# modified from https://github.com/Haiyang-W/DSVT from math import ceil import torch from torch import nn from .utils import (PositionEmbeddingLearned, get_continous_inds, get_inner_win_inds_cuda, get_pooling_index, get_window_coors) class DSVTInputLayer(nn.Module): ''' This class converts the output of vfe to dsvt input. We do in this class: 1. Window partition: partition voxels to non-overlapping windows. 2. Set partition: generate non-overlapped and size-equivalent local sets within each window. 3. Pre-compute the downsample information between two consecutive stages. 4. Pre-compute the position embedding vectors. Args: sparse_shape (tuple[int, int, int]): Shape of input space (xdim, ydim, zdim). window_shape (list[list[int, int, int]]): Window shapes (winx, winy, winz) in different stages. Length: stage_num. downsample_stride (list[list[int, int, int]]): Downsample strides between two consecutive stages. Element i is [ds_x, ds_y, ds_z], which is used between stage_i and stage_{i+1}. Length: stage_num - 1. dim_model (list[int]): Number of input channels for each stage. Length: stage_num. set_info (list[list[int, int]]): A list of set config for each stage. Eelement i contains [set_size, block_num], where set_size is the number of voxel in a set and block_num is the number of blocks for stage i. Length: stage_num. hybrid_factor (list[int, int, int]): Control the window shape in different blocks. e.g. for block_{0} and block_{1} in stage_0, window shapes are [win_x, win_y, win_z] and [win_x * h[0], win_y * h[1], win_z * h[2]] respectively. shift_list (list): Shift window. Length: stage_num. normalize_pos (bool): Whether to normalize coordinates in position embedding. ''' def __init__(self, sparse_shape, window_shape, downsample_stride, dim_model, set_info, hybrid_factor, shift_list, normalize_pos): super().__init__() self.sparse_shape = sparse_shape self.window_shape = window_shape self.downsample_stride = downsample_stride self.dim_model = dim_model self.set_info = set_info self.stage_num = len(self.dim_model) self.hybrid_factor = hybrid_factor self.window_shape = [[ self.window_shape[s_id], [ self.window_shape[s_id][coord_id] * self.hybrid_factor[coord_id] for coord_id in range(3) ] ] for s_id in range(self.stage_num)] self.shift_list = shift_list self.normalize_pos = normalize_pos self.num_shifts = [ 2, ] * len(self.window_shape) self.sparse_shape_list = [self.sparse_shape] # compute sparse shapes for each stage for ds_stride in self.downsample_stride: last_sparse_shape = self.sparse_shape_list[-1] self.sparse_shape_list.append( (ceil(last_sparse_shape[0] / ds_stride[0]), ceil(last_sparse_shape[1] / ds_stride[1]), ceil(last_sparse_shape[2] / ds_stride[2]))) # position embedding layers self.posembed_layers = nn.ModuleList() for i in range(len(self.set_info)): input_dim = 3 if self.sparse_shape_list[i][-1] > 1 else 2 stage_posembed_layers = nn.ModuleList() for j in range(self.set_info[i][1]): block_posembed_layers = nn.ModuleList() for s in range(self.num_shifts[i]): block_posembed_layers.append( PositionEmbeddingLearned(input_dim, self.dim_model[i])) stage_posembed_layers.append(block_posembed_layers) self.posembed_layers.append(stage_posembed_layers) def forward(self, batch_dict): ''' Args: bacth_dict (dict): The dict contains the following keys - voxel_features (Tensor[float]): Voxel features after VFE with shape (N, dim_model[0]), where N is the number of input voxels. - voxel_coords (Tensor[int]): Shape of (N, 4), corresponding voxel coordinates of each voxels. Each row is (batch_id, z, y, x). - ... Returns: voxel_info (dict): The dict contains the following keys - voxel_coors_stage{i} (Tensor[int]): Shape of (N_i, 4). N is the number of voxels in stage_i. Each row is (batch_id, z, y, x). - set_voxel_inds_stage{i}_shift{j} (Tensor[int]): Set partition index with shape (2, set_num, set_info[i][0]). 