import torch import torch.nn as nn import os import numpy as np import os.path as osp import matplotlib.pyplot as plt from ..base_model import BaseModel from ..modules import SetBlockWrapper, HorizontalPoolingPyramid, PackSequenceWrapper, SeparateFCs, SeparateBNNecks, conv1x1, conv3x3, BasicBlock2D, BasicBlockP3D, BasicBlock3D import torch.nn.functional as F from einops import rearrange import copy import cv2 from kornia import morphology as morph blocks_map = { '2d': BasicBlock2D, 'p3d': BasicBlockP3D, '3d': BasicBlock3D } class MultiGaitpp(BaseModel): def build_network(self, model_cfg): in_C, B, C = model_cfg['Backbone']['in_channels'], model_cfg['Backbone']['blocks'], model_cfg['Backbone']['C'] self.part1 = model_cfg['Backbone']['part1_channel'] self.part2 = model_cfg['Backbone']['part2_channel'] self.inplanes = 32 * C self.part1_layer0 = SetBlockWrapper(nn.Sequential( conv3x3(self.part1, self.inplanes, 1), nn.BatchNorm2d(self.inplanes), nn.ReLU(inplace=True) )) self.part2_layer0 = SetBlockWrapper(nn.Sequential( conv3x3(self.part2, self.inplanes, 1), nn.BatchNorm2d(self.inplanes), nn.ReLU(inplace=True) )) self.part1_layer1 = SetBlockWrapper(self.make_layer(BasicBlock2D, 32 * C, stride=[1, 1], blocks_num=B[0], mode='2d')) self.part2_layer1 = copy.deepcopy(self.part1_layer1) self.fusion = CatFusion(256) self.part1_layer2 = self.make_layer(BasicBlockP3D, 64 * C, stride=[2, 2], blocks_num=B[1], mode='p3d') self.part2_layer2 = copy.deepcopy(self.part1_layer2) self.layer2 = copy.deepcopy(self.part1_layer2) self.part1_layer3 = self.make_layer(BasicBlockP3D, 128 * C, stride=[2, 2], blocks_num=B[2], mode='p3d') self.part2_layer3 = copy.deepcopy(self.part1_layer3) self.layer3 = copy.deepcopy(self.part1_layer3) self.layer4 = self.make_layer(BasicBlockP3D, 256 * C, stride=[1, 1], blocks_num=B[3], mode='p3d') self.crossattn1 = CrossAttention(64) self.FCs = SeparateFCs(16, 256*C, 128*C) self.BNNecks = SeparateBNNecks(16, 128*C, class_num=model_cfg['SeparateBNNecks']['class_num']) self.TP = PackSequenceWrapper(torch.max) self.HPP = HorizontalPoolingPyramid(bin_num=[16]) def make_layer(self, block, planes, stride, blocks_num, mode='2d'): if max(stride) > 1 or self.inplanes != planes * block.expansion: if mode == '3d': downsample = nn.Sequential(nn.Conv3d(self.inplanes, planes * block.expansion, kernel_size=[1, 1, 1], stride=stride, padding=[0, 0, 0], bias=False), nn.BatchNorm3d(planes * block.expansion)) elif mode == '2d': downsample = nn.Sequential(conv1x1(self.inplanes, planes * block.expansion, stride=stride), nn.BatchNorm2d(planes * block.expansion)) elif mode == 'p3d': downsample = nn.Sequential(nn.Conv3d(self.inplanes, planes * block.expansion, kernel_size=[1, 1, 1], stride=[1, *stride], padding=[0, 0, 0], bias=False), nn.BatchNorm3d(planes * block.expansion)) else: raise TypeError('xxx') else: downsample = lambda x: x layers = [block(self.inplanes, planes, stride=stride, downsample=downsample)] self.inplanes = planes * block.expansion s = [1, 1] if mode in ['2d', 'p3d'] else [1, 1, 1] for i in range(1, blocks_num): layers.append( block(self.inplanes, planes, stride=s) ) return nn.Sequential(*layers) def