• 【手把手反内卷】开创全新AI多模态任务一视听分割：代码实践、优化教程（二）

    前言

    理论部分请看上一篇文章：

    简要概述：我们要知道图像中哪个物体在发声如下视频演示：

    gif 不能发出声音，大家脑补一下场景中有很多车，只有这辆120在发出声音，所以分割出发出声音的物体。

     

     

     这是一位歌手时而唱歌，时而弹琴场景，只弹琴时，不分割人体，唱歌时，分割人体。

     

    代码相对路径介绍（我的版本，非官方）

     

    大家可以通过下载我的百度网盘（附带全部数据和代码），也可以下载官方代码，但不含数据，只能申请得到。

    训练

    先看train.py

    看下面代码的help里面。

    1. parser.add_argument("--session_name", default="MS3", type=str, help="使用MS3是对数据里的Multi-sources下的数据进行训练，是多声源数据，也就是，可能同时有多个物体发声")
    2. parser.add_argument("--visual_backbone", default="resnet", type=str,
    3.                     help="use resnet50 or pvt-v2 as the visual backbone")
    4. ​
    5. parser.add_argument("--train_batch_size", default=4, type=int)
    6. parser.add_argument("--val_batch_size", default=1, type=int)
    7. parser.add_argument("--max_epoches", default=5, type=int)
    8. parser.add_argument("--lr", default=0.0001, type=float)
    9. parser.add_argument("--num_workers", default=0, type=int)
    10. parser.add_argument("--wt_dec", default=5-4, type=float)
    11. ​
    12. parser.add_argument('--masked_av_flag', action='store_true', default=True,
    13.                     help='使用作者论文里说的loss： sa/masked_va loss')
    14. parser.add_argument("--lambda_1", default=0.5, type=float, help='均衡系数weight for balancing l4 loss')
    15. parser.add_argument("--masked_av_stages", default=[0, 1, 2, 3], nargs='+', type=int,
    16.                     help='作者的设置compute sa/masked_va loss in which stages: [0, 1, 2, 3]')
    17. parser.add_argument('--threshold_flag', action='store_true', default=False,
    18.                     help='whether thresholding the generated masks')
    19. parser.add_argument("--mask_pooling_type", default='avg', type=str, help='the manner to downsample predicted masks')
    20. parser.add_argument('--norm_fea_flag', action='store_true', default=False, help='音频标准化normalize audio-visual features')
    21. parser.add_argument('--closer_flag', action='store_true', default=False, help='use closer loss for masked_va loss')
    22. parser.add_argument('--euclidean_flag', action='store_true', default=False,
    23.                     help='use euclidean distance for masked_va loss')
    24. parser.add_argument('--kl_flag', action='store_true', default=True, help='KL散度 use kl loss for masked_va loss')
    25. ​
    26. parser.add_argument("--load_s4_params", action='store_true', default=False,
    27.                     help='use S4 parameters for initilization')
    28. parser.add_argument("--trained_s4_model_path", type=str, default='', help='pretrained S4 model')
    29. ​
    30. parser.add_argument("--tpavi_stages", default=[0, 1, 2, 3], nargs='+', type=int,
    31.                     help='tpavi模块 add tpavi block in which stages: [0, 1, 2, 3]')
    32. parser.add_argument("--tpavi_vv_flag", action='store_true', default=False, help='视觉自注意visual-visual self-attention')
    33. parser.add_argument("--tpavi_va_flag", action='store_true', default=True, help='视听交叉注意visual-audio cross-attention')
    34. ​
    35. parser.add_argument("--weights", type=str, default='', help='初始训练预训练模型，可以不写path of trained model')
    36. parser.add_argument('--log_dir', default='./train_logs', type=str)

    大家根据train.sh就可以训练

    代码细节

    接下来会根据设置你要的视觉特征提取backbone，语音的默认使用vggish特征提取。

    1. if (args.visual_backbone).lower() == "resnet":
    2.     from model import ResNet_AVSModel as AVSModel
    3. ​
    4.     print('==> Use ResNet50 as the visual backbone...')
    5. elif (args.visual_backbone).lower() == "pvt":
    6.     from model import PVT_AVSModel as AVSModel
    7. ​
    8.     print('==> Use pvt-v2 as the visual backbone...')
    9. else:
    10.     raise NotImplementedError("only support the resnet50 and pvt-v2")

