thirdparty/learning3d/utils/curvenet_util.py

"""
@Author: Yue Wang
@Contact: yuewangx@mit.edu
@File: pointnet_util.py
@Time: 2018/10/13 10:39 PM

Modified by 
@Author: Tiange Xiang
@Contact: txia7609@uni.sydney.edu.au
@Time: 2021/01/21 3:10 PM
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from time import time
import numpy as np
from .model_common_utils import (
    knn,
    square_distance,
    index_points,
    farthest_point_sample,
    query_ball_point,
)

def sample_and_group(npoint, radius, nsample, xyz, points, returnfps=False):
    """
    Input:
        npoint:
        radius:
        nsample:
        xyz: input points position data, [B, N, 3]
        points: input points data, [B, N, D]
    Return:
        new_xyz: sampled points position data, [B, npoint, nsample, 3]
        new_points: sampled points data, [B, npoint, nsample, 3+D]
    """
    new_xyz = index_points(xyz, farthest_point_sample(xyz, npoint, start_with_first_point=True))
    torch.cuda.empty_cache()

    idx = query_ball_point(radius, nsample, xyz, new_xyz, get_cnt=False)
    torch.cuda.empty_cache()

    new_points = index_points(points, idx)
    torch.cuda.empty_cache()

    if returnfps:
        return new_xyz, new_points, idx
    else:
        return new_xyz, new_points

def batched_index_select(input, dim, index):
	views = [input.shape[0]] + \
		[1 if i != dim else -1 for i in range(1, len(input.shape))]
	expanse = list(input.shape)
	expanse[0] = -1
	expanse[dim] = -1
	index = index.view(views).expand(expanse)
	return torch.gather(input, dim, index)

def gumbel_softmax(logits, dim, temperature=1):
    """
    ST-gumple-softmax w/o random gumbel samplings
    input: [*, n_class]
    return: flatten --> [*, n_class] an one-hot vector
    """
    y = F.softmax(logits / temperature, dim=dim)

    shape = y.size()
    _, ind = y.max(dim=-1)
    y_hard = torch.zeros_like(y).view(-1, shape[-1])
    y_hard.scatter_(1, ind.view(-1, 1), 1)
    y_hard = y_hard.view(*shape)

    y_hard = (y_hard - y).detach() + y
    return y_hard

class Walk(nn.Module):
    '''
    Walk in the cloud
    '''
    def __init__(self, in_channel, k, curve_num, curve_length):
        super(Walk, self).__init__()
        self.curve_num = curve_num
        self.curve_length = curve_length
        self.k = k

        self.agent_mlp = nn.Sequential(
            nn.Conv2d(in_channel * 2,
                        1,
                        kernel_size=1,
                        bias=False), nn.BatchNorm2d(1))
        self.momentum_mlp = nn.Sequential(
            nn.Conv1d(in_channel * 2,
                        2,
                        kernel_size=1,
                        bias=False), nn.BatchNorm1d(2))

    def crossover_suppression(self, cur, neighbor, bn, n, k):
        # cur: bs*n, 3
        # neighbor: bs*n, 3, k
        neighbor = neighbor.detach()
        cur = cur.unsqueeze(-1).detach()
        dot = torch.bmm(cur.transpose(1,2), neighbor) # bs*n, 1, k
        norm1 = torch.norm(cur, dim=1, keepdim=True)
        norm2 = torch.norm(neighbor, dim=1, keepdim=True)
        divider = torch.clamp(norm1 * norm2, min=1e-8)
        ans = torch.div(dot, divider).squeeze() # bs*n, k

        # normalize to [0, 1]
        ans = 1. + ans
        ans = torch.clamp(ans, 0., 1.0)

        return ans.detach()

    def forward(self, xyz, x, adj, cur):
        bn, c, tot_points = x.size()
        device = x.device

        # raw point coordinates
        xyz = xyz.transpose(1,2).contiguous # bs, n, 3

        # point features
        x = x.transpose(1,2).contiguous() # bs, n, c

        flatten_x = x.view(bn * tot_points, -1)
        batch_offset = torch.arange(0, bn, device=device).detach() * tot_points