2 indicates x-axis partition and y-axis partition. - set_voxel_mask_stage{i}_shift{i} (Tensor[bool]): Key mask used in set attention with shape (2, set_num, set_info[i][0]). - pos_embed_stage{i}_block{i}_shift{i} (Tensor[float]): Position embedding vectors with shape (N_i, dim_model[i]). N_i is the number of remain voxels in stage_i; - pooling_mapping_index_stage{i} (Tensor[int]): Pooling region index used in pooling operation between stage_{i-1} and stage_{i} with shape (N_{i-1}). - pooling_index_in_pool_stage{i} (Tensor[int]): Index inner region with shape (N_{i-1}). Combined with pooling_mapping_index_stage{i}, we can map each voxel in satge_{i-1} to pooling_preholder_feats_stage{i}, which are input of downsample operation. - pooling_preholder_feats_stage{i} (Tensor[int]): Preholder features initial with value 0. Shape of (N_{i}, downsample_stride[i-1].prob(), d_moel[i-1]), where prob() returns the product of all elements. - ... ''' voxel_feats = batch_dict['voxel_features'] voxel_coors = batch_dict['voxel_coords'].long() voxel_info = {} voxel_info['voxel_feats_stage0'] = voxel_feats.clone() voxel_info['voxel_coors_stage0'] = voxel_coors.clone() for stage_id in range(self.stage_num): # window partition of corresponding stage-map voxel_info = self.window_partition(voxel_info, stage_id) # generate set id of corresponding stage-map voxel_info = self.get_set(voxel_info, stage_id) for block_id in range(self.set_info[stage_id][1]): for shift_id in range(self.num_shifts[stage_id]): layer_name = f'pos_embed_stage{stage_id}_block{block_id}_shift{shift_id}' # noqa: E501 pos_name = f'coors_in_win_stage{stage_id}_shift{shift_id}' voxel_info[layer_name] = self.get_pos_embed( voxel_info[pos_name], stage_id, block_id, shift_id) # compute pooling information if stage_id < self.stage_num - 1: voxel_info = self.subm_pooling(voxel_info, stage_id) return voxel_info @torch.no_grad() def subm_pooling(self, voxel_info, stage_id): # x,y,z stride cur_stage_downsample = self.downsample_stride[stage_id] # batch_win_coords is from 1 of x, y batch_win_inds, _, index_in_win, batch_win_coors = get_pooling_index( voxel_info[f'voxel_coors_stage{stage_id}'], self.sparse_shape_list[stage_id], cur_stage_downsample) # compute pooling mapping index unique_batch_win_inds, contiguous_batch_win_inds = torch.unique( batch_win_inds, return_inverse=True) voxel_info[ f'pooling_mapping_index_stage{stage_id+1}'] = \ contiguous_batch_win_inds # generate empty placeholder features placeholder_prepool_feats = voxel_info['voxel_feats_stage0'].new_zeros( (len(unique_batch_win_inds), torch.prod(torch.IntTensor(cur_stage_downsample)).item(), self.dim_model[stage_id])) voxel_info[f'pooling_index_in_pool_stage{stage_id+1}'] = index_in_win voxel_info[ f'pooling_preholder_feats_stage{stage_id+1}'] = \ placeholder_prepool_feats # compute pooling coordinates unique, inverse = unique_batch_win_inds.clone( ), contiguous_batch_win_inds.clone() perm = torch.arange( inverse.size(0), dtype=inverse.dtype, device=inverse.device) inverse, perm = inverse.flip([0]), perm.flip([0]) perm = inverse.new_empty(unique.size(0)).scatter_(0, inverse, perm) pool_coors = batch_win_coors[perm] voxel_info[f'voxel_coors_stage{stage_id+1}'] = pool_coors return voxel_info def