    数据读取部分：

    1. class MS3Dataset(Dataset):
    2.     """Dataset for multiple sound source segmentation"""
    3.     def __init__(self, split='train'):
    4.         super(MS3Dataset, self).__init__()
    5.         self.split = split
    6.         self.mask_num = 5
    7.         df_all = pd.read_csv(cfg.DATA.ANNO_CSV, sep=',')
    8.         self.df_split = df_all[df_all['split'] == split]
    9.         print("{}/{} videos are used for {}".format(len(self.df_split), len(df_all), self.split))
    10.         self.img_transform = transforms.Compose([
    11.             transforms.ToTensor(),
    12.             transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
    13.         ])
    14.         self.mask_transform = transforms.Compose([
    15.             transforms.ToTensor(),
    16.         ])
    17. ​
    18. ​
    19. ​
    20.     def __getitem__(self, index):
    21.         df_one_video = self.df_split.iloc[index]
    22.         video_name = df_one_video[0]
    23.         img_base_path =  os.path.join(cfg.DATA.DIR_IMG, video_name)
    24.         audio_lm_path = os.path.join(cfg.DATA.DIR_AUDIO_LOG_MEL, self.split, video_name + '.pkl')
    25.         mask_base_path = os.path.join(cfg.DATA.DIR_MASK, self.split, video_name)
    26.         audio_log_mel = load_audio_lm(audio_lm_path)
    27.         # audio_lm_tensor = torch.from_numpy(audio_log_mel)
    28.         imgs, masks = [], []
    29.         for img_id in range(1, 6):
    30.             img = load_image_in_PIL_to_Tensor(os.path.join(img_base_path, "%s.mp4_%d.png"%(video_name, img_id)), transform=self.img_transform)
    31.             imgs.append(img)
    32.         for mask_id in range(1, self.mask_num + 1):
    33.             mask = load_image_in_PIL_to_Tensor(os.path.join(mask_base_path, "%s_%d.png"%(video_name, mask_id)), transform=self.mask_transform, mode='P')
    34.             masks.append(mask)
    35.         imgs_tensor = torch.stack(imgs, dim=0)
    36.         masks_tensor = torch.stack(masks, dim=0)
    37. ​
    38.         return imgs_tensor, audio_log_mel, masks_tensor, video_name
    39. ​
    40.     def __len__(self):
    41.         return len(self.df_split)

    可以看到，一次读取5张图，我看了视频，都是5秒的，说明作者一次训练一个视频，每个视频每秒的帧和GT和语音合并训练。

    1. for n_iter, batch_data in enumerate(train_dataloader):
    2.     imgs, audio, mask, _ = batch_data  # [bs, 5, 3, 224, 224], [bs, 5, 1, 96, 64], [bs, 5 or 1, 1, 224, 224]
    3. ​
    4.     imgs = imgs.cuda()
    5.     audio = audio.cuda()
    6.     mask = mask.cuda()
    7.     B, frame, C, H, W = imgs.shape
    8.     imgs = imgs.view(B * frame, C, H, W)
    9.     mask_num = 5
    10.     mask = mask.view(B * mask_num, 1, H, W)
    11.     audio = audio.view(-1, audio.shape[2], audio.shape[3], audio.shape[4])  # [B*T, 1, 96, 64]
    12.     with torch.no_grad():
    13.         audio_feature = audio_backbone(audio)  # [B*T, 128]
    14. ​
    15.     output, v_map_list, a_fea_list = model(imgs, audio_feature)  # [bs*5, 1, 224, 224]
    16.     loss, loss_dict = IouSemanticAwareLoss(output, mask, a_fea_list, v_map_list, \
    17.                                            sa_loss_flag=args.masked_av_flag, lambda_1=args.lambda_1,
    18.                                            count_stages=args.masked_av_stages, \
    19.                                            mask_pooling_type=args.mask_pooling_type,
    20.                                            threshold=args.threshold_flag, norm_fea=args.norm_fea_flag, \
    21.                                            closer_flag=args.closer_flag, euclidean_flag=args.euclidean_flag,
    22.                                            kl_flag=args.kl_flag)
    23. ​
    24.     avg_meter_total_loss.add({'total_loss': loss.item()})
    25.     avg_meter_iou_loss.add({'iou_loss': loss_dict['iou_loss']})
    26.     avg_meter_sa_loss.add({'sa_loss': loss_dict['sa_loss']})
    27. ​
    28.     optimizer.zero_grad()
    29.     loss.backward()
    30.     optimizer.step()
    31. ​
    32.     global_step += 1
    33.     if (global_step - 1) % 20 == 0:
    34.         train_log = 'Iter:%5d/%5d, Total_Loss:%.4f, iou_loss:%.4f, sa_loss:%.4f, lr: %.4f' % (
    35.             global_step - 1, max_step, avg_meter_total_loss.pop('total_loss'),
    36.             avg_meter_iou_loss.pop('iou_loss'), avg_meter_sa_loss.pop('sa_loss'),
    37.             optimizer.param_groups[0]['lr'])

