【视频理解】TSN Temporal Segment Networks: Towards Good Practices for Deep Action Recognition 笔记

Temporal Segment Networks: Towards Good Practices for Deep Action Recognition

https://arxiv.org/abs/1608.00859
https://github.com/yjxiong/temporal-segment-networks
https://github.com/yjxiong/tsn-pytorch

时间分段网络（TSN，Two-Stream）
结合了稀疏的时间采样策略和视频级别的监督，可以使用整个动作视频而不只是一个视频片段的信息进行高效的学习。测试时采用主流方法进行。
ActivityNet 2016竞赛的冠军(93.2% mAP)、HMDB51 ( 69.4%)、UCF101 (94.2%)

Paper

Method

视频中的长距离时序依赖

视频运动(motion)信息的处理和设法融合表象和运动信息是解决视频理解任务的关键

目前做动作识别的两大主流方法是3D卷积和two-stream，但这里两种方案能捕获的仅是视频中的短距离时序依赖。为了捕获长距离时序依赖，这些方法通常需要密采样视频片段clip(时序动作定位里，将视频分帧后，采用多尺度滑动窗口，比如滑动窗口为64，也就是每64帧图片为一个视频clip，视频分为若干个clip )

采用稀疏采样，利用整个视频的信息（相邻的帧有信息冗余）

TSN把视频分成3段，每个片段均匀地随机采样一个视频片段，并使用双流网络得到视频片段属于各类得分(softmax之前的值)，之后把不同片段得分取平均，最后通过softmax输出。下图K个spatial convnet的参数是共享的，K个temporal convnet的参数也是共享的

Details

Input

• RGB
  video中的某一帧
• RGB difference
  相邻两帧的差，可以用来表达动作信息
• optical flow
• warped optical flow

Modality

• RGB/optical flow
  1:1.5
• RGB/optical flow/warped optical flow
  1:1:0.5

Data argument

• Random cropping
• Horizontal flipping
• Corner cropping
  四角+中心，防止网络只关心中心位置
• Scale and ratio jittering

other skills

• Cross-modality pre-training
  • spatial ConvNets
    用在ImageNet预训练模型对双流网络进行初始化
  • temporal ConvNets
    交叉预训练，将图像领域的预训练模型迁移到光流领域
• Partial BN with dropout
  除了第一个之外的所有BN层的均值和标准差参数固定

Results

Pytorch

from torch import nn

from ops.basic_ops import ConsensusModule, Identity
from transforms import *
from torch.nn.init import normal, constant

class TSN(nn.Module):
    def __init__(self, num_class, num_segments, modality,
                 base_model='resnet101', new_length=None,
                 consensus_type='avg', before_softmax=True,
                 dropout=0.8,
                 crop_num=1, partial_bn=True):
        super(TSN, self).__init__()
        self.modality = modality
        self.num_segments = num_segments
        self.reshape = True
        self.before_softmax = before_softmax
        self.dropout = dropout
        self.crop_num = crop_num
        self.consensus_type = consensus_type
        if not before_softmax and consensus_type != 'avg':
            raise ValueError("Only avg consensus can be used after Softmax")

        if new_length is None:
            self.new_length = 1 if modality == "RGB" else 5
        else:
            self.new_length = new_length

        print(("""
Initializing TSN with base model: {}.
TSN Configurations:
    input_modality:     {}
    num_segments:       {}
    new_length:         {}
    consensus_module:   {}
    dropout_ratio:      {}
        """.format(base_model, self.modality, self.num_segments, self.new_length, consensus_type, self.dropout)))

        self._prepare_base_model(base_model)

        feature_dim = self._prepare_tsn(num_class)

        if self.modality == 'Flow':
            print("Converting the ImageNet model to a flow init model")
            self.base_model = self._construct_flow_model(self.base_model)
            print("Done. Flow model ready...")
        elif self.modality == 'RGBDiff':
            print("Converting the ImageNet model to RGB+Diff init model")
            self.base_model = self._construct_diff_model(self.base_model)
            print("Done. RGBDiff model ready.")

        self.consensus = ConsensusModule(consensus_type)

        if not self.before_softmax:
            self.softmax = nn.Softmax()

        self._enable_pbn = partial_bn
        if partial_bn:
            self.partialBN(True)

    def _prepare_tsn(self, num_class):
        feature_dim = getattr(self.base_model, self.base_model.last_layer_name).in_features
        if self.dropout == 0:
            setattr(self.base_model, self.base_model.last_layer_name, nn.Linear(feature_dim, num_class))
            self.new_fc = None
        else:
            setattr(self.base_model, self.base_model.last_layer_name, nn.Dropout(p=self.dropout))
            self.new_fc = nn.Linear(feature_dim, num_class)

        std = 0.001
        if self.new_fc is None:
            normal(getattr(self.base_model, self.base_model.last_layer_name).weight, 0, std)
            constant(getattr(self.base_model, self.base_model.last_layer_name).bias, 0)
        else:
            normal(self.new_fc.weight, 0, std)
            constant(self.new_fc.bias, 0)
        return feature_dim

