# 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 import itertools from collections.abc import Callable, Iterable, Mapping, Sequence from typing import Any import numpy as np import torch import torch.nn.functional as F from monai.data.meta_tensor import MetaTensor from monai.data.utils import compute_importance_map, dense_patch_slices, get_valid_patch_size from monai.utils import ( BlendMode, PytorchPadMode, convert_data_type, convert_to_dst_type, ensure_tuple, ensure_tuple_rep, fall_back_tuple, look_up_option, optional_import, ) tqdm, _ = optional_import("tqdm", name="tqdm") _nearest_mode = "nearest-exact" __all__ = ["sliding_window_inference"] def sliding_window_inference( inputs: torch.Tensor | MetaTensor, roi_size: Sequence[int] | int, sw_batch_size: int, predictor: Callable[..., torch.Tensor | Sequence[torch.Tensor] | dict[Any, torch.Tensor]], overlap: Sequence[float] | float = 0.25, mode: BlendMode | str = BlendMode.CONSTANT, sigma_scale: Sequence[float] | float = 0.125, padding_mode: PytorchPadMode | str = PytorchPadMode.CONSTANT, cval: float = 0.0, sw_device: torch.device | str | None = None, device: torch.device | str | None = None, progress: bool = False, roi_weight_map: torch.Tensor | None = None, process_fn: Callable | None = None, buffer_steps: int | None = None, buffer_dim: int = -1, with_coord: bool = False, *args: Any, **kwargs: Any, ) -> torch.Tensor | tuple[torch.Tensor, ...] | dict[Any, torch.Tensor]: """ Sliding window inference on `inputs` with `predictor`. The outputs of `predictor` could be a tensor, a tuple, or a dictionary of tensors. Each output in the tuple or dict value is allowed to have different resolutions with respect to the input. e.g., the input patch spatial size is [128,128,128], the output (a tuple of two patches) patch sizes could be ([128,64,256], [64,32,128]). In this case, the parameter `overlap` and `roi_size` need to be carefully chosen to ensure the output ROI is still an integer. If the predictor's input and output spatial sizes are not equal, we recommend choosing the parameters so that `overlap*roi_size*output_size/input_size` is an integer (for each spatial dimension). When roi_size is larger than the inputs' spatial size, the input image are padded during inference. To maintain the same spatial sizes, the output image will be cropped to the original input size. Args: inputs: input image to be processed (assuming NCHW[D]) roi_size: the spatial window size for inferences. When its components have None or non-positives, the corresponding inputs dimension will be used. if the components of the `roi_size` are non-positive values, the transform will use the corresponding components of img size. For example, `roi_size=(32, -1)` will be adapted to `(32, 64)` if the second spatial dimension size of img is `64`. sw_batch_size: the batch size to run window slices. predictor: given input tensor ``patch_data`` in shape NCHW[D], The outputs of the function call ``predictor(patch_data)`` should be a tensor, a tuple, or a dictionary with Tensor values. Each output in the tuple or dict value should have the same batch_size, i.e. NM'H'W'[D']; where H'W'[D'] represents the output patch's spatial size, M is the number of output channels, N is `sw_batch_size`, e.g., the input shape is (7, 1, 128,128,128), the output could be a tuple of two tensors, with shapes: ((7, 5, 128, 64, 256), (7, 4, 64, 32, 128)). In this case, the parameter `overlap` and `roi_size` need to be carefully chosen to ensure the scaled output ROI sizes are still integers. If the `predictor`'s input and output spatial sizes are different, we recommend choosing the parameters so that ``overlap*roi_size*zoom_scale`` is an integer for each dimension. overlap: Amount of overlap between scans along each spatial dimension, defaults to ``0.25``. mode: {``"constant"``, ``"gaussian"``} How to blend output of overlapping windows. Defaults to ``"constant"``. - ``"constant``": gives equal weight to all predictions. - ``"gaussian``": gives less weight to predictions on edges of windows. sigma_scale: the standard deviation coefficient of the Gaussian window when `mode` is ``"gaussian"``. Default: 0.125. Actual window sigma is ``sigma_scale`` * ``dim_size``. When sigma_scale is a sequence of floats, the values denote sigma_scale at the corresponding spatial