        # indices of neighbors for the starting points
        tmp_adj = (adj + batch_offset.view(-1,1,1)).view(adj.size(0)*adj.size(1),-1) #bs, n, k
    
        # batch flattened indices for teh starting points
        flatten_cur = (cur + batch_offset.view(-1,1,1)).view(-1)

        curves = []
        flatten_curve_idxs = [flatten_cur.unsqueeze(1)]

        # one step at a time
        for step in range(self.curve_length):

            if step == 0:
                # get starting point features using flattend indices
                starting_points =  flatten_x[flatten_cur, :].contiguous()
                pre_feature = starting_points.view(bn, self.curve_num, -1, 1).transpose(1,2) # bs * n, c
            else:
                # dynamic momentum
                cat_feature = torch.cat((cur_feature.squeeze(-1), pre_feature.squeeze(-1)),dim=1)
                att_feature = F.softmax(self.momentum_mlp(cat_feature),dim=1).view(bn, 1, self.curve_num, 2) # bs, 1, n, 2
                cat_feature = torch.cat((cur_feature, pre_feature),dim=-1) # bs, c, n, 2
                
                # update curve descriptor
                pre_feature = torch.sum(cat_feature * att_feature, dim=-1, keepdim=True) # bs, c, n
                pre_feature_cos =  pre_feature.transpose(1,2).contiguous().view(bn * self.curve_num, -1)

            pick_idx = tmp_adj[flatten_cur] # bs*n, k
            
            # get the neighbors of current points
            pick_values = flatten_x[pick_idx.view(-1),:]

            # reshape to fit crossover suppresion below
            pick_values_cos = pick_values.view(bn * self.curve_num, self.k, c)
            pick_values = pick_values_cos.view(bn, self.curve_num, self.k, c)
            pick_values_cos = pick_values_cos.transpose(1,2).contiguous()
            
            pick_values = pick_values.permute(0,3,1,2) # bs, c, n, k

            pre_feature_expand = pre_feature.expand_as(pick_values)
            
            # concat current point features with curve descriptors
            pre_feature_expand = torch.cat((pick_values, pre_feature_expand),dim=1)
            
            # which node to pick next?
            pre_feature_expand = self.agent_mlp(pre_feature_expand) # bs, 1, n, k

            if step !=0:
                # cross over supression
                d = self.crossover_suppression(cur_feature_cos - pre_feature_cos,
                                               pick_values_cos - cur_feature_cos.unsqueeze(-1), 
                                               bn, self.curve_num, self.k)
                d = d.view(bn, self.curve_num, self.k).unsqueeze(1) # bs, 1, n, k
                pre_feature_expand = torch.mul(pre_feature_expand, d)

            pre_feature_expand = gumbel_softmax(pre_feature_expand, -1) #bs, 1, n, k

            cur_feature = torch.sum(pick_values * pre_feature_expand, dim=-1, keepdim=True) # bs, c, n, 1

            cur_feature_cos = cur_feature.transpose(1,2).contiguous().view(bn * self.curve_num, c)

            cur = torch.argmax(pre_feature_expand, dim=-1).view(-1, 1) # bs * n, 1

            flatten_cur = batched_index_select(pick_idx, 1, cur).squeeze() # bs * n

            # collect curve progress
            curves.append(cur_feature)
            flatten_curve_idxs.append(flatten_cur.unsqueeze(1))

        return torch.cat(curves,dim=-1), torch.cat(flatten_curve_idxs, dim=1)


class Attention_block(nn.Module):
    '''
    Used in attention U-Net.
    '''
    def __init__(self,F_g,F_l,F_int):
        super(Attention_block,self).__init__()
        self.W_g = nn.Sequential(
            nn.Conv1d(F_g, F_int, kernel_size=1,stride=1,padding=0,bias=True),
            nn.BatchNorm1d(F_int)
            )

        self.W_x = nn.Sequential(
            nn.Conv1d(F_l, F_int, kernel_size=1,stride=1,padding=0,bias=True),
            nn.BatchNorm1d(F_int)
        )