    def _prepare_base_model(self, base_model):

        if 'resnet' in base_model or 'vgg' in base_model:
            self.base_model = getattr(torchvision.models, base_model)(True)
            self.base_model.last_layer_name = 'fc'
            self.input_size = 224
            self.input_mean = [0.485, 0.456, 0.406]
            self.input_std = [0.229, 0.224, 0.225]

            if self.modality == 'Flow':
                self.input_mean = [0.5]
                self.input_std = [np.mean(self.input_std)]
            elif self.modality == 'RGBDiff':
                self.input_mean = [0.485, 0.456, 0.406] + [0] * 3 * self.new_length
                self.input_std = self.input_std + [np.mean(self.input_std) * 2] * 3 * self.new_length
        elif base_model == 'BNInception':
            import tf_model_zoo
            self.base_model = getattr(tf_model_zoo, base_model)()
            self.base_model.last_layer_name = 'fc'
            self.input_size = 224
            self.input_mean = [104, 117, 128]
            self.input_std = [1]

            if self.modality == 'Flow':
                self.input_mean = [128]
            elif self.modality == 'RGBDiff':
                self.input_mean = self.input_mean * (1 + self.new_length)

        elif 'inception' in base_model:
            import tf_model_zoo
            self.base_model = getattr(tf_model_zoo, base_model)()
            self.base_model.last_layer_name = 'classif'
            self.input_size = 299
            self.input_mean = [0.5]
            self.input_std = [0.5]
        else:
            raise ValueError('Unknown base model: {}'.format(base_model))

    def train(self, mode=True):
        """
        Override the default train() to freeze the BN parameters
        :return:
        """
        super(TSN, self).train(mode)
        count = 0
        if self._enable_pbn:
            print("Freezing BatchNorm2D except the first one.")
            for m in self.base_model.modules():
                if isinstance(m, nn.BatchNorm2d):
                    count += 1
                    if count >= (2 if self._enable_pbn else 1):
                        m.eval()

                        # shutdown update in frozen mode
                        m.weight.requires_grad = False
                        m.bias.requires_grad = False

    def partialBN(self, enable):
        self._enable_pbn = enable

    def get_optim_policies(self):
        first_conv_weight = []
        first_conv_bias = []
        normal_weight = []
        normal_bias = []
        bn = []

        conv_cnt = 0
        bn_cnt = 0
        for m in self.modules():
            if isinstance(m, torch.nn.Conv2d) or isinstance(m, torch.nn.Conv1d):
                ps = list(m.parameters())
                conv_cnt += 1
                if conv_cnt == 1:
                    first_conv_weight.append(ps[0])
                    if len(ps) == 2:
                        first_conv_bias.append(ps[1])
                else:
                    normal_weight.append(ps[0])
                    if len(ps) == 2:
                        normal_bias.append(ps[1])
            elif isinstance(m, torch.nn.Linear):
                ps = list(m.parameters())
                normal_weight.append(ps[0])
                if len(ps) == 2:
                    normal_bias.append(ps[1])
                  
            elif isinstance(m, torch.nn.BatchNorm1d):
                bn.extend(list(m.parameters()))
            elif isinstance(m, torch.nn.BatchNorm2d):
                bn_cnt += 1
                # later BN's are frozen
                if not self._enable_pbn or bn_cnt == 1:
                    bn.extend(list(m.parameters()))
            elif len(m._modules) == 0:
                if len(list(m.parameters())) > 0:
                    raise ValueError("New atomic module type: {}. Need to give it a learning policy".format(type(m)))

        return [
            {'params': first_conv_weight, 'lr_mult': 5 if self.modality == 'Flow' else 1, 'decay_mult': 1,
             'name': "first_conv_weight"},
            {'params': first_conv_bias, 'lr_mult': 10 if self.modality == 'Flow' else 2, 'decay_mult': 0,
             'name': "first_conv_bias"},
            {'params': normal_weight, 'lr_mult': 1, 'decay_mult': 1,
             'name': "normal_weight"},
            {'params': normal_bias, 'lr_mult': 2, 'decay_mult': 0,
             'name': "normal_bias"},
            {'params': bn, 'lr_mult': 1, 'decay_mult': 0,
             'name': "BN scale/shift"},
        ]

    def forward(self, input):
        sample_len = (3 if self.modality == "RGB" else 2) * self.new_length

        if self.modality == 'RGBDiff':
            sample_len = 3 * self.new_length
            input = self._get_diff(input)

        base_out = self.base_model(input.view((-1, sample_len) + input.size()[-2:]))

        if self.dropout > 0:
            base_out = self.new_fc(base_out)

        if not self.before_softmax:
            base_out = self.softmax(base_out)
        if self.reshape:
            base_out = base_out.view((-1, self.num_segments) + base_out.size()[1:])

