data/pathml/inference/inference.py

"""
Copyright 2023, Dana-Farber Cancer Institute and Weill Cornell Medicine
License: GNU GPL 2.0
"""

import os

import numpy as np
import onnx
import onnxruntime
import requests
import torch

import pathml
import pathml.preprocessing.transforms as Transforms
from pathml.inference.mesmer_utils import (
    deep_watershed,
    format_output_mesmer,
    mesmer_preprocess,
)


def remove_initializer_from_input(model_path, new_path):
    """Removes initializers from HaloAI ONNX models
    Taken from https://github.com/microsoft/onnxruntime/blob/main/tools/python/remove_initializer_from_input.py

    Args:
        model_path (str): path to ONNX model,
        new_path (str): path to save adjusted model w/o initializers,

    Returns:
        ONNX model w/o initializers to run inference using PathML
    """

    model = onnx.load(model_path)

    inputs = model.graph.input
    name_to_input = {}
    for onnx_input in inputs:
        name_to_input[onnx_input.name] = onnx_input

    for initializer in model.graph.initializer:
        if initializer.name in name_to_input:
            inputs.remove(name_to_input[initializer.name])

    onnx.save(model, new_path)


def check_onnx_clean(model_path):
    """Checks if the model has had it's initalizers removed from input graph.
    Adapted from from https://github.com/microsoft/onnxruntime/blob/main/tools/python/remove_initializer_from_input.py

    Args:
        model_path (str): path to ONNX model,

    Returns:
        Boolean if there are initializers in input graph.
    """

    model = onnx.load(model_path)

    inputs = model.graph.input
    name_to_input = {}
    for onnx_input in inputs:
        name_to_input[onnx_input.name] = onnx_input

    for initializer in model.graph.initializer:
        if initializer.name in name_to_input:
            return True


def convert_pytorch_onnx(
    model, dummy_tensor, model_name, opset_version=10, input_name="data"
):
    """Converts a Pytorch Model to ONNX
    Adjusted from https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html

    You need to define the model class and load the weights before exporting. See URL above for full steps.

    Args:
        model_path (torch.nn.Module Model): Pytorch model to be converted,
        dummy_tensor (torch.tensor): dummy input tensor that is an example of what will be passed into the model,
        model_name (str): name of ONNX model created with .onnx at the end,
        opset_version (int): which opset version you want to use to export
        input_name (str): name assigned to dummy_tensor

    Returns:
        Exports ONNX model converted from Pytorch
    """

    if not isinstance(model, torch.nn.Module):
        raise ValueError(
            f"The model is not of type torch.nn.Module. Received {type(model)}."
        )

    if not torch.is_tensor(dummy_tensor):
        raise ValueError(
            f"The dummy tensor needs to be a torch tensor. Received {type(dummy_tensor)}."
        )

    torch.onnx.export(
        model,
        dummy_tensor,
        model_name,
        export_params=True,
        opset_version=opset_version,
        do_constant_folding=True,
        input_names=[input_name],
    )


# Base class
class InferenceBase(Transforms.Transform):
    """
    Base class for all ONNX Models.
    Each transform must operate on a Tile.
    """

    def __init__(self):
        self.model_card = {
            "name": None,
            "num_classes": None,
            "model_type": None,
            "notes": None,
            "model_input_notes": None,
            "model_output_notes": None,
            "citation": None,
        }

    def __repr__(self):
        return "Base class for all ONNX models"

    def get_model_card(self):
        """Returns model card."""
        return self.model_card

    def set_name(self, name):
        """
        Sets the "name" parameter in the model card.

        Args:
            name (str): Name for the model
        """
        self.model_card["name"] = name

    def set_num_classes(self, num):
        """
        Sets the "num_classes" parameter in the model card.

        Args:
            num (int): Number of classes your model predicts
        """
        self.model_card["num_classes"] = num

    def set_model_type(self, model_type):
        """
        Sets the "model_type" parameter in the model card.

        Args:
            model_type (str): Type of model, e.g. "segmentation"
        """
        self.model_card["model_type"] = model_type

    def set_notes(self, note):
        """
        Sets the "notes" parameter in the model card.

        Args:
            note (str): Any extra information you want to put in the model card
        """
        self.model_card["notes"] = note

    def set_model_input_notes(self, note):
        """
        Sets the "model_input_notes" parameter in the model card.

        Args:
            note (str): Comments on the model input
        """
        self.model_card["model_input_notes"] = note

    def set_model_output_notes(self, note):
        """
        Sets the "model_output_notes" parameter in the model card.

        Args:
            note (str): Comments on the model output
        """
        self.model_card["model_output_notes"] = note

    def set_citation(self, citation):
        """
        Sets the "citation" parameter in the model card.

