults to 0 for use with a single model.
+            """
+            self.model_path: Optional[str] = model_path
+            self.model_content: Optional[bytes] = model_content
+            self.params_path: Optional[str] = params_path
+            self.params_content: Optional[bytes] = params_content
+            self.tile: int = model_index
+            self.secondary_memory = secondary_memory
+            self.flash = flash
+            self.opList: List[str] = []
+            self.pathToContent()
+            self.modelToOpList()
+
+        def modelToOpList(self) -> None:
+            """! Generates operator list from model."""
+
+            # Load model
+            buffer = self.model_content
+            model = Model.GetRootAsModel(buffer, 0)
+            self.opList = []
+
+            # Iterate through operators in model and add operators to opList
+            for y in range(0, model.Subgraphs(0).OperatorsLength()):
+                opcode = model.OperatorCodes(
+                    model.Subgraphs(0).Operators(y).OpcodeIndex()
+                )
+                # If custom opcode parse string
+                if opcode.BuiltinCode() == 32:
+                    self.opList.append(str(opcode.CustomCode()).strip("b'"))
+                # If built in op code, decode
+                else:
+                    self.opList.append(opcode2name(opcode.BuiltinCode()))
+
+        def pathToContent(self) -> None:
+            """! Reads model and params paths to content (byte arrays)"""
+
+            # Check if path exists but not content
+            if self.model_content is None and self.model_path is not None:
+                with open(self.model_path, "rb") as input_fd:
+                    self.model_content = input_fd.read()
+
+            # Check if params_path exists but not params_content
+            if self.params_content is None and self.params_path is not None:
+                with open(self.params_path, "rb") as input_fd2:
+                    self.params_content = input_fd2.read()
+
+            # If params_content is None, set to empty byte array
+            if self.params_content is None:
+                self.params_content = bytes([])

xmos_ai_tools-1.3.2.dev180.dist-info/METADATA

@@ -0,0 +1,33 @@
+Metadata-Version: 2.1
+Name: xmos_ai_tools
+Version: 1.3.2.dev180
+Summary: XMOS AI Tools
+Home-page: https://github.com/xmos/ai_tools
+Author: XMOS
+Author-email: support@xmos.com
+License: LICENSE.txt
+Keywords: tensorflow binarized neural networks
+Classifier: License :: Other/Proprietary License
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Education
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Software Development
+Classifier: Topic :: Software Development :: Libraries
+Classifier: Topic :: Software Development :: Libraries :: Python Modules
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: numpy<2.0
+Requires-Dist: tflite>=2.4.0
+
+Documentation
+-------------
+
+Click [here](https://github.com/xmos/ai_tools/blob/7d0536c84f7f2e689cb6a11fd2f5e62b3e4dfeee/README.md) for documentation on using xmos-ai-tools to deploy AI models on xcore.