    可以看到，训练很简单，先load图像5帧view合并在一起，再获取语音特征，送入模型。然后计算损失和Iou得分。

    输入模型的数据分为两部分，图像帧【bs*5, 3, 224, 224】，乘以5意思是每个视频有5帧，第二部分是语音帧，维度相似。

    1. class Pred_endecoder(nn.Module):
    2.     # resnet based encoder decoder
    3.     def __init__(self, channel=256, config=None, tpavi_stages=[], tpavi_vv_flag=False, tpavi_va_flag=True):
    4.         super(Pred_endecoder, self).__init__()
    5.         self.cfg = config
    6.         self.tpavi_stages = tpavi_stages
    7.         self.tpavi_vv_flag = tpavi_vv_flag
    8.         self.tpavi_va_flag = tpavi_va_flag
    9. ​
    10.         self.resnet = B2_ResNet()
    11.         self.relu = nn.ReLU(inplace=True)
    12. ​
    13.         self.conv4 = self._make_pred_layer(Classifier_Module, [3, 6, 12, 18], [3, 6, 12, 18], channel, 2048)
    14.         self.conv3 = self._make_pred_layer(Classifier_Module, [3, 6, 12, 18], [3, 6, 12, 18], channel, 1024)
    15.         self.conv2 = self._make_pred_layer(Classifier_Module, [3, 6, 12, 18], [3, 6, 12, 18], channel, 512)
    16.         self.conv1 = self._make_pred_layer(Classifier_Module, [3, 6, 12, 18], [3, 6, 12, 18], channel, 256)
    17. ​
    18.         self.path4 = FeatureFusionBlock(channel)
    19.         self.path3 = FeatureFusionBlock(channel)
    20.         self.path2 = FeatureFusionBlock(channel)
    21.         self.path1 = FeatureFusionBlock(channel)
    22. ​
    23.         for i in self.tpavi_stages:
    24.             setattr(self, f"tpavi_b{i + 1}", TPAVIModule(in_channels=channel, mode='dot'))
    25.             print("==> Build TPAVI block...")
    26. ​
    27.         self.output_conv = nn.Sequential(
    28.             nn.Conv2d(channel, 128, kernel_size=3, stride=1, padding=1),
    29.             Interpolate(scale_factor=2, mode="bilinear"),
    30.             nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1),
    31.             nn.ReLU(True),
    32.             nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
    33.         )
    34. ​
    35.         if self.training:
    36.             self.initialize_weights()
    37. ​
    38.     def pre_reshape_for_tpavi(self, x):
    39.         # x: [B*5, C, H, W]
    40.         _, C, H, W = x.shape
    41.         x = x.reshape(-1, 5, C, H, W)
    42.         x = x.permute(0, 2, 1, 3, 4).contiguous()  # [B, C, T, H, W]
    43.         return x
    44. ​
    45.     def post_reshape_for_tpavi(self, x):
    46.         # x: [B, C, T, H, W]
    47.         # return: [B*T, C, H, W]
    48.         _, C, _, H, W = x.shape
    49.         x = x.permute(0, 2, 1, 3, 4)  # [B, T, C, H, W]
    50.         x = x.view(-1, C, H, W)
    51.         return x
    52. ​
    53.     def tpavi_vv(self, x, stage):
    54.         # x: visual, [B*T, C=256, H, W]
    55.         tpavi_b = getattr(self, f'tpavi_b{stage + 1}')
    56.         x = self.pre_reshape_for_tpavi(x)  # [B, C, T, H, W]
    57.         x, _ = tpavi_b(x)  # [B, C, T, H, W]
    58.         x = self.post_reshape_for_tpavi(x)  # [B*T, C, H, W]
    59.         return x
    60. ​
    61.     def tpavi_va(self, x, audio, stage):
    62.         # x: visual, [B*T, C=256, H, W]
    63.         # audio: [B*T, 128]
    64.         # ra_flag: return audio feature list or not
    65.         tpavi_b = getattr(self, f'tpavi_b{stage + 1}')
    66.         audio = audio.view(-1, 5, audio.shape[-1])  # [B, T, 128]
    67.         x = self.pre_reshape_for_tpavi(x)  # [B, C, T, H, W]
    68.         x, a = tpavi_b(x, audio)  # [B, C, T, H, W], [B, T, C]
    69.         x = self.post_reshape_for_tpavi(x)  # [B*T, C, H, W]
    70.         return x, a
    71. ​
    72.     def _make_pred_layer(self, block, dilation_series, padding_series, NoLabels, input_channel):
    73.         return block(dilation_series, padding_series, NoLabels, input_channel)
    74. ​
    75.     def forward(self, x, audio_feature=None):
    76.         x = self.resnet.conv1(x)
    77.         x = self.resnet.bn1(x)
    78.         x = self.resnet.relu(x)
    79.         x = self.resnet.maxpool(x)