        Args:
            citation (str): Citation for the model
        """
        self.model_card["citation"] = citation

    def reshape(self, image):
        """standard reshaping of tile image"""
        # flip dimensions
        # follows convention used here https://github.com/Dana-Farber-AIOS/pathml/blob/master/pathml/ml/dataset.py

        if image.ndim == 3:
            # swap axes from HWC to CHW
            image = image.transpose(2, 0, 1)
            # add a dimesion bc onnx models usually have batch size as first dim: e.g. (1, channel, height, width)
            image = np.expand_dims(image, axis=0)

            return image
        else:
            # in this case, we assume that we have XYZCT channel order
            # so we swap axes to TCZYX for batching
            # note we are not adding a dim here for batch bc we assume that subsetting will create a batch "placeholder" dim
            image = image.T

            return image

    def F(self, target):
        """functional implementation"""
        raise NotImplementedError

    def apply(self, tile):
        """modify Tile object in-place"""
        raise NotImplementedError


# class to handle local onnx models
class Inference(InferenceBase):
    """Transformation to run inferrence on ONNX model.

    Assumptions:
        - The ONNX model has been cleaned by `remove_initializer_from_input` first

    Args:
        model_path (str): path to ONNX model w/o initializers,
        input_name (str): name of the input the ONNX model accepts, default = "data"
        num_classes (int): number of classes you are predicting
        model_type (str): type of model, e.g. "segmentation"
        local (bool): True if the model is stored locally, default = "True"
    """

    def __init__(
        self,
        model_path=None,
        input_name="data",
        num_classes=None,
        model_type=None,
        local=True,
    ):
        super().__init__()

        self.input_name = input_name
        self.num_classes = num_classes
        self.model_type = model_type
        self.local = local

        if self.local:
            # using a local onnx model
            self.model_path = model_path
        else:
            # if using a model from the model zoo, set the local path to a temp file
            self.model_path = "temp.onnx"

        # fill in parts of the model_card with the following info
        self.model_card["num_classes"] = self.num_classes
        self.model_card["model_type"] = self.model_type

        # check if there are initializers in input graph if using a local model
        if local:
            if check_onnx_clean(model_path):
                raise ValueError(
                    "The ONNX model still has graph initializers in the input graph. Use `remove_initializer_from_input` to remove them."
                )
        else:
            pass

    def __repr__(self):
        if self.local:
            return f"Class to handle ONNX model locally stored at {self.model_path}"
        else:
            return f"Class to handle a {self.model_card['name']} from the PathML model zoo."

    def inference(self, image):
        # reshape the image
        image = self.reshape(image)

        # load fixed model
        onnx_model = onnx.load(self.model_path)

        # check tile dimensions match ONNX input dimensions
        input_node = onnx_model.graph.input

        dimensions = []
        for input in input_node:
            if input.name == self.input_name:
                input_shape = input.type.tensor_type.shape.dim
                for dim in input_shape:
                    dimensions.append(dim.dim_value)

        assert (
            image.shape[-1] == dimensions[-1] and image.shape[-2] == dimensions[-2]
        ), f"expecting tile shape of {dimensions[-2]} by {dimensions[-1]}, got {image.shape[-2]} by {image.shape[-1]}"

        # check onnx model
        onnx.checker.check_model(onnx_model)

        # start an inference session
        ort_sess = onnxruntime.InferenceSession(self.model_path)

        # create model output, returns a list
        model_output = ort_sess.run(None, {self.input_name: image.astype("f")})

        return model_output

    def F(self, image):
        # run inference function
        prediction_map = self.inference(image)

        # single task model
        if len(prediction_map) == 1:
            # return first and only prediction array in the list
            return prediction_map[0]

        # multi task model
        else:
            # concatenate prediction results
            # assumes that the tasks all output prediction arrays of same dimension on H and W
            result_array = np.concatenate(prediction_map, axis=1)
            return result_array

    def apply(self, tile):
        tile.image = self.F(tile.image)


class HaloAIInference(Inference):
    """Transformation to run inferrence on HALO AI ONNX model.