        self.psi = nn.Sequential(
            nn.Conv1d(F_int, 1, kernel_size=1,stride=1,padding=0,bias=True),
            nn.BatchNorm1d(1),
            nn.Sigmoid()
        )

    def forward(self,g,x):
        g1 = self.W_g(g)
        x1 = self.W_x(x)
        psi = F.leaky_relu(g1+x1, negative_slope=0.2)
        psi = self.psi(psi)

        return psi, 1. - psi


class LPFA(nn.Module):
    def __init__(self, in_channel, out_channel, k, mlp_num=2, initial=False):
        super(LPFA, self).__init__()
        self.k = k
        self.initial = initial

        if not initial:
            self.xyz2feature = nn.Sequential(
                        nn.Conv2d(9, in_channel, kernel_size=1, bias=False),
                        nn.BatchNorm2d(in_channel))

        self.mlp = []
        for _ in range(mlp_num):
            self.mlp.append(nn.Sequential(nn.Conv2d(in_channel, out_channel, 1, bias=False),
                                 nn.BatchNorm2d(out_channel),
                                 nn.LeakyReLU(0.2)))
            in_channel = out_channel
        self.mlp = nn.Sequential(*self.mlp)        

    def forward(self, x, xyz, idx=None):
        x = self.group_feature(x, xyz, idx)
        x = self.mlp(x)

        if self.initial:
            x = x.max(dim=-1, keepdim=False)[0]
        else:
            x = x.mean(dim=-1, keepdim=False)

        return x

    def group_feature(self, x, xyz, idx):
        batch_size, num_dims, num_points = x.size()
        device = x.device

        if idx is None:
            idx = knn(xyz, k=self.k, add_one_to_k=True)[:,:,:self.k]  # (batch_size, num_points, k)

        idx_base = torch.arange(0, batch_size, device=device).view(-1, 1, 1) * num_points
        idx = idx + idx_base
        idx = idx.view(-1)

        xyz = xyz.transpose(2, 1).contiguous() # bs, n, 3
        point_feature = xyz.view(batch_size * num_points, -1)[idx, :]
        point_feature = point_feature.view(batch_size, num_points, self.k, -1)  # bs, n, k, 3
        points = xyz.view(batch_size, num_points, 1, 3).expand(-1, -1, self.k, -1)  # bs, n, k, 3

        point_feature = torch.cat((points, point_feature, point_feature - points),
                                dim=3).permute(0, 3, 1, 2).contiguous()

        if self.initial:
            return point_feature

        x = x.transpose(2, 1).contiguous() # bs, n, c
        feature = x.view(batch_size * num_points, -1)[idx, :]
        feature = feature.view(batch_size, num_points, self.k, num_dims)  #bs, n, k, c
        x = x.view(batch_size, num_points, 1, num_dims)
        feature = feature - x

        feature = feature.permute(0, 3, 1, 2).contiguous()
        point_feature = self.xyz2feature(point_feature)  #bs, c, n, k
        feature = F.leaky_relu(feature + point_feature, 0.2)
        return feature #bs, c, n, k


class PointNetFeaturePropagation(nn.Module):
    def __init__(self, in_channel, mlp, att=None):
        super(PointNetFeaturePropagation, self).__init__()
        self.mlp_convs = nn.ModuleList()
        self.mlp_bns = nn.ModuleList()
        last_channel = in_channel
        self.att = None
        if att is not None:
            self.att = Attention_block(F_g=att[0],F_l=att[1],F_int=att[2])
        
        for out_channel in mlp:
            self.mlp_convs.append(nn.Conv1d(last_channel, out_channel, 1))
            self.mlp_bns.append(nn.BatchNorm1d(out_channel))
            last_channel = out_channel

    def forward(self, xyz1, xyz2, points1, points2):
        """
        Input:
            xyz1: input points position data, [B, C, N]
            xyz2: sampled input points position data, [B, C, S], skipped xyz
            points1: input points data, [B, D, N]
            points2: input points data, [B, D, S], skipped features
        Return:
            new_points: upsampled points data, [B, D', N]
        """
        xyz1 = xyz1.permute(0, 2, 1)
        xyz2 = xyz2.permute(0, 2, 1)