    80.         x1 = self.resnet.layer1(x)  # BF x 256  x 56 x 56
    81.         x2 = self.resnet.layer2(x1)  # BF x 512  x 28 x 28
    82.         x3 = self.resnet.layer3_1(x2)  # BF x 1024 x 14 x 14
    83.         x4 = self.resnet.layer4_1(x3)  # BF x 2048 x  7 x  7
    84.         # print(x1.shape, x2.shape, x3.shape, x4.shape)
    85. ​
    86.         conv1_feat = self.conv1(x1)  # BF x 256 x 56 x 56
    87.         conv2_feat = self.conv2(x2)  # BF x 256 x 28 x 28
    88.         conv3_feat = self.conv3(x3)  # BF x 256 x 14 x 14
    89.         conv4_feat = self.conv4(x4)  # BF x 256 x  7 x  7
    90.         # print(conv1_feat.shape, conv2_feat.shape, conv3_feat.shape, conv4_feat.shape)
    91. ​
    92.         feature_map_list = [conv1_feat, conv2_feat, conv3_feat, conv4_feat]
    93.         a_fea_list = [None] * 4
    94. ​
    95.         if len(self.tpavi_stages) > 0:
    96.             if (not self.tpavi_vv_flag) and (not self.tpavi_va_flag):
    97.                 raise Exception('tpavi_vv_flag and tpavi_va_flag cannot be False at the same time if len(tpavi_stages)>0, \
    98.                     tpavi_vv_flag is for video self-attention while tpavi_va_flag indicates the standard version (audio-visual attention)')
    99.             for i in self.tpavi_stages:
    100.                 tpavi_count = 0
    101.                 conv_feat = torch.zeros_like(feature_map_list[i]).cuda()
    102.                 if self.tpavi_vv_flag:
    103.                     conv_feat_vv = self.tpavi_vv(feature_map_list[i], stage=i)
    104.                     conv_feat += conv_feat_vv
    105.                     tpavi_count += 1
    106.                 if self.tpavi_va_flag:
    107.                     conv_feat_va, a_fea = self.tpavi_va(feature_map_list[i], audio_feature, stage=i)
    108.                     conv_feat += conv_feat_va
    109.                     tpavi_count += 1
    110.                     a_fea_list[i] = a_fea
    111.                 conv_feat /= tpavi_count
    112.                 feature_map_list[i] = conv_feat  # update features of stage-i which conduct TPAVI
    113. ​
    114.         conv4_feat = self.path4(feature_map_list[3])  # BF x 256 x 14 x 14
    115.         conv43 = self.path3(conv4_feat, feature_map_list[2])  # BF x 256 x 28 x 28
    116.         conv432 = self.path2(conv43, feature_map_list[1])  # BF x 256 x 56 x 56
    117.         conv4321 = self.path1(conv432, feature_map_list[0])  # BF x 256 x 112 x 112
    118.         # print(conv4_feat.shape, conv43.shape, conv432.shape, conv4321.shape)
    119. ​
    120.         pred = self.output_conv(conv4321)  # BF x 1 x 224 x 224
    121.         # print(pred.shape)
    122. ​
    123.         return pred, feature_map_list, a_fea_list
    124. ​
    125.     def initialize_weights(self):
    126.         res50 = models.resnet50(pretrained=False)
    127.         resnet50_dict = torch.load(self.cfg.TRAIN.PRETRAINED_RESNET50_PATH)
    128.         res50.load_state_dict(resnet50_dict)
    129.         pretrained_dict = res50.state_dict()
    130.         # print(pretrained_dict.keys())
    131.         all_params = {}
    132.         for k, v in self.resnet.state_dict().items():
    133.             if k in pretrained_dict.keys():
    134.                 v = pretrained_dict[k]
    135.                 all_params[k] = v
    136.             elif '_1' in k:
    137.                 name = k.split('_1')[0] + k.split('_1')[1]
    138.                 v = pretrained_dict[name]
    139.                 all_params[k] = v
    140.             elif '_2' in k:
    141.                 name = k.split('_2')[0] + k.split('_2')[1]
    142.                 v = pretrained_dict[name]
    143.                 all_params[k] = v
    144.         assert len(all_params.keys()) == len(self.resnet.state_dict().keys())
    145.         self.resnet.load_state_dict(all_params)
    146.         print(f'==> Load pretrained ResNet50 parameters from {self.cfg.TRAIN.PRETRAINED_RESNET50_PATH}')
    ​

