def __init__(self):
super(Model, self).__init__()
# some configurations for the model
n_tc_timesteps = config.cfg.MODEL.N_TC_TIMESTEPS #32
backbone_name = config.cfg.MODEL.BACKBONE_CNN #'i3d_pytorch_charades_rgb'
feature_name = config.cfg.MODEL.BACKBONE_FEATURE #'mixed_5c'
n_tc_layers = config.cfg.MODEL.N_TC_LAYERS #2
n_classes = config.cfg.MODEL.N_CLASSES #157
is_dilated = config.cfg.MODEL.MULTISCALE_TYPE #采用的多核策略类型'ks'
OutputActivation = Sigmoid if config.cfg.MODEL.CLASSIFICATION_TYPE == 'ml' else LogSoftmax
n_channels_in, channel_h, channel_w = utils.get_model_feat_maps_info(
backbone_name, feature_name) #1024, 7, 7
n_groups = int(n_channels_in / 128.0) #8
input_shape = (None, n_channels_in, n_tc_timesteps, channel_h,
channel_w
) # (None, C, T, H, W),(None, 1024, 32, 7, 7)其中T是自己设定的
self._input_shape = input_shape #(None, 1024, 32, 7, 7)
# define 4 layers of timeception
self.timeception = timeception_pytorch.Timeception(
input_shape, n_tc_layers, n_groups, is_dilated)
# get number of output channels after timeception
n_channels_in = self.timeception.n_channels_out
# define layers for classifier
self.do1 = Dropout(0.5)
self.l1 = Linear(n_channels_in, 512)
self.bn1 = BatchNorm1d(512)
self.ac1 = LeakyReLU(0.2)
self.do2 = Dropout(0.25)
self.l2 = Linear(512, n_classes)
self.ac2 = OutputActivation()
def __init__(self):
super(Model, self).__init__()
# some configurations for the model
n_tc_timesteps = config.cfg.MODEL.N_TC_TIMESTEPS # 输入到timeception的timesteps
backbone_name = config.cfg.MODEL.BACKBONE_CNN
feature_name = config.cfg.MODEL.BACKBONE_FEATURE
n_tc_layers = config.cfg.MODEL.N_TC_LAYERS # timeception的层数
n_classes = config.cfg.MODEL.N_CLASSES # 157
is_dilated = config.cfg.MODEL.MULTISCALE_TYPE # 多尺度的类型（空洞卷积/多尺度卷积核）
OutputActivation = Sigmoid if config.cfg.MODEL.CLASSIFICATION_TYPE == 'ml' else LogSoftmax
n_channels_in, channel_h, channel_w = utils.get_model_feat_maps_info(
backbone_name, feature_name)
n_groups = int(n_channels_in / 128.0) # groups的数量
# timeception层的输入shape
input_shape = (None, n_channels_in, n_tc_timesteps, channel_h,
channel_w) # (C, T, H, W)
self._input_shape = input_shape
# define 4 layers of timeception
self.timeception = timeception_pytorch.Timeception(
input_shape, n_tc_layers, n_groups, is_dilated) # (C, T, H, W)
# get number of output channels after timeception
n_channels_in = self.timeception.n_channels_out
# define layers for classifier
self.do1 = Dropout(0.5)
self.l1 = Linear(n_channels_in, 512)
self.bn1 = BatchNorm1d(512)
self.ac1 = LeakyReLU(0.2) #negative_slope：控制负斜率的角度，默认等于0.01
self.do2 = Dropout(0.25)
self.l2 = Linear(512, n_classes)
self.ac2 = OutputActivation()
import numpy as np import torch as T from nets import timeception_pytorch # i3d first input layer has shape (10, 3, 36, 224, 224) # (batch_size, channel, time, height, width) # define input tensor (from i3d last layer before logic) input = T.tensor(np.zeros((10, 1024, 5, 7, 7)), dtype=T.float32) # define 4 layers of timeception module = timeception_pytorch.Timeception(input.size(), n_layers=2) # feedforward the input to the timeception layers tensor = module(input) # the output is (10, 1600, 1, 7, 7) print(tensor.size())