    Assumptions:
        - Assumes that the ONNX model returns a tensor in which there is one prediction map for each class
        - For example, if there are 5 classes, the ONNX model will output a (1, 5, Height, Weight) tensor
        - If you select to argmax the classes, the class assumes a softmax or sigmoid has already been applied
        - HaloAI ONNX models always have 20 class maps so you need to index into the first x maps if you have x classes


    Args:
        model_path (str): path to HaloAI ONNX model w/o initializers,
        input_name (str): name of the input the ONNX model accepts, default = "data"
        num_classes (int): number of classes you are predicting
        model_type (str): type of model, e.g. "segmentation"
        local (bool): True if the model is stored locally, default = "True"
    """

    def __init__(
        self,
        model_path=None,
        input_name="data",
        num_classes=None,
        model_type=None,
        local=True,
    ):
        super().__init__(model_path, input_name, num_classes, model_type, local)

        self.model_card["num_classes"] = self.num_classes
        self.model_card["model_type"] = self.model_type

    def __repr__(self):
        return f"Class to handle HALO AI ONNX model locally stored at {self.model_path}"

    def F(self, image):
        prediction_map = self.inference(image)

        prediction_map = prediction_map[0][:, 0 : self.num_classes, :, :]

        return prediction_map

    def apply(self, tile):
        tile.image = self.F(tile.image)


# class to handle remote onnx models
class RemoteTestHoverNet(Inference):
    """Transformation to run inference on ONNX model.

    Citation for model:
    Pocock J, Graham S, Vu QD, Jahanifar M, Deshpande S, Hadjigeorghiou G, Shephard A, Bashir RM, Bilal M, Lu W, Epstein D.
    TIAToolbox as an end-to-end library for advanced tissue image analytics. Communications medicine. 2022 Sep 24;2(1):120.

    Args:
        model_path (str): temp file name to download onnx from huggingface, do not change
        input_name (str): name of the input the ONNX model accepts, default = "data", do not change
        num_classes (int): number of classes you are predicting, do not change
        model_type (str): type of model, e.g. "segmentation", do not change
        local (bool): True if the model is stored locally, default = "True", do not change
    """

    def __init__(
        self,
        model_path="temp.onnx",
        input_name="data",
        num_classes=5,
        model_type="Segmentation",
        local=False,
    ):
        super().__init__(model_path, input_name, num_classes, model_type, local)

        # specify URL of the model in PathML public repository
        url = "https://huggingface.co/pathml/test/resolve/main/hovernet_fast_tiatoolbox_fixed.onnx"

        # download model, save as temp.onnx
        with open(self.model_path, "wb") as out_file:
            content = requests.get(url, stream=True).content
            out_file.write(content)

        self.model_card["num_classes"] = self.num_classes
        self.model_card["model_type"] = self.model_type
        self.model_card["name"] = "Tiabox HoverNet Test"
        self.model_card["model_input_notes"] = "Accepts tiles of 256 x 256"
        self.model_card["citation"] = (
            "Pocock J, Graham S, Vu QD, Jahanifar M, Deshpande S, Hadjigeorghiou G, Shephard A, Bashir RM, Bilal M, Lu W, Epstein D. TIAToolbox as an end-to-end library for advanced tissue image analytics. Communications medicine. 2022 Sep 24;2(1):120."
        )

    def __repr__(self):
        return "Class to handle remote TIAToolBox HoverNet test ONNX. See model card for citation."

    def apply(self, tile):
        tile.image = self.F(tile.image)

    def remove(self):
        # remove the temp.onnx model
        os.remove(self.model_path)


class RemoteMesmer(Inference):
    """
    Transformation to run inference on ONNX Mesmer model.

    Citation for model:
    Greenwald NF, Miller G, Moen E, Kong A, Kagel A, Dougherty T, Fullaway CC, McIntosh BJ, Leow KX, Schwartz MS, Pavelchek C.
    Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning.
    Nature biotechnology. 2022 Apr;40(4):555-65.

    Args:
        model_path (str): temp file name to download onnx from huggingface, do not change
        input_name (str): name of the input the ONNX model accepts, default = "data", do not change
        num_classes (int): number of classes you are predicting, do not change
        model_type (str): type of model, e.g. "segmentation", do not change
        local (bool): True if the model is stored locally, default = "True", do not change
        nuclear_channel(int): channel that defines cell nucleus
        cytoplasm_channel(int): channel that defines cell membrane or cytoplasm
        image_resolution(float): pixel resolution of image in microns. Currently only supports 0.5
        preprocess_kwargs(dict): keyword arguemnts to pass to pre-processing function
        postprocess_kwargs_nuclear(dict): keyword arguments to pass to post-processing function
        postprocess_kwargs_whole_cell(dict): keyword arguments to pass to post-processing function
    """