    网络部分很简单，模型的定义没什么亮点，我们看forward里面的代码：

    1. def forward(self, x, audio_feature=None):  #  输入图像帧和音频梅尔图经过vggish 的特征图。
    2.     x = self.resnet.conv1(x)
    3.     x = self.resnet.bn1(x)
    4.     x = self.resnet.relu(x)
    5.     x = self.resnet.maxpool(x)
    6.     x1 = self.resnet.layer1(x)  # BF x 256  x 56 x 56
    7.     x2 = self.resnet.layer2(x1)  # BF x 512  x 28 x 28
    8.     x3 = self.resnet.layer3_1(x2)  # BF x 1024 x 14 x 14
    9.     x4 = self.resnet.layer4_1(x3)  # BF x 2048 x  7 x  7  先进行resnet特征提取
    10.     # print(x1.shape, x2.shape, x3.shape, x4.shape)
    11. ​
    12.     conv1_feat = self.conv1(x1)  # BF x 256 x 56 x 56   维度转换一下
    13.     conv2_feat = self.conv2(x2)  # BF x 256 x 28 x 28
    14.     conv3_feat = self.conv3(x3)  # BF x 256 x 14 x 14
    15.     conv4_feat = self.conv4(x4)  # BF x 256 x  7 x  7
    16.     # print(conv1_feat.shape, conv2_feat.shape, conv3_feat.shape, conv4_feat.shape)
    17. ​
    18.     feature_map_list = [conv1_feat, conv2_feat, conv3_feat, conv4_feat]
    19.     a_fea_list = [None] * 4
    20. ​
    21.     if len(self.tpavi_stages) > 0:   # 做几次tpavi模块，论文中是4次
    22.         if (not self.tpavi_vv_flag) and (not self.tpavi_va_flag):
    23.             raise Exception('tpavi_vv_flag and tpavi_va_flag cannot be False at the same time if len(tpavi_stages)>0, \
    24.                 tpavi_vv_flag is for video self-attention while tpavi_va_flag indicates the standard version (audio-visual attention)')
    25.         for i in self.tpavi_stages:
    26.             tpavi_count = 0
    27.             conv_feat = torch.zeros_like(feature_map_list[i]).cuda()
    28.             if self.tpavi_vv_flag:
    29.                 conv_feat_vv = self.tpavi_vv(feature_map_list[i], stage=i)
    30.                 conv_feat += conv_feat_vv
    31.                 tpavi_count += 1
    32.             if self.tpavi_va_flag:
    33.                 # tpavi模块
    34.                 conv_feat_va, a_fea = self.tpavi_va(feature_map_list[i], audio_feature, stage=i)  
    35.                 conv_feat += conv_feat_va
    36.                 tpavi_count += 1
    37.                 a_fea_list[i] = a_fea
    38.             conv_feat /= tpavi_count
    39.             feature_map_list[i] = conv_feat  # update features of stage-i which conduct TPAVI
    40. ​
    41.     conv4_feat = self.path4(feature_map_list[3])  # BF x 256 x 14 x 14  # 解码
    42.     conv43 = self.path3(conv4_feat, feature_map_list[2])  # BF x 256 x 28 x 28
    43.     conv432 = self.path2(conv43, feature_map_list[1])  # BF x 256 x 56 x 56
    44.     conv4321 = self.path1(conv432, feature_map_list[0])  # BF x 256 x 112 x 112
    45.     # print(conv4_feat.shape, conv43.shape, conv432.shape, conv4321.shape)
    46. ​
    47.     pred = self.output_conv(conv4321)  # BF x 1 x 224 x 224
    48.     # print(pred.shape)
    49. ​
    50.     return pred, feature_map_list, a_fea_list

