(self, cfg, neighbour_matrix=None)
| 114 | |
| 115 | class ScoreNet(nn.Module): |
| 116 | def __init__(self, cfg, neighbour_matrix=None): |
| 117 | super(ScoreNet, self).__init__() |
| 118 | self.joint_ch = cfg.scorenet.joint_ch |
| 119 | self.joints = cfg.hyponet.num_joints |
| 120 | self.num_twists = cfg.hyponet.num_twists |
| 121 | self.num_item = self.num_twists+self.joints |
| 122 | self.num_blocks = cfg.scorenet.num_blocks |
| 123 | self.dropout_rate = 0.25 |
| 124 | self.neighbour_matrix = neighbour_matrix |
| 125 | self.mask = self.init_mask(self.neighbour_matrix[0]) |
| 126 | self.mask_twist = self.init_mask(self.neighbour_matrix[1]) |
| 127 | self.mask_joints = self.init_mask(self.neighbour_matrix[2]) |
| 128 | self.atten_knn = cfg.scorenet.atten_knn |
| 129 | self.local_ch = cfg.hrnet.local_ch |
| 130 | self.parents_idx = torch.tensor(parents[1:1+self.num_twists]) |
| 131 | self.child_idx = torch.arange(start=1,end=1+self.num_twists) |
| 132 | self.emb_h , self.emb_w = int(cfg.hrnet.image_size[0]/ 32), int(cfg.hrnet.image_size[1]/ 32) |
| 133 | self.mask_list = [] |
| 134 | assert len(self.atten_knn) == self.num_blocks |
| 135 | for i in range(self.num_blocks): |
| 136 | mask_i = np.linalg.matrix_power(neighbour_matrix[3], self.atten_knn[i]) |
| 137 | mask_i = np.array(mask_i!=0, dtype=np.float32) |
| 138 | mask_i = self.init_mask(mask_i) |
| 139 | mask_i = 1 - mask_i |
| 140 | self.mask_list.append(mask_i.bool()) |
| 141 | self.mask_list_cross_atten = [None for i in range(self.num_blocks)] |
| 142 | |
| 143 | self.ch = self.joint_ch + self.local_ch |
| 144 | |
| 145 | |
| 146 | # first layer |
| 147 | # joints |
| 148 | self.linear_start_j = nn.Linear(self.joints*3, self.joints*self.joint_ch) |
| 149 | self.bn_start_j = nn.GroupNorm(32, num_channels=self.joints*self.joint_ch) |
| 150 | self.activation_start_j = nn.LeakyReLU(negative_slope=0.2) |
| 151 | self.dropout_start_j = nn.Dropout(p=self.dropout_rate) |
| 152 | # twist |
| 153 | self.linear_start_t = nn.Linear(self.num_twists*2, self.num_twists*self.joint_ch) |
| 154 | self.bn_start_t = nn.GroupNorm(32, num_channels=self.num_twists*self.joint_ch) |
| 155 | self.activation_start_t = nn.LeakyReLU(negative_slope=0.2) |
| 156 | self.dropout_start_t = nn.Dropout(p=self.dropout_rate) |
| 157 | |
| 158 | self.linear_mlp = nn.Sequential( |
| 159 | nn.Linear(self.num_item*self.ch, 1024), |
| 160 | nn.GroupNorm(32, num_channels=1024), |
| 161 | nn.LeakyReLU(negative_slope=0.2), |
| 162 | nn.Dropout(p=self.dropout_rate), |
| 163 | nn.Linear(1024,1) |
| 164 | ) |
| 165 | #blocks |
| 166 | self.blocks = nn.ModuleList([DecoderLayer(cfg,self.ch) for i in range(self.num_blocks)]) |
| 167 | |
| 168 | def init_mask(self,neighbour_matrix): |
| 169 | """ |
no test coverage detected