get_set(self, voxel_info, stage_id): ''' This is one of the core operation of DSVT. Given voxels' window ids and relative-coords inner window, we partition them into window-bounded and size-equivalent local sets. To make it clear and easy to follow, we do not use loop to process two shifts. Args: voxel_info (dict): The dict contains the following keys - batch_win_inds_s{i} (Tensor[float]): Windows indices of each voxel with shape (N), computed by 'window_partition'. - coors_in_win_shift{i} (Tensor[int]): Relative-coords inner window of each voxel with shape (N, 3), computed by 'window_partition'. Each row is (z, y, x). - ... Returns: See from 'forward' function. ''' batch_win_inds_shift0 = voxel_info[ f'batch_win_inds_stage{stage_id}_shift0'] coors_in_win_shift0 = voxel_info[ f'coors_in_win_stage{stage_id}_shift0'] set_voxel_inds_shift0 = self.get_set_single_shift( batch_win_inds_shift0, stage_id, shift_id=0, coors_in_win=coors_in_win_shift0) voxel_info[ f'set_voxel_inds_stage{stage_id}_shift0'] = set_voxel_inds_shift0 # compute key masks, voxel duplication must happen continuously prefix_set_voxel_inds_s0 = torch.roll( set_voxel_inds_shift0.clone(), shifts=1, dims=-1) prefix_set_voxel_inds_s0[:, :, 0] = -1 set_voxel_mask_s0 = (set_voxel_inds_shift0 == prefix_set_voxel_inds_s0) voxel_info[ f'set_voxel_mask_stage{stage_id}_shift0'] = set_voxel_mask_s0 batch_win_inds_shift1 = voxel_info[ f'batch_win_inds_stage{stage_id}_shift1'] coors_in_win_shift1 = voxel_info[ f'coors_in_win_stage{stage_id}_shift1'] set_voxel_inds_shift1 = self.get_set_single_shift( batch_win_inds_shift1, stage_id, shift_id=1, coors_in_win=coors_in_win_shift1) voxel_info[ f'set_voxel_inds_stage{stage_id}_shift1'] = set_voxel_inds_shift1 # compute key masks, voxel duplication must happen continuously prefix_set_voxel_inds_s1 = torch.roll( set_voxel_inds_shift1.clone(), shifts=1, dims=-1) prefix_set_voxel_inds_s1[:, :, 0] = -1 set_voxel_mask_s1 = (set_voxel_inds_shift1 == prefix_set_voxel_inds_s1) voxel_info[ f'set_voxel_mask_stage{stage_id}_shift1'] = set_voxel_mask_s1 return voxel_info def get_set_single_shift(self, batch_win_inds, stage_id, shift_id=None, coors_in_win=None): device = batch_win_inds.device # the number of voxels assigned to a set voxel_num_set = self.set_info[stage_id][0] # max number of voxels in a window max_voxel = self.window_shape[stage_id][shift_id][ 0] * self.window_shape[stage_id][shift_id][1] * self.window_shape[ stage_id][shift_id][2] # get unique set indices contiguous_win_inds = torch.unique( batch_win_inds, return_inverse=True)[1] voxelnum_per_win = torch.bincount(contiguous_win_inds) win_num = voxelnum_per_win.shape[0] setnum_per_win_float = voxelnum_per_win / voxel_num_set setnum_per_win = torch.ceil(setnum_per_win_float).long() set_win_inds, set_inds_in_win = get_continous_inds(setnum_per_win) # compution of Eq.3 in 'DSVT: Dynamic Sparse Voxel Transformer with # Rotated Sets' - https://arxiv.org/abs/2301.06051, # for each window, we can get voxel indices belong to different sets. offset_idx = set_inds_in_win[:, None].repeat( 1, voxel_num_set) * voxel_num_set base_idx = torch.arange(0, voxel_num_set, 1, device=device) base_select_idx = offset_idx + base_idx base_select_idx = base_select_idx * voxelnum_per_win[ set_win_inds][:, None] base_select_idx = base_select_idx.double() / ( setnum_per_win[set_win_inds] * voxel_num_set)[:, None].double() base_select_idx = torch.floor(base_select_idx) # obtain unique indices in whole space select_idx = base_select_idx select_idx = select_idx + set_win_inds.view(-1, 