    可以看到要经过一个TPAVI模块，是蛮复杂的模块：

    1. class TPAVIModule(nn.Module):
    2.     def __init__(self, in_channels, inter_channels=None, mode='dot', 
    3.                  dimension=3, bn_layer=True):
    4.         """
    5.         args:
    6.             in_channels: original channel size (1024 in the paper)
    7.             inter_channels: channel size inside the block if not specifed reduced to half (512 in the paper)
    8.             mode: supports Gaussian, Embedded Gaussian, Dot Product, and Concatenation 
    9.             dimension: can be 1 (temporal), 2 (spatial), 3 (spatiotemporal)
    10.             bn_layer: whether to add batch norm
    11.         """
    12.         super(TPAVIModule, self).__init__()
    13. ​
    14.         assert dimension in [1, 2, 3]
    15.         
    16.         if mode not in ['gaussian', 'embedded', 'dot', 'concatenate']:
    17.             raise ValueError('`mode` must be one of `gaussian`, `embedded`, `dot` or `concatenate`')
    18.             
    19.         self.mode = mode
    20.         self.dimension = dimension
    21. ​
    22.         self.in_channels = in_channels
    23.         self.inter_channels = inter_channels
    24. ​
    25.         # the channel size is reduced to half inside the block
    26.         if self.inter_channels is None:
    27.             self.inter_channels = in_channels // 2
    28.             if self.inter_channels == 0:
    29.                 self.inter_channels = 1
    30.         
    31.         ## add align channel
    32.         self.align_channel = nn.Linear(128, in_channels)
    33.         self.norm_layer=nn.LayerNorm(in_channels)
    34. ​
    35.         # assign appropriate convolutional, max pool, and batch norm layers for different dimensions
    36.         if dimension == 3:
    37.             conv_nd = nn.Conv3d
    38.             max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2))
    39.             bn = nn.BatchNorm3d
    40.         elif dimension == 2:
    41.             conv_nd = nn.Conv2d
    42.             max_pool_layer = nn.MaxPool2d(kernel_size=(2, 2))
    43.             bn = nn.BatchNorm2d
    44.         else:
    45.             conv_nd = nn.Conv1d
    46.             max_pool_layer = nn.MaxPool1d(kernel_size=(2))
    47.             bn = nn.BatchNorm1d
    48. ​
    49.         # function g in the paper which goes through conv. with kernel size 1
    50.         self.g = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1)
    51. ​
    52.         if bn_layer:
    53.             self.W_z = nn.Sequential(
    54.                     conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels, kernel_size=1),
    55.                     bn(self.in_channels)
    56.                 )
    57.             nn.init.constant_(self.W_z[1].weight, 0)
    58.             nn.init.constant_(self.W_z[1].bias, 0)
    59.         else:
    60.             self.W_z = conv_nd(in_channels=self.inter_channels, out_channels=self.in_channels, kernel_size=1)
    61. ​
    62.             nn.init.constant_(self.W_z.weight, 0)
    63.             nn.init.constant_(self.W_z.bias, 0)
    64. ​
    65.         # define theta and phi for all operations except gaussian
    66.         if self.mode == "embedded" or self.mode == "dot" or self.mode == "concatenate":
    67.             self.theta = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1)
    68.             self.phi = conv_nd(in_channels=self.in_channels, out_channels=self.inter_channels, kernel_size=1)
    69.         
    70.         if self.mode == "concatenate":
    71.             self.W_f = nn.Sequential(
    72.                     nn.Conv2d(in_channels=self.inter_channels * 2, out_channels=1, kernel_size=1),
