--- a
+++ b/mmaction/models/heads/trn_head.py
@@ -0,0 +1,211 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import itertools
+
+import numpy as np
+import torch
+import torch.nn as nn
+from mmcv.cnn import normal_init
+
+from ..builder import HEADS
+from .base import BaseHead
+
+
+class RelationModule(nn.Module):
+    """Relation Module of TRN.
+
+    Args:
+        hidden_dim (int): The dimension of hidden layer of MLP in relation
+            module.
+        num_segments (int): Number of frame segments.
+        num_classes (int): Number of classes to be classified.
+    """
+
+    def __init__(self, hidden_dim, num_segments, num_classes):
+        super().__init__()
+        self.hidden_dim = hidden_dim
+        self.num_segments = num_segments
+        self.num_classes = num_classes
+        bottleneck_dim = 512
+        self.classifier = nn.Sequential(
+            nn.ReLU(),
+            nn.Linear(self.num_segments * self.hidden_dim, bottleneck_dim),
+            nn.ReLU(), nn.Linear(bottleneck_dim, self.num_classes))
+
+    def init_weights(self):
+        # Use the default kaiming_uniform for all nn.linear layers.
+        pass
+
+    def forward(self, x):
+        # [N, num_segs * hidden_dim]
+        x = x.view(x.size(0), -1)
+        x = self.classifier(x)
+        return x
+
+
+class RelationModuleMultiScale(nn.Module):
+    """Relation Module with Multi Scale of TRN.
+
+    Args:
+        hidden_dim (int): The dimension of hidden layer of MLP in relation
+            module.
+        num_segments (int): Number of frame segments.
+        num_classes (int): Number of classes to be classified.
+    """
+
+    def __init__(self, hidden_dim, num_segments, num_classes):
+        super().__init__()
+        self.hidden_dim = hidden_dim
+        self.num_segments = num_segments
+        self.num_classes = num_classes
+
+        # generate the multiple frame relations
+        self.scales = range(num_segments, 1, -1)
+
+        self.relations_scales = []
+        self.subsample_scales = []
+        max_subsample = 3
+        for scale in self.scales:
+            # select the different frame features for different scales
+            relations_scale = list(
+                itertools.combinations(range(self.num_segments), scale))
+            self.relations_scales.append(relations_scale)
+            # sample `max_subsample` relation_scale at most
+            self.subsample_scales.append(
+                min(max_subsample, len(relations_scale)))
+        assert len(self.relations_scales[0]) == 1
+
+        bottleneck_dim = 256
+        self.fc_fusion_scales = nn.ModuleList()
+        for scale in self.scales:
+            fc_fusion = nn.Sequential(
+                nn.ReLU(), nn.Linear(scale * self.hidden_dim, bottleneck_dim),
+                nn.ReLU(), nn.Linear(bottleneck_dim, self.num_classes))
+            self.fc_fusion_scales.append(fc_fusion)
+
+    def init_weights(self):
+        # Use the default kaiming_uniform for all nn.linear layers.
+        pass
+
+    def forward(self, x):
+        # the first one is the largest scale
+        act_all = x[:, self.relations_scales[0][0], :]
+        act_all = act_all.view(
+            act_all.size(0), self.scales[0] * self.hidden_dim)
+        act_all = self.fc_fusion_scales[0](act_all)
+
+        for scaleID in range(1, len(self.scales)):
+            # iterate over the scales
+            idx_relations_randomsample = np.random.choice(
+                len(self.relations_scales[scaleID]),
+                self.subsample_scales[scaleID],
+                replace=False)
+            for idx in idx_relations_randomsample:
+                act_relation = x[:, self.relations_scales[scaleID][idx], :]
+                act_relation = act_relation.view(
+                    act_relation.size(0),
+                    self.scales[scaleID] * self.hidden_dim)
+                act_relation = self.fc_fusion_scales[scaleID](act_relation)
+                act_all += act_relation
+        return act_all
+
+
+@HEADS.register_module()
+class TRNHead(BaseHead):
+    """Class head for TRN.