        points2 = points2.permute(0, 2, 1)
        B, N, C = xyz1.shape
        _, S, _ = xyz2.shape

        if S == 1:
            interpolated_points = points2.repeat(1, N, 1)
        else:
            dists = square_distance(xyz1, xyz2)
            dists, idx = dists.sort(dim=-1)
            dists, idx = dists[:, :, :3], idx[:, :, :3]  # [B, N, 3]

            dist_recip = 1.0 / (dists + 1e-8)
            norm = torch.sum(dist_recip, dim=2, keepdim=True)
            weight = dist_recip / norm
            interpolated_points = torch.sum(index_points(points2, idx) * weight.view(B, N, 3, 1), dim=2)

        # skip attention
        if self.att is not None:
           psix, psig = self.att(interpolated_points.permute(0, 2, 1), points1)
           points1 = points1 * psix
           
        if points1 is not None:
            points1 = points1.permute(0, 2, 1)
            new_points = torch.cat([points1, interpolated_points], dim=-1)
        else:
            new_points = interpolated_points

        new_points = new_points.permute(0, 2, 1)

        for i, conv in enumerate(self.mlp_convs):
            bn = self.mlp_bns[i]
            new_points = F.leaky_relu(bn(conv(new_points)), 0.2)

        return new_points


class CIC(nn.Module):
    def __init__(self, npoint, radius, k, in_channels, output_channels, bottleneck_ratio=2, mlp_num=2, curve_config=None):
        super(CIC, self).__init__()
        self.in_channels = in_channels
        self.output_channels = output_channels
        self.bottleneck_ratio = bottleneck_ratio
        self.radius = radius
        self.k = k
        self.npoint = npoint

        planes = in_channels // bottleneck_ratio

        self.use_curve = curve_config is not None
        if self.use_curve:
            self.curveaggregation = CurveAggregation(planes)
            self.curvegrouping = CurveGrouping(planes, k, curve_config[0], curve_config[1])

        self.conv1 = nn.Sequential(
            nn.Conv1d(in_channels,
                      planes,
                      kernel_size=1,
                      bias=False),
            nn.BatchNorm1d(in_channels // bottleneck_ratio),
            nn.LeakyReLU(negative_slope=0.2, inplace=True))

        self.conv2 = nn.Sequential(
            nn.Conv1d(planes, output_channels, kernel_size=1, bias=False),
            nn.BatchNorm1d(output_channels))

        if in_channels != output_channels:
            self.shortcut = nn.Sequential(
                nn.Conv1d(in_channels,
                          output_channels,
                          kernel_size=1,
                          bias=False),
                nn.BatchNorm1d(output_channels))

        self.relu = nn.LeakyReLU(negative_slope=0.2, inplace=True)

        self.maxpool = MaskedMaxPool(npoint, radius, k)

        self.lpfa = LPFA(planes, planes, k, mlp_num=mlp_num, initial=False)

    def forward(self, xyz, x):
 
        # max pool
        if xyz.size(-1) != self.npoint:
            xyz, x = self.maxpool(
                xyz.transpose(1, 2).contiguous(), x)
            xyz = xyz.transpose(1, 2)

        shortcut = x
        x = self.conv1(x)  # bs, c', n

        idx = knn(xyz, self.k, add_one_to_k=True)

        if self.use_curve:
            # curve grouping
            curves, flatten_curve_idxs = self.curvegrouping(x, xyz, idx[:,:,1:]) # avoid self-loop

