| 68 | self.unet = Unet() |
| 69 | |
| 70 | def forward(self, x, scale=[4, 2, 1], timestep=0.5, returnflow=False): |
| 71 | timestep = (x[:, :1].clone() * 0 + 1) * timestep |
| 72 | img0 = x[:, :3] |
| 73 | img1 = x[:, 3:6] |
| 74 | gt = x[:, 6:] # In inference time, gt is None |
| 75 | flow_list = [] |
| 76 | merged = [] |
| 77 | mask_list = [] |
| 78 | warped_img0 = img0 |
| 79 | warped_img1 = img1 |
| 80 | flow = None |
| 81 | loss_distill = 0 |
| 82 | stu = [self.block0, self.block1, self.block2] |
| 83 | for i in range(3): |
| 84 | if flow != None: |
| 85 | flow_d, mask_d = stu[i]( |
| 86 | torch.cat((img0, img1, timestep, warped_img0, warped_img1, mask), 1), flow, scale=scale[i] |
| 87 | ) |
| 88 | flow = flow + flow_d |
| 89 | mask = mask + mask_d |
| 90 | else: |
| 91 | flow, mask = stu[i](torch.cat((img0, img1, timestep), 1), None, scale=scale[i]) |
| 92 | mask_list.append(torch.sigmoid(mask)) |
| 93 | flow_list.append(flow) |
| 94 | warped_img0 = warp(img0, flow[:, :2]) |
| 95 | warped_img1 = warp(img1, flow[:, 2:4]) |
| 96 | merged_student = (warped_img0, warped_img1) |
| 97 | merged.append(merged_student) |
| 98 | if gt.shape[1] == 3: |
| 99 | flow_d, mask_d = self.block_tea( |
| 100 | torch.cat((img0, img1, timestep, warped_img0, warped_img1, mask, gt), 1), flow, scale=1 |
| 101 | ) |
| 102 | flow_teacher = flow + flow_d |
| 103 | warped_img0_teacher = warp(img0, flow_teacher[:, :2]) |
| 104 | warped_img1_teacher = warp(img1, flow_teacher[:, 2:4]) |
| 105 | mask_teacher = torch.sigmoid(mask + mask_d) |
| 106 | merged_teacher = warped_img0_teacher * mask_teacher + warped_img1_teacher * (1 - mask_teacher) |
| 107 | else: |
| 108 | flow_teacher = None |
| 109 | merged_teacher = None |
| 110 | for i in range(3): |
| 111 | merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i]) |
| 112 | if gt.shape[1] == 3: |
| 113 | loss_mask = ( |
| 114 | ((merged[i] - gt).abs().mean(1, True) > (merged_teacher - gt).abs().mean(1, True) + 0.01) |
| 115 | .float() |
| 116 | .detach() |
| 117 | ) |
| 118 | loss_distill += (((flow_teacher.detach() - flow_list[i]) ** 2).mean(1, True) ** 0.5 * loss_mask).mean() |
| 119 | if returnflow: |
| 120 | return flow |
| 121 | else: |
| 122 | c0 = self.contextnet(img0, flow[:, :2]) |
| 123 | c1 = self.contextnet(img1, flow[:, 2:4]) |
| 124 | tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1) |
| 125 | res = tmp[:, :3] * 2 - 1 |
| 126 | merged[2] = torch.clamp(merged[2] + res, 0, 1) |
| 127 | return flow_list, mask_list[2], merged, flow_teacher, merged_teacher, loss_distill |