    73.                     nn.ReLU()
    74.                 )
    75. ​
    76.             
    77.     def forward(self, x, audio=None):
    78.         """
    79.         args:
    80.             x: (N, C, T, H, W) for dimension=3; (N, C, H, W) for dimension 2; (N, C, T) for dimension 1
    81.             audio: (N, T, C)
    82.         """
    83. ​
    84.         audio_temp = 0
    85.         batch_size, C = x.size(0), x.size(1)
    86.         if audio is not None:
    87.             # print('==> audio.shape', audio.shape)
    88.             H, W = x.shape[-2], x.shape[-1]
    89.             audio_temp = self.align_channel(audio) # [bs, T, C]
    90.             audio = audio_temp.permute(0, 2, 1) # [bs, C, T]
    91.             audio = audio.unsqueeze(-1).unsqueeze(-1) # [bs, C, T, 1, 1]
    92.             audio = audio.repeat(1, 1, 1, H, W) # [bs, C, T, H, W]
    93.         else:
    94.             audio = x
    95. ​
    96.         # (N, C, THW)
    97.         g_x = self.g(x).view(batch_size, self.inter_channels, -1) # [bs, C, THW]
    98.         # print('g_x.shape', g_x.shape)
    99.         # g_x = x.view(batch_size, C, -1)  # [bs, C, THW]
    100.         g_x = g_x.permute(0, 2, 1) # [bs, THW, C]
    101. ​
    102.         if self.mode == "gaussian":
    103.             theta_x = x.view(batch_size, self.in_channels, -1)
    104.             phi_x = audio.view(batch_size, self.in_channels, -1)
    105.             theta_x = theta_x.permute(0, 2, 1)
    106.             f = torch.matmul(theta_x, phi_x)
    107. ​
    108.         elif self.mode == "embedded" or self.mode == "dot":
    109.             theta_x = self.theta(x).view(batch_size, self.inter_channels, -1) # [bs, C', THW]
    110.             phi_x = self.phi(audio).view(batch_size, self.inter_channels, -1) # [bs, C', THW]
    111.             theta_x = theta_x.permute(0, 2, 1) # [bs, THW, C']
    112.             f = torch.matmul(theta_x, phi_x) # [bs, THW, THW]
    113. ​
    114.         elif self.mode == "concatenate":
    115.             theta_x = self.theta(x).view(batch_size, self.inter_channels, -1, 1)
    116.             phi_x = self.phi(audio).view(batch_size, self.inter_channels, 1, -1)
    117.             
    118.             h = theta_x.size(2)
    119.             w = phi_x.size(3)
    120.             theta_x = theta_x.repeat(1, 1, 1, w)
    121.             phi_x = phi_x.repeat(1, 1, h, 1)
    122.             
    123.             concat = torch.cat([theta_x, phi_x], dim=1)
    124.             f = self.W_f(concat)
    125.             f = f.view(f.size(0), f.size(2), f.size(3))
    126.         
    127.         if self.mode == "gaussian" or self.mode == "embedded":
    128.             f_div_C = F.softmax(f, dim=-1)
    129.         elif self.mode == "dot" or self.mode == "concatenate":
    130.             N = f.size(-1) # number of position in x
    131.             f_div_C = f / N  # [bs, THW, THW]
    132.         
    133.         y = torch.matmul(f_div_C, g_x) # [bs, THW, C]
    134.         
    135.         # contiguous here just allocates contiguous chunk of memory
    136.         y = y.permute(0, 2, 1).contiguous() # [bs, C, THW]
    137.         y = y.view(batch_size, self.inter_channels, *x.size()[2:]) # [bs, C', T, H, W]
    138.         
    139.         W_y = self.W_z(y)  # [bs, C, T, H, W]
    140.         # residual connection
    141.         z = W_y + x #  # [bs, C, T, H, W]
    142. ​
    143.         # add LayerNorm
    144.         z =  z.permute(0, 2, 3, 4, 1) # [bs, T, H, W, C]
    145.         z = self.norm_layer(z)
    146.         z = z.permute(0, 4, 1, 2, 3) # [bs, C, T, H, W]
    147.         
    148.         return z, audio_temp