            # curve aggregation
            x = self.curveaggregation(x, curves)
        else:
            flatten_curve_idxs = None

        x = self.lpfa(x, xyz, idx=idx[:,:,:self.k]) #bs, c', n, k

        x = self.conv2(x)  # bs, c, n

        if self.in_channels != self.output_channels:
            shortcut = self.shortcut(shortcut)

        x = self.relu(x + shortcut)
        return xyz, x, flatten_curve_idxs


class CurveAggregation(nn.Module):
    def __init__(self, in_channel):
        super(CurveAggregation, self).__init__()
        self.in_channel = in_channel
        mid_feature = in_channel // 2
        self.conva = nn.Conv1d(in_channel,
                               mid_feature,
                               kernel_size=1,
                               bias=False)
        self.convb = nn.Conv1d(in_channel,
                               mid_feature,
                               kernel_size=1,
                               bias=False)
        self.convc = nn.Conv1d(in_channel,
                               mid_feature,
                               kernel_size=1,
                               bias=False)
        self.convn = nn.Conv1d(mid_feature,
                               mid_feature,
                               kernel_size=1,
                               bias=False)
        self.convl = nn.Conv1d(mid_feature,
                               mid_feature,
                               kernel_size=1,
                               bias=False)
        self.convd = nn.Sequential(
            nn.Conv1d(mid_feature * 2,
                      in_channel,
                      kernel_size=1,
                      bias=False),
            nn.BatchNorm1d(in_channel))
        self.line_conv_att = nn.Conv2d(in_channel,
                                       1,
                                       kernel_size=1,
                                       bias=False)

    def forward(self, x, curves):
        curves_att = self.line_conv_att(curves)  # bs, 1, c_n, c_l

        curver_inter = torch.sum(curves * F.softmax(curves_att, dim=-1), dim=-1)  #bs, c, c_n
        curves_intra = torch.sum(curves * F.softmax(curves_att, dim=-2), dim=-2)  #bs, c, c_l

        curver_inter = self.conva(curver_inter) # bs, mid, n
        curves_intra = self.convb(curves_intra) # bs, mid ,n

        x_logits = self.convc(x).transpose(1, 2).contiguous()
        x_inter = F.softmax(torch.bmm(x_logits, curver_inter), dim=-1) # bs, n, c_n
        x_intra = F.softmax(torch.bmm(x_logits, curves_intra), dim=-1) # bs, l, c_l
        

        curver_inter = self.convn(curver_inter).transpose(1, 2).contiguous()
        curves_intra = self.convl(curves_intra).transpose(1, 2).contiguous()

        x_inter = torch.bmm(x_inter, curver_inter)
        x_intra = torch.bmm(x_intra, curves_intra)

        curve_features = torch.cat((x_inter, x_intra),dim=-1).transpose(1, 2).contiguous()
        x = x + self.convd(curve_features)

        return F.leaky_relu(x, negative_slope=0.2)


class CurveGrouping(nn.Module):
    def __init__(self, in_channel, k, curve_num, curve_length):
        super(CurveGrouping, self).__init__()
        self.curve_num = curve_num
        self.curve_length = curve_length
        self.in_channel = in_channel
        self.k = k

        self.att = nn.Conv1d(in_channel, 1, kernel_size=1, bias=False)

        self.walk = Walk(in_channel, k, curve_num, curve_length)

    def forward(self, x, xyz, idx):
        # starting point selection in self attention style
        x_att = torch.sigmoid(self.att(x))
        x = x * x_att

        _, start_index = torch.topk(x_att,
                                    self.curve_num,
                                    dim=2,
                                    sorted=False)
        start_index = start_index.squeeze(1).unsqueeze(2)

        curves, flatten_curve_idxs = self.walk(xyz, x, idx, start_index)  #bs, c, c_n, c_l
        
        return curves, flatten_curve_idxs


class MaskedMaxPool(nn.Module):
    def __init__(self, npoint, radius, k):
        super(MaskedMaxPool, self).__init__()
        self.npoint = npoint
        self.radius = radius
        self.k = k

    def forward(self, xyz, features):
        sub_xyz, neighborhood_features = sample_and_group(self.npoint, self.radius, self.k, xyz, features.transpose(1,2))

        neighborhood_features = neighborhood_features.permute(0, 3, 1, 2).contiguous()
        sub_features = F.max_pool2d(
            neighborhood_features, kernel_size=[1, neighborhood_features.shape[3]]
        )  # bs, c, n, 1
        sub_features = torch.squeeze(sub_features, -1)  # bs, c, n
        return sub_xyz, sub_features