dimensions. padding_mode: {``"constant"``, ``"reflect"``, ``"replicate"``, ``"circular"``} Padding mode for ``inputs``, when ``roi_size`` is larger than inputs. Defaults to ``"constant"`` See also: https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html cval: fill value for 'constant' padding mode. Default: 0 sw_device: device for the window data. By default the device (and accordingly the memory) of the `inputs` is used. Normally `sw_device` should be consistent with the device where `predictor` is defined. device: device for the stitched output prediction. By default the device (and accordingly the memory) of the `inputs` is used. If for example set to device=torch.device('cpu') the gpu memory consumption is less and independent of the `inputs` and `roi_size`. Output is on the `device`. progress: whether to print a `tqdm` progress bar. roi_weight_map: pre-computed (non-negative) weight map for each ROI. If not given, and ``mode`` is not `constant`, this map will be computed on the fly. process_fn: process inference output and adjust the importance map per window buffer_steps: the number of sliding window iterations along the ``buffer_dim`` to be buffered on ``sw_device`` before writing to ``device``. (Typically, ``sw_device`` is ``cuda`` and ``device`` is ``cpu``.) default is None, no buffering. For the buffer dim, when spatial size is divisible by buffer_steps*roi_size, (i.e. no overlapping among the buffers) non_blocking copy may be automatically enabled for efficiency. buffer_dim: the spatial dimension along which the buffers are created. 0 indicates the first spatial dimension. Default is -1, the last spatial dimension. with_coord: whether to pass the window coordinates to ``predictor``. Default is False. If True, the signature of ``predictor`` should be ``predictor(patch_data, patch_coord, ...)``. args: optional args to be passed to ``predictor``. kwargs: optional keyword args to be passed to ``predictor``. Note: - input must be channel-first and have a batch dim, supports N-D sliding window. """ buffered = buffer_steps is not None and buffer_steps > 0 num_spatial_dims = len(inputs.shape) - 2 if buffered: if buffer_dim < -num_spatial_dims or buffer_dim > num_spatial_dims: raise ValueError(f"buffer_dim must be in [{-num_spatial_dims}, {num_spatial_dims}], got {buffer_dim}.") if buffer_dim < 0: buffer_dim += num_spatial_dims overlap = ensure_tuple_rep(overlap, num_spatial_dims) for o in overlap: if o < 0 or o >= 1: raise ValueError(f"overlap must be >= 0 and < 1, got {overlap}.") compute_dtype = inputs.dtype # determine image spatial size and batch size # Note: all input images must have the same image size and batch size batch_size, _, *image_size_ = inputs.shape device = device or inputs.device sw_device = sw_device or inputs.device condition = kwargs.pop("condition", None) temp_meta = None if isinstance(inputs, MetaTensor): temp_meta = MetaTensor([]).copy_meta_from(inputs, copy_attr=False) inputs = convert_data_type(inputs, torch.Tensor, wrap_sequence=True)[0] roi_size = fall_back_tuple(roi_size, image_size_) # in case that image size is smaller than roi size image_size = tuple(max(image_size_[i], roi_size[i]) for i in range(num_spatial_dims)) pad_size = [] for k in range(len(inputs.shape) - 1, 1, -1): diff = max(roi_size[k - 2] - inputs.shape[k], 0) half = diff // 2 pad_size.extend([half, diff - half]) if any(pad_size): inputs = F.pad(inputs, pad=pad_size, mode=look_up_option(padding_mode, PytorchPadMode), value=cval) if condition is not None: condition = F.pad(condition, pad=pad_size, mode=look_up_option(padding_mode, PytorchPadMode), value=cval) # Store all slices scan_interval = _get_scan_interval(image_size, roi_size, num_spatial_dims, overlap) slices = dense_patch_slices(image_size, roi_size, scan_interval, return_slice=not buffered) num_win = len(slices) # number of windows per image total_slices = num_win * batch_size # total number of windows windows_range: Iterable if not buffered: non_blocking = False windows_range = range(0, total_slices, sw_batch_size) else: slices, n_per_batch, b_slices, windows_range = _create_buffered_slices( slices, batch_size, sw_batch_size, buffer_dim, buffer_steps ) non_blocking, _ss = torch.cuda.is_available(), -1 for x in b_slices[:n_per_batch]: if x[1] < _ss: # detect overlapping slices