        output = self.consensus(base_out)
        return output.squeeze(1)

    def _get_diff(self, input, keep_rgb=False):
        input_c = 3 if self.modality in ["RGB", "RGBDiff"] else 2
        input_view = input.view((-1, self.num_segments, self.new_length + 1, input_c,) + input.size()[2:])
        if keep_rgb:
            new_data = input_view.clone()
        else:
            new_data = input_view[:, :, 1:, :, :, :].clone()

        for x in reversed(list(range(1, self.new_length + 1))):
            if keep_rgb:
                new_data[:, :, x, :, :, :] = input_view[:, :, x, :, :, :] - input_view[:, :, x - 1, :, :, :]
            else:
                new_data[:, :, x - 1, :, :, :] = input_view[:, :, x, :, :, :] - input_view[:, :, x - 1, :, :, :]

        return new_data


    def _construct_flow_model(self, base_model):
        # modify the convolution layers
        # Torch models are usually defined in a hierarchical way.
        # nn.modules.children() return all sub modules in a DFS manner
        modules = list(self.base_model.modules())
        first_conv_idx = list(filter(lambda x: isinstance(modules[x], nn.Conv2d), list(range(len(modules)))))[0]
        conv_layer = modules[first_conv_idx]
        container = modules[first_conv_idx - 1]

        # modify parameters, assume the first blob contains the convolution kernels
        params = [x.clone() for x in conv_layer.parameters()]
        kernel_size = params[0].size()
        new_kernel_size = kernel_size[:1] + (2 * self.new_length, ) + kernel_size[2:]
        new_kernels = params[0].data.mean(dim=1, keepdim=True).expand(new_kernel_size).contiguous()

        new_conv = nn.Conv2d(2 * self.new_length, conv_layer.out_channels,
                             conv_layer.kernel_size, conv_layer.stride, conv_layer.padding,
                             bias=True if len(params) == 2 else False)
        new_conv.weight.data = new_kernels
        if len(params) == 2:
            new_conv.bias.data = params[1].data # add bias if neccessary
        layer_name = list(container.state_dict().keys())[0][:-7] # remove .weight suffix to get the layer name

        # replace the first convlution layer
        setattr(container, layer_name, new_conv)
        return base_model

    def _construct_diff_model(self, base_model, keep_rgb=False):
        # modify the convolution layers
        # Torch models are usually defined in a hierarchical way.
        # nn.modules.children() return all sub modules in a DFS manner
        modules = list(self.base_model.modules())
        first_conv_idx = list(filter(lambda x: isinstance(modules[x], nn.Conv2d), list(range(len(modules)))))[0]
        conv_layer = modules[first_conv_idx]
        container = modules[first_conv_idx - 1]

        # modify parameters, assume the first blob contains the convolution kernels
        params = [x.clone() for x in conv_layer.parameters()]
        kernel_size = params[0].size()
        if not keep_rgb:
            new_kernel_size = kernel_size[:1] + (3 * self.new_length,) + kernel_size[2:]
            new_kernels = params[0].data.mean(dim=1, keepdim=True).expand(new_kernel_size).contiguous()
        else:
            new_kernel_size = kernel_size[:1] + (3 * self.new_length,) + kernel_size[2:]
            new_kernels = torch.cat((params[0].data, params[0].data.mean(dim=1, keepdim=True).expand(new_kernel_size).contiguous()),
                                    1)
            new_kernel_size = kernel_size[:1] + (3 + 3 * self.new_length,) + kernel_size[2:]

        new_conv = nn.Conv2d(new_kernel_size[1], conv_layer.out_channels,
                             conv_layer.kernel_size, conv_layer.stride, conv_layer.padding,
                             bias=True if len(params) == 2 else False)
        new_conv.weight.data = new_kernels
        if len(params) == 2:
            new_conv.bias.data = params[1].data  # add bias if neccessary
        layer_name = list(container.state_dict().keys())[0][:-7]  # remove .weight suffix to get the layer name

        # replace the first convolution layer
        setattr(container, layer_name, new_conv)
        return base_model

    @property
    def crop_size(self):
        return self.input_size

    @property
    def scale_size(self):
        return self.input_size * 256 // 224

    def get_augmentation(self):
        if self.modality == 'RGB':
            return torchvision.transforms.Compose([GroupMultiScaleCrop(self.input_size, [1, .875, .75, .66]),
                                                   GroupRandomHorizontalFlip(is_flow=False)])
        elif self.modality == 'Flow':
            return torchvision.transforms.Compose([GroupMultiScaleCrop(self.input_size, [1, .875, .75]),
                                                   GroupRandomHorizontalFlip(is_flow=True)])
        elif self.modality == 'RGBDiff':
            return torchvision.transforms.Compose([GroupMultiScaleCrop(self.input_size, [1, .875, .75]),
                                                   GroupRandomHorizontalFlip(is_flow=False)])