SegMamba/monai/visualize/gradient_based.py

# Copyright (c) MONAI Consortium
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#     http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import annotations

from functools import partial
from typing import Any, Callable

import torch

from monai.networks.utils import replace_modules_temp
from monai.utils.module import optional_import
from monai.visualize.class_activation_maps import ModelWithHooks

trange, has_trange = optional_import("tqdm", name="trange")

__all__ = ["VanillaGrad", "SmoothGrad", "GuidedBackpropGrad", "GuidedBackpropSmoothGrad"]


class _AutoGradReLU(torch.autograd.Function):

    @staticmethod
    def forward(ctx, x):
        pos_mask = (x > 0).type_as(x)
        output = torch.mul(x, pos_mask)
        ctx.save_for_backward(x, output)
        return output

    @staticmethod
    def backward(ctx, grad_output):
        x, _ = ctx.saved_tensors
        pos_mask_1 = (x > 0).type_as(grad_output)
        pos_mask_2 = (grad_output > 0).type_as(grad_output)
        y = torch.mul(grad_output, pos_mask_1)
        grad_input = torch.mul(y, pos_mask_2)
        return grad_input


class _GradReLU(torch.nn.Module):
    """
    A customized ReLU with the backward pass imputed for guided backpropagation (https://arxiv.org/abs/1412.6806).
    """

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        out: torch.Tensor = _AutoGradReLU.apply(x)
        return out


class VanillaGrad:
    """
    Given an input image ``x``, calling this class will perform the forward pass, then set to zero
    all activations except one (defined by ``index``) and propagate back to the image to achieve a gradient-based
    saliency map.

    If ``index`` is None, argmax of the output logits will be used.

    See also:

        - Simonyan et al. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
          (https://arxiv.org/abs/1312.6034)
    """

    def __init__(self, model: torch.nn.Module) -> None:
        if not isinstance(model, ModelWithHooks):  # Convert to model with hooks if necessary
            self._model = ModelWithHooks(model, target_layer_names=(), register_backward=True)
        else:
            self._model = model

    @property
    def model(self):
        return self._model.model

    @model.setter
    def model(self, m):
        if not isinstance(m, ModelWithHooks):  # regular model as ModelWithHooks
            self._model.model = m
        else:
            self._model = m  # replace the ModelWithHooks

    def get_grad(
        self, x: torch.Tensor, index: torch.Tensor | int | None, retain_graph: bool = True, **kwargs: Any
    ) -> torch.Tensor:
        if x.shape[0] != 1:
            raise ValueError("expect batch size of 1")
        x.requires_grad = True

        self._model(x, class_idx=index, retain_graph=retain_graph, **kwargs)
        grad: torch.Tensor = x.grad.detach()  # type: ignore
        return grad

    def __call__(self, x: torch.Tensor, index: torch.Tensor | int | None = None, **kwargs: Any) -> torch.Tensor:
        return self.get_grad(x, index, **kwargs)


class SmoothGrad(VanillaGrad):
    """
    Compute averaged sensitivity map based on ``n_samples`` (Gaussian additive) of noisy versions
    of the input image ``x``.

    See also:

        - Smilkov et al. SmoothGrad: removing noise by adding noise https://arxiv.org/abs/1706.03825
    """

    def __init__(
        self,
        model: torch.nn.Module,
        stdev_spread: float = 0.15,
        n_samples: int = 25,
        magnitude: bool = True,
        verbose: bool = True,
    ) -> None:
        super().__init__(model)
        self.stdev_spread = stdev_spread
        self.n_samples = n_samples
        self.magnitude = magnitude
        self.range: Callable
        if verbose and has_trange:
            self.range = partial(trange, desc=f"Computing {self.__class__.__name__}")
        else:
            self.range = range

    def __call__(self, x: torch.Tensor, index: torch.Tensor | int | None = None, **kwargs: Any) -> torch.Tensor:
        stdev = (self.stdev_spread * (x.max() - x.min())).item()
        total_gradients = torch.zeros_like(x)
        for _ in self.range(self.n_samples):
            # create noisy image
            noise = torch.normal(0, stdev, size=x.shape, dtype=torch.float32, device=x.device)
            x_plus_noise = x + noise
            x_plus_noise = x_plus_noise.detach()

            # get gradient and accumulate
            grad = self.get_grad(x_plus_noise, index, **kwargs)
            total_gradients += (grad * grad) if self.magnitude else grad

        # average
        if self.magnitude:
            total_gradients = total_gradients**0.5

        return total_gradients / self.n_samples


class GuidedBackpropGrad(VanillaGrad):
    """
    Based on Springenberg and Dosovitskiy et al. https://arxiv.org/abs/1412.6806,
    compute gradient-based saliency maps by backpropagating positive gradients and inputs (see ``_AutoGradReLU``).

    See also:

        - Springenberg and Dosovitskiy et al. Striving for Simplicity: The All Convolutional Net
          (https://arxiv.org/abs/1412.6806)
    """

    def __call__(self, x: torch.Tensor, index: torch.Tensor | int | None = None, **kwargs: Any) -> torch.Tensor:
        with replace_modules_temp(self.model, "relu", _GradReLU(), strict_match=False):
            return super().__call__(x, index, **kwargs)


class GuidedBackpropSmoothGrad(SmoothGrad):
    """
    Compute gradient-based saliency maps based on both ``GuidedBackpropGrad`` and ``SmoothGrad``.
    """

    def __call__(self, x: torch.Tensor, index: torch.Tensor | int | None = None, **kwargs: Any) -> torch.Tensor:
        with replace_modules_temp(self.model, "relu", _GradReLU(), strict_match=False):
            return super().__call__(x, index, **kwargs)