non_blocking = False break _ss = x[2] # Create window-level importance map valid_patch_size = get_valid_patch_size(image_size, roi_size) if valid_patch_size == roi_size and (roi_weight_map is not None): importance_map_ = roi_weight_map else: try: valid_p_size = ensure_tuple(valid_patch_size) importance_map_ = compute_importance_map( valid_p_size, mode=mode, sigma_scale=sigma_scale, device=sw_device, dtype=compute_dtype ) if len(importance_map_.shape) == num_spatial_dims and not process_fn: importance_map_ = importance_map_[None, None] # adds batch, channel dimensions except Exception as e: raise RuntimeError( f"patch size {valid_p_size}, mode={mode}, sigma_scale={sigma_scale}, device={device}\n" "Seems to be OOM. Please try smaller patch size or mode='constant' instead of mode='gaussian'." ) from e importance_map_ = convert_data_type(importance_map_, torch.Tensor, device=sw_device, dtype=compute_dtype)[0] # stores output and count map output_image_list, count_map_list, sw_device_buffer, b_s, b_i = [], [], [], 0, 0 # type: ignore # for each patch for slice_g in tqdm(windows_range) if progress else windows_range: slice_range = range(slice_g, min(slice_g + sw_batch_size, b_slices[b_s][0] if buffered else total_slices)) unravel_slice = [ [slice(idx // num_win, idx // num_win + 1), slice(None)] + list(slices[idx % num_win]) for idx in slice_range ] if sw_batch_size > 1: win_data = torch.cat([inputs[win_slice] for win_slice in unravel_slice]).to(sw_device) if condition is not None: win_condition = torch.cat([condition[win_slice] for win_slice in unravel_slice]).to(sw_device) kwargs["condition"] = win_condition else: win_data = inputs[unravel_slice[0]].to(sw_device) if condition is not None: win_condition = condition[unravel_slice[0]].to(sw_device) kwargs["condition"] = win_condition if with_coord: seg_prob_out = predictor(win_data, unravel_slice, *args, **kwargs) else: seg_prob_out = predictor(win_data, *args, **kwargs) # convert seg_prob_out to tuple seg_tuple, this does not allocate new memory. dict_keys, seg_tuple = _flatten_struct(seg_prob_out) if process_fn: seg_tuple, w_t = process_fn(seg_tuple, win_data, importance_map_) else: w_t = importance_map_ if len(w_t.shape) == num_spatial_dims: w_t = w_t[None, None] w_t = w_t.to(dtype=compute_dtype, device=sw_device) if buffered: c_start, c_end = b_slices[b_s][1:] if not sw_device_buffer: k = seg_tuple[0].shape[1] # len(seg_tuple) > 1 is currently ignored sp_size = list(image_size) sp_size[buffer_dim] = c_end - c_start sw_device_buffer = [torch.zeros(size=[1, k, *sp_size], dtype=compute_dtype, device=sw_device)] for p, s in zip(seg_tuple[0], unravel_slice): offset = s[buffer_dim + 2].start - c_start s[buffer_dim + 2] = slice(offset, offset + roi_size[buffer_dim]) s[0] = slice(0, 1) sw_device_buffer[0][s] += p * w_t b_i += len(unravel_slice) if b_i < b_slices[b_s][0]: continue else: sw_device_buffer = list(seg_tuple) for ss in range(len(sw_device_buffer)): b_shape = sw_device_buffer[ss].shape seg_chns, seg_shape = b_shape[1], b_shape[2:] z_scale = None if not buffered and seg_shape != roi_size: z_scale = [out_w_i / float(in_w_i) for out_w_i, in_w_i in zip(seg_shape, roi_size)] w_t = F.interpolate(w_t, seg_shape, mode=_nearest_mode) if len(output_image_list) <= ss: output_shape = [batch_size, seg_chns] output_shape += [int(_i * _z) for _i, _z in zip(image_size, z_scale)] if z_scale else list(image_size) # allocate memory to store the full output and the count for overlapping parts new_tensor: Callable = torch.empty if non_blocking else torch.zeros # type: ignore output_image_list.append(new_tensor(output_shape, dtype=compute_dtype, device=device)) count_map_list.append(torch.zeros([1, 1] + output_shape[2:], dtype=compute_dtype, device=device)) w_t_ = w_t.to(device) for __s in slices: if z_scale is not None: __s = tuple(slice(int(_si.start * z_s), int(_si.stop * z_s)) for _si, z_s in zip(__s, z_scale)) count_map_list[-1][(slice(None), slice(None), *__s)] += w_t_ if buffered: o_slice = [slice(None)] * len(inputs.shape) o_slice[buffer_dim + 2] = slice(c_start, c_end) img_b = b_s // n_per_batch # image batch index o_slice[0] = slice(img_b, img_b + 1) if non_blocking: output_image_list[0][o_slice].copy_(sw_device_buffer[0], non_blocking=non_blocking) else: output_image_list[0][o_slice] += sw_device_buffer[0].to(device=device) else: sw_device_buffer[ss] *= w_t sw_device_buffer[ss] = sw_device_buffer[ss].to(device) _compute_coords(unravel_slice, z_scale, output_image_list[ss], sw_device_buffer[ss]) sw_device_buffer = [] if buffered: b_s += 1 if non_blocking: torch.cuda.current_stream().synchronize() # account for any overlapping sections for ss in range(len(output_image_list)): output_image_list[ss] /= count_map_list.pop(0) # remove padding if image_size smaller than roi_size if any(pad_size): kwargs.update({"pad_size": pad_size}) for ss, output_i in enumerate(output_image_list): zoom_scale = [_shape_d / _roi_size_d for _shape_d, _roi_size_d in zip(output_i.shape[2:], roi_size)] final_slicing: list[slice] = [] for sp in range(num_spatial_dims): si = num_spatial_dims - sp - 1 slice_dim = slice( int(round(pad_size[sp * 2] * zoom_scale[si])), int(round((pad_size[sp * 2] + image_size_[si]) * zoom_scale[si])), ) final_slicing.insert(0, slice_dim) output_image_list[ss] = output_i[(slice(None), slice(None), *final_slicing)] final_output = _pack_struct(output_image_list, dict_keys) if temp_meta is not None: final_output = convert_to_dst_type(final_output, temp_meta, device=device)[0] else: final_output = convert_to_dst_type(final_output, inputs, device=device)[0] return final_output # type: ignore def _create_buffered_slices(slices, batch_size, sw_batch_size, buffer_dim, buffer_steps): """rearrange slices for buffering""" slices_np = np.asarray(slices) slices_np = slices_np[np.argsort(slices_np[:, buffer_dim, 0], kind="mergesort")] slices = [tuple(slice(c[0], c[1]) for c in i) for i in slices_np] slices_np = slices_np[:, buffer_dim] _, _, _b_lens = np.unique(slices_np[:, 0], return_counts=True, return_index=True) b_ends = np.cumsum(_b_lens).tolist() # possible buffer flush boundaries x = [0, *b_ends][:: min(len(b_ends), int(buffer_steps))] if x[-1] < b_ends[-1]: x.append(b_ends[-1]) n_per_batch = len(x) - 1 windows_range = [ range(b * x[-1] + x[i], b * x[-1] + x[i + 1], sw_batch_size) for b in range(batch_size) for i in range(n_per_batch) ] b_slices = [] for _s, _r in enumerate(windows_range): s_s = slices_np[windows_range[_s - 1].stop % len(slices) if _s > 0 else 0, 0] s_e = slices_np[(_r.stop - 1) % len(slices), 1] b_slices.append((_r.stop, s_s, s_e)) # buffer index, slice start, slice end windows_range = itertools.chain(*windows_range) # type: ignore return slices, n_per_batch, b_slices, windows_range def _compute_coords(coords, z_scale, out, patch): """sliding window batch spatial scaling indexing for multi-resolution outputs.""" for original_idx, p in zip(coords, patch): idx_zm = list(original_idx) # 4D for 2D image, 5D for 3D image if z_scale: for axis in range(2, len(idx_zm)): idx_zm[axis] = slice( int(original_idx[axis].start * z_scale[axis - 2]), int(original_idx[axis].stop * z_scale[axis - 2]) ) out[idx_zm] += p def _get_scan_interval( image_size: Sequence[int], roi_size: Sequence[int], num_spatial_dims: int, overlap: Sequence[float] ) -> tuple[int, ...]: """ Compute scan interval according to the image size, roi size and overlap. Scan interval will be `int((1 - overlap) * roi_size)`, if interval is 0, use 1 instead to make sure sliding window works. """ if len(image_size) != num_spatial_dims: raise ValueError(f"len(image_size) {len(image_size)} different from spatial dims {num_spatial_dims}.") if len(roi_size) != num_spatial_dims: raise ValueError(f"len(roi_size) {len(roi_size)} different from spatial dims {num_spatial_dims}.") scan_interval = [] for i, o in zip(range(num_spatial_dims), overlap): if roi_size[i] == image_size[i]: scan_interval.append(int(roi_size[i])) else: interval = int(roi_size[i] * (1 - o)) scan_interval.append(interval if interval > 0 else 1) return tuple(scan_interval) def _flatten_struct(seg_out): dict_keys = None seg_probs: tuple[torch.Tensor, ...] if isinstance(seg_out, torch.Tensor): seg_probs = (seg_out,) elif isinstance(seg_out, Mapping): dict_keys = sorted(seg_out.keys()) # track predictor's output keys seg_probs = tuple(seg_out[k] for k in dict_keys) else: seg_probs = ensure_tuple(seg_out) return dict_keys, seg_probs def _pack_struct(seg_out, dict_keys=None): if dict_keys is not None: return dict(zip(dict_keys, seg_out)) if isinstance(seg_out, (list, tuple)) and len(seg_out) == 1: return seg_out[0] return ensure_tuple(seg_out)