    def __init__(
        self,
        model_path="temp.onnx",
        input_name="data",
        num_classes=3,
        model_type="Segmentation",
        local=False,
        nuclear_channel=None,
        cytoplasm_channel=None,
        image_resolution=0.5,
        preprocess_kwargs=None,
        postprocess_kwargs_nuclear=None,
        postprocess_kwargs_whole_cell=None,
    ):
        super().__init__(model_path, input_name, num_classes, model_type, local)
        assert isinstance(
            nuclear_channel, int
        ), "nuclear_channel must be an int indicating index"
        assert isinstance(
            cytoplasm_channel, int
        ), "cytoplasm_channel must be an int indicating index"
        self.nuclear_channel = nuclear_channel
        self.cytoplasm_channel = cytoplasm_channel
        self.image_resolution = image_resolution
        self.preprocess_kwargs = preprocess_kwargs if preprocess_kwargs else {}
        self.postprocess_kwargs_nuclear = (
            postprocess_kwargs_nuclear if postprocess_kwargs_nuclear else {}
        )
        self.postprocess_kwargs_whole_cell = (
            postprocess_kwargs_whole_cell if postprocess_kwargs_whole_cell else {}
        )

        # specify URL of the model in PathML public repository
        url = "https://huggingface.co/pathml/test/resolve/main/mesmer.onnx"

        # download model, save as temp.onnx
        with open(self.model_path, "wb") as out_file:
            content = requests.get(url, stream=True).content
            out_file.write(content)

        self.model_card["num_classes"] = self.num_classes
        self.model_card["model_type"] = self.model_type
        self.model_card["name"] = "Deepcell's Mesmer"
        self.model_card["model_input_notes"] = (
            "Accepts tiles of 256 x 256, resolution must be 0.5. Unlike other inference classes, segmentation maps are saved to tile.masks."
        )
        self.model_card["citation"] = (
            "Greenwald NF, Miller G, Moen E, Kong A, Kagel A, Dougherty T, Fullaway CC, McIntosh BJ, Leow KX, Schwartz MS, Pavelchek C. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nature biotechnology. 2022 Apr;40(4):555-65."
        )

        print(self.model_card["model_input_notes"])

        if not (self.image_resolution == 0.5):  # pragma: no cover
            print("The model only works with images of resolution 0.5.")

    def __repr__(self):
        return "Class to handle remote Mesmer Model from Deepcell. See model card for citation."

    def remove(self):
        # remove the temp.onnx model
        os.remove(self.model_path)

    def inference(self, image):
        # load fixed model
        onnx_model = onnx.load(self.model_path)

        # check tile dimensions match ONNX input dimensions
        input_node = onnx_model.graph.input

        dimensions = []
        for input in input_node:
            if input.name == self.input_name:
                input_shape = input.type.tensor_type.shape.dim
                for dim in input_shape:
                    dimensions.append(dim.dim_value)

        # check onnx model
        onnx.checker.check_model(onnx_model)

        # start an inference session
        ort_sess = onnxruntime.InferenceSession(self.model_path)

        # create model output, returns a list
        model_output = ort_sess.run(None, {self.input_name: image.astype("f")})

        return model_output

    def F(self, image):
        img = image.copy()
        if len(img.shape) not in [3, 4]:
            raise ValueError(
                f"input image has shape {img.shape}. supported image shapes are x,y,c or batch,x,y,c."
            )  # pragma: no cover
        if len(img.shape) == 3:
            img = np.expand_dims(img, axis=0)
        if img.shape[1] != 256 and img.shape[2] != 256:
            raise ValueError(
                f"input image has shape {img.shape}. currently, we only support image shapes that are (256,256,c) or (batch,256,256,c)."
            )  # pragma: no cover
        nuc_cytoplasm = np.stack(
            (img[:, :, :, self.nuclear_channel], img[:, :, :, self.cytoplasm_channel]),
            axis=-1,
        )

        # get pre-processing output
        pre_processed_output = mesmer_preprocess(
            nuc_cytoplasm, **self.preprocess_kwargs
        )

        # run infernece
        output = self.inference(pre_processed_output)

        # reformat output
        output = format_output_mesmer(output)

        # post-processing
        label_images_cell = deep_watershed(
            output["whole-cell"], **self.postprocess_kwargs_whole_cell
        )

        label_images_nucleus = deep_watershed(
            output["nuclear"], **self.postprocess_kwargs_nuclear
        )

        return np.squeeze(label_images_cell, axis=0), np.squeeze(
            label_images_nucleus, axis=0
        )

    def apply(self, tile):
        assert isinstance(
            tile, pathml.core.tile.Tile
        ), f"tile is type {type(tile)} but must be pathml.core.tile.Tile"
        assert (
            tile.slide_type.stain == "Fluor"
        ), f"Tile has slide_type.stain='{tile.slide_type.stain}', but must be 'Fluor'"

        cell_segmentation, nuclear_segmentation = self.F(tile.image)
        tile.masks["cell_segmentation"] = cell_segmentation
        tile.masks["nuclear_segmentation"] = nuclear_segmentation