+
+    Args:
+        num_classes (int): Number of classes to be classified.
+        in_channels (int): Number of channels in input feature.
+        num_segments (int): Number of frame segments. Default: 8.
+        loss_cls (dict): Config for building loss. Default:
+            dict(type='CrossEntropyLoss')
+        spatial_type (str): Pooling type in spatial dimension. Default: 'avg'.
+        relation_type (str): The relation module type. Choices are 'TRN' or
+            'TRNMultiScale'. Default: 'TRNMultiScale'.
+        hidden_dim (int): The dimension of hidden layer of MLP in relation
+            module. Default: 256.
+        dropout_ratio (float): Probability of dropout layer. Default: 0.8.
+        init_std (float): Std value for Initiation. Default: 0.001.
+        kwargs (dict, optional): Any keyword argument to be used to initialize
+            the head.
+    """
+
+    def __init__(self,
+                 num_classes,
+                 in_channels,
+                 num_segments=8,
+                 loss_cls=dict(type='CrossEntropyLoss'),
+                 spatial_type='avg',
+                 relation_type='TRNMultiScale',
+                 hidden_dim=256,
+                 dropout_ratio=0.8,
+                 init_std=0.001,
+                 **kwargs):
+        super().__init__(num_classes, in_channels, loss_cls, **kwargs)
+
+        self.num_classes = num_classes
+        self.in_channels = in_channels
+        self.num_segments = num_segments
+        self.spatial_type = spatial_type
+        self.relation_type = relation_type
+        self.hidden_dim = hidden_dim
+        self.dropout_ratio = dropout_ratio
+        self.init_std = init_std
+
+        if self.relation_type == 'TRN':
+            self.consensus = RelationModule(self.hidden_dim, self.num_segments,
+                                            self.num_classes)
+        elif self.relation_type == 'TRNMultiScale':
+            self.consensus = RelationModuleMultiScale(self.hidden_dim,
+                                                      self.num_segments,
+                                                      self.num_classes)
+        else:
+            raise ValueError(f'Unknown Relation Type {self.relation_type}!')
+
+        if self.dropout_ratio != 0:
+            self.dropout = nn.Dropout(p=self.dropout_ratio)
+        else:
+            self.dropout = None
+        self.fc_cls = nn.Linear(self.in_channels, self.hidden_dim)
+
+        if self.spatial_type == 'avg':
+            # use `nn.AdaptiveAvgPool2d` to adaptively match the in_channels.
+            self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        else:
+            self.avg_pool = None
+
+    def init_weights(self):
+        """Initiate the parameters from scratch."""
+        normal_init(self.fc_cls, std=self.init_std)
+        self.consensus.init_weights()
+
+    def forward(self, x, num_segs):
+        """Defines the computation performed at every call.
+
+        Args:
+            x (torch.Tensor): The input data.
+            num_segs (int): Useless in TRNHead. By default, `num_segs`
+                is equal to `clip_len * num_clips * num_crops`, which is
+                automatically generated in Recognizer forward phase and
+                useless in TRN models. The `self.num_segments` we need is a
+                hyper parameter to build TRN models.
+        Returns:
+            torch.Tensor: The classification scores for input samples.
+        """
+        # [N * num_segs, in_channels, 7, 7]
+        if self.avg_pool is not None:
+            x = self.avg_pool(x)
+        # [N * num_segs, in_channels, 1, 1]
+        x = torch.flatten(x, 1)
+        # [N * num_segs, in_channels]
+        if self.dropout is not None:
+            x = self.dropout(x)
+
+        # [N, num_segs, hidden_dim]
+        cls_score = self.fc_cls(x)
+        cls_score = cls_score.view((-1, self.num_segments) +
+                                   cls_score.size()[1:])
+
+        # [N, num_classes]
+        cls_score = self.consensus(cls_score)
+        return cls_score