1) * max_voxel # this function will return unordered inner window indices of # each voxel inner_voxel_inds = get_inner_win_inds_cuda(contiguous_win_inds) global_voxel_inds = contiguous_win_inds * max_voxel + inner_voxel_inds _, order1 = torch.sort(global_voxel_inds) # get y-axis partition results global_voxel_inds_sorty = contiguous_win_inds * max_voxel + \ coors_in_win[:, 1] * self.window_shape[stage_id][shift_id][0] * \ self.window_shape[stage_id][shift_id][2] + coors_in_win[:, 2] * \ self.window_shape[stage_id][shift_id][2] + \ coors_in_win[:, 0] _, order2 = torch.sort(global_voxel_inds_sorty) inner_voxel_inds_sorty = -torch.ones_like(inner_voxel_inds) inner_voxel_inds_sorty.scatter_( dim=0, index=order2, src=inner_voxel_inds[order1] ) # get y-axis ordered inner window indices of each voxel voxel_inds_in_batch_sorty = inner_voxel_inds_sorty + max_voxel * \ contiguous_win_inds voxel_inds_padding_sorty = -1 * torch.ones( (win_num * max_voxel), dtype=torch.long, device=device) voxel_inds_padding_sorty[voxel_inds_in_batch_sorty] = torch.arange( 0, voxel_inds_in_batch_sorty.shape[0], dtype=torch.long, device=device) set_voxel_inds_sorty = voxel_inds_padding_sorty[select_idx.long()] # get x-axis partition results global_voxel_inds_sortx = contiguous_win_inds * max_voxel + \ coors_in_win[:, 2] * self.window_shape[stage_id][shift_id][1] * \ self.window_shape[stage_id][shift_id][2] + \ coors_in_win[:, 1] * self.window_shape[stage_id][shift_id][2] + \ coors_in_win[:, 0] _, order2 = torch.sort(global_voxel_inds_sortx) inner_voxel_inds_sortx = -torch.ones_like(inner_voxel_inds) inner_voxel_inds_sortx.scatter_( dim=0, index=order2, src=inner_voxel_inds[order1] ) # get x-axis ordered inner window indices of each voxel voxel_inds_in_batch_sortx = inner_voxel_inds_sortx + max_voxel * \ contiguous_win_inds voxel_inds_padding_sortx = -1 * torch.ones( (win_num * max_voxel), dtype=torch.long, device=device) voxel_inds_padding_sortx[voxel_inds_in_batch_sortx] = torch.arange( 0, voxel_inds_in_batch_sortx.shape[0], dtype=torch.long, device=device) set_voxel_inds_sortx = voxel_inds_padding_sortx[select_idx.long()] all_set_voxel_inds = torch.stack( (set_voxel_inds_sorty, set_voxel_inds_sortx), dim=0) return all_set_voxel_inds @torch.no_grad() def window_partition(self, voxel_info, stage_id): for i in range(2): batch_win_inds, coors_in_win = get_window_coors( voxel_info[f'voxel_coors_stage{stage_id}'], self.sparse_shape_list[stage_id], self.window_shape[stage_id][i], i == 1, self.shift_list[stage_id][i]) voxel_info[ f'batch_win_inds_stage{stage_id}_shift{i}'] = batch_win_inds voxel_info[f'coors_in_win_stage{stage_id}_shift{i}'] = coors_in_win return voxel_info def get_pos_embed(self, coors_in_win, stage_id, block_id, shift_id): ''' Args: coors_in_win: shape=[N, 3], order: z, y, x ''' # [N,] window_shape = self.window_shape[stage_id][shift_id] embed_layer = self.posembed_layers[stage_id][block_id][shift_id] if len(window_shape) == 2: ndim = 2 win_x, win_y = window_shape win_z = 0 elif window_shape[-1] == 1: ndim = 2 win_x, win_y = window_shape[:2] win_z = 0 else: win_x, win_y, win_z = window_shape ndim = 3 assert coors_in_win.size(1) == 3 z, y, x = coors_in_win[:, 0] - win_z / 2,\ coors_in_win[:, 1] - win_y / 2,\ coors_in_win[:, 2] - win_x / 2 if self.normalize_pos: x = x / win_x * 2 * 3.1415 # [-pi, pi] y = y / win_y * 2 * 3.1415 # [-pi, pi] z = z / win_z * 2 * 3.1415 # [-pi, pi] if ndim == 2: location = torch.stack((x, y), dim=-1) else: location = torch.stack((x, y, z), dim=-1) pos_embed = embed_layer(location) return pos_embed