    代码看着复杂，其实是作者做了很多的模块选择以及代码的通道转换，实际最后的操作就是几个1* 1 *1 3D卷积，咱不用想也知道，3d卷积来做时序的特征提取。然后做一些累乘累加操作。

    1. if dimension == 3:
    2.     conv_nd = nn.Conv3d
    3.     max_pool_layer = nn.MaxPool3d(kernel_size=(1, 2, 2))
    4.     bn = nn.BatchNorm3d
    5. elif dimension == 2:
    6.     conv_nd = nn.Conv2d
    7.     max_pool_layer = nn.MaxPool2d(kernel_size=(2, 2))
    8.     bn = nn.BatchNorm2d
    9. else:
    10.     conv_nd = nn.Conv1d
    11.     max_pool_layer = nn.MaxPool1d(kernel_size=(2))
    12.     bn = nn.BatchNorm1d

    最后经过几个解码器，将特征图转为一维度：

    1. conv4_feat = self.path4(feature_map_list[3])  # BF x 256 x 14 x 14
    2. conv43 = self.path3(conv4_feat, feature_map_list[2])  # BF x 256 x 28 x 28
    3. conv432 = self.path2(conv43, feature_map_list[1])  # BF x 256 x 56 x 56
    4. conv4321 = self.path1(conv432, feature_map_list[0])  # BF x 256 x 112 x 112
    5. # print(conv4_feat.shape, conv43.shape, conv432.shape, conv4321.shape)
    6. ​
    7. pred = self.output_conv(conv4321)  # BF x 1 x 224 x 224

    可以看到【BF x 1 x 224 x 224】这个1维度的变化，就是网络的一个回归预测部分。最后输出的bs *frame 张1 * 224 *224 的图，就是我们最后输出的图（经过argmax等操作后显示成0，1分类），就变成了预测的mask图，

    大家可以看我的预测图：

     

    测试

    先看看ms3_meta_data.csv 的数据

     

    可以看到，一共有三份数据：训练、验证和测试集，我们训练好模型后，可以使用test.py 进行测试，测试效果会放在test_log文件夹。会去测试，test文件夹里的数据。运行测试代码，改一下训练好的模型路径就可以看到结果。

    测试某个视频

    点开avsbench_data/det/det的raw_videos/里面放你想测试的videos，建议5s，因为要切5帧，除非你改代码。

    然后运行preprocess_scripts/preprocess_ms3.py,这是为了生成语音的梅尔图，和切帧，会保存到raw_videos同级。

    接着运行detect.py(在train.py 同级)就可以针对你的视频，推理了。

    实时检测，这个代码我还在写，稍等。

    代码所有的链接（本地文件不能上传，只能提供原始github）：https://github.com/OpenNLPLab/AVSBench

    最后

    近期我会录制视频，过一遍原理和代码和训练推理，大家关注一下~

  • 相关阅读:
    Kubernetes中gRpc的服务发现
    Java版 招投标系统简介 招投标系统源码 java招投标系统 招投标系统功能设计
    人工智能领域的机器学习方法给我们的带来了哪些好处？
    为什么要学C语言及C语言存在的意义
    学习潘海东博士的《潮汐调和分析原理和应用》
    LibreOffice怎么打开导航页面
    精品基于Uniapp+SSM实现的公园植物介绍APP
    最简单的pdf加水印方法，五个步骤很详细
    elasticsearch 7.x的 常用命令
    【leetcode】开密码锁
  • 原文地址：https://blog.csdn.net/qq_46098574/article/details/126255334