forward(self, inputs): ipts, labs, _, _, seqL = inputs if len(ipts[0].size()) == 4: ipts = ipts[0].unsqueeze(1) else: ipts = ipts[0].transpose(1, 2).contiguous() part1 = ipts[:, :self.part1, ...] part2 = ipts[:, self.part1:, ...] del ipts part2 = self.part2_layer0(part2) part2 = self.part2_layer1(part2) part1 = self.part1_layer0(part1) part1 = self.part1_layer1(part1) out, attn1, attn2, attn_co = self.crossattn1(part2,part1) part2 = self.part2_layer2(part2*attn1) part1 = self.part1_layer2(part1*attn2) out = self.layer2(out) part2 = self.part2_layer3(part2) part1 = self.part1_layer3(part1) out = self.layer3(out) out = self.fusion([part1, out, part2]) out = self.layer4(out) outs = self.TP(out, seqL, options={"dim": 2})[0] # [n, c, h, w] feat = self.HPP(outs) # [n, c, p] embed_1 = self.FCs(feat) # [n, c, p] embed_2, logits = self.BNNecks(embed_1) # [n, c, p] embed = embed_1 retval = { 'training_feat': { 'triplet': {'embeddings': embed_1, 'labels': labs}, 'softmax': {'logits': logits, 'labels': labs} }, 'visual_summary': { }, 'inference_feat': { 'embeddings': embed } } return retval class CatFusion(nn.Module): def __init__(self, in_channels=64): super(CatFusion, self).__init__() self.conv = SetBlockWrapper( nn.Sequential( conv1x1(in_channels * 3, in_channels), ) ) def forward(self, feat_list): ''' sil_feat: [n, c, s, h, w] map_feat: [n, c, s, h, w] ''' # print(feat_list.shape) feats = torch.cat(feat_list, dim=1) retun = self.conv(feats) return retun class CrossAttention(nn.Module): def __init__(self, in_channels=64, squeeze_ratio=16, h=32, w=22): super(CrossAttention, self).__init__() hidden_dim = int(in_channels / squeeze_ratio) self.TP_mean = PackSequenceWrapper(torch.mean) self.conv2 = SetBlockWrapper(nn.Sequential( conv1x1(in_channels, hidden_dim), nn.BatchNorm2d(hidden_dim), nn.ReLU(inplace=True), conv1x1(hidden_dim, hidden_dim), nn.BatchNorm2d(hidden_dim), nn.ReLU(inplace=True), conv1x1(hidden_dim, in_channels), )) self.conv1 = SetBlockWrapper(nn.Sequential( conv1x1(in_channels, hidden_dim), nn.BatchNorm2d(hidden_dim), nn.ReLU(inplace=True), conv1x1(hidden_dim, hidden_dim), nn.BatchNorm2d(hidden_dim), nn.ReLU(inplace=True), conv1x1(hidden_dim, in_channels), )) self.kernel = torch.ones((3,3)) def channel_normalization(self, masked_attn): min_vals = masked_attn.min(dim=1, keepdim=True).values max_vals = masked_attn.max(dim=1, keepdim=True).values min_vals = min_vals.expand_as(masked_attn) max_vals = max_vals.expand_as(masked_attn) attn_norm = (masked_attn - min_vals) / (max_vals - min_vals + 1e-6) attn_norm = attn_norm.clamp(0, 1) return attn_norm def forward(self, x1, x2): ''' x1 [n, c, s, h, w] x2 [n, c, s, h, w] shape ''' t = x2.size(2) attn_x2 = self.conv2(x2) n, c, t, h, w = attn_x2.size() attn_x1 = self.conv1(x1) # [n, c, h, w] attn_x = torch.stack((attn_x1, attn_x2), dim=1) attn_x = F.softmax(attn_x, dim=1) attn_x1_softmax = attn_x[:, 0, ...] attn_x2_softmax = attn_x[:, 1, ...] attn_ = torch.min(attn_x1_softmax,attn_x2_softmax) #* mask attn = self.channel_normalization(attn_) attn_co = rearrange(attn, 'n c s h w -> (n s) c h w') return (x1+x2)/2 * attn, (1.-attn)*attn_x1_softmax, (1.-attn)*attn_x2_softmax, attn_co #87.2