4. 深入探究在 Kinetics400 数据集上训练 I3D 模型

这是一个使用 Gluon CV 工具包的视频动作识别教程，它是一个循序渐进的示例。读者应该具备深度学习的基础知识，并熟悉 Gluon API。新用户可以先学习 A 60-minute Gluon Crash Course。您可以 立即开始训练 或 `深入探究`_。

立即开始训练

注意

您可以跳过本教程，因为训练脚本是完整的，可以直接启动。

下载 完整 Python 脚本: train_recognizer.py

更多训练命令行选项，请运行 python train_recognizer.py -h 请查看 模型库 以获取复现预训练模型的训练命令。

网络结构

首先，让我们导入所需的 Python 库。

from __future__ import division

import argparse, time, logging, os, sys, math

import numpy as np
import mxnet as mx
import gluoncv as gcv
from mxnet import gluon, nd, init, context
from mxnet import autograd as ag
from mxnet.gluon import nn
from mxnet.gluon.data.vision import transforms

from gluoncv.data.transforms import video
from gluoncv.data import Kinetics400
from gluoncv.model_zoo import get_model
from gluoncv.utils import makedirs, LRSequential, LRScheduler, split_and_load, TrainingHistory

本教程我们选用一个广泛采用的模型，I3D-InceptionV1。 I3D (Inflated 3D Networks) 是一个广泛采用的 3D 视频分类网络。它使用 3D 卷积直接从视频中学习时空信息。提出 I3D 是为了通过从 2D 模型“膨胀”而来改进 C3D 模型。我们不仅可以重用 2D 模型的架构（例如 ResNet、Inception），还可以从 2D 预训练模型引导模型权重。通过这种方式，训练 3D 网络进行视频分类变得可行，并且取得了更好的结果。

# number of GPUs to use
num_gpus = 1
ctx = [mx.gpu(i) for i in range(num_gpus)]

# Get the model i3d_inceptionv1_kinetics400 with 400 output classes, without pre-trained weights
net = get_model(name='i3d_inceptionv1_kinetics400', nclass=400)
net.collect_params().reset_ctx(ctx)
print(net)

输出

Downloading /root/.mxnet/models/googlenet-c7c89366.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/googlenet-c7c89366.zip...

  0%|          | 0/13680 [00:00<?, ?KB/s]
  3%|2         | 373/13680 [00:00<00:04, 2821.98KB/s]
 14%|#4        | 1928/13680 [00:00<00:01, 7991.17KB/s]
 52%|#####2    | 7162/13680 [00:00<00:00, 25107.90KB/s]
 96%|#########6| 13184/13680 [00:00<00:00, 37586.58KB/s]
100%|##########| 13680/13680 [00:00<00:00, 28878.21KB/s]
i3d_inceptionv11_conv0_weight is done with shape:  (64, 3, 7, 7, 7)
i3d_inceptionv11_batchnorm0_gamma is done with shape:  (64,)
i3d_inceptionv11_batchnorm0_beta is done with shape:  (64,)
i3d_inceptionv11_batchnorm0_running_mean is done with shape:  (64,)
i3d_inceptionv11_batchnorm0_running_var is done with shape:  (64,)
i3d_inceptionv11_conv1_weight is done with shape:  (64, 64, 1, 1, 1)
i3d_inceptionv11_batchnorm1_gamma is done with shape:  (64,)
i3d_inceptionv11_batchnorm1_beta is done with shape:  (64,)
i3d_inceptionv11_batchnorm1_running_mean is done with shape:  (64,)
i3d_inceptionv11_batchnorm1_running_var is done with shape:  (64,)
i3d_inceptionv11_conv2_weight is done with shape:  (192, 64, 3, 3, 3)
i3d_inceptionv11_batchnorm2_gamma is done with shape:  (192,)
i3d_inceptionv11_batchnorm2_beta is done with shape:  (192,)
i3d_inceptionv11_batchnorm2_running_mean is done with shape:  (192,)
i3d_inceptionv11_batchnorm2_running_var is done with shape:  (192,)
i3d_inceptionv11_Mixed_3a_conv0_weight is done with shape:  (64, 192, 1, 1, 1)
i3d_inceptionv11_Mixed_3a_batchnorm0_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_3a_batchnorm0_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_3a_batchnorm0_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_3a_batchnorm0_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_3a_conv1_weight is done with shape:  (96, 192, 1, 1, 1)
i3d_inceptionv11_Mixed_3a_batchnorm1_gamma is done with shape:  (96,)
i3d_inceptionv11_Mixed_3a_batchnorm1_beta is done with shape:  (96,)
i3d_inceptionv11_Mixed_3a_batchnorm1_running_mean is done with shape:  (96,)
i3d_inceptionv11_Mixed_3a_batchnorm1_running_var is done with shape:  (96,)
i3d_inceptionv11_Mixed_3a_conv2_weight is done with shape:  (128, 96, 3, 3, 3)
i3d_inceptionv11_Mixed_3a_batchnorm2_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_3a_batchnorm2_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_3a_batchnorm2_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_3a_batchnorm2_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_3a_conv3_weight is done with shape:  (16, 192, 1, 1, 1)
i3d_inceptionv11_Mixed_3a_batchnorm3_gamma is done with shape:  (16,)
i3d_inceptionv11_Mixed_3a_batchnorm3_beta is done with shape:  (16,)
i3d_inceptionv11_Mixed_3a_batchnorm3_running_mean is done with shape:  (16,)
i3d_inceptionv11_Mixed_3a_batchnorm3_running_var is done with shape:  (16,)
i3d_inceptionv11_Mixed_3a_conv4_weight is done with shape:  (32, 16, 3, 3, 3)
i3d_inceptionv11_Mixed_3a_batchnorm4_gamma is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_batchnorm4_beta is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_batchnorm4_running_mean is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_batchnorm4_running_var is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_conv5_weight is done with shape:  (32, 192, 1, 1, 1)
i3d_inceptionv11_Mixed_3a_batchnorm5_gamma is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_batchnorm5_beta is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_batchnorm5_running_mean is done with shape:  (32,)
i3d_inceptionv11_Mixed_3a_batchnorm5_running_var is done with shape:  (32,)
i3d_inceptionv11_Mixed_3b_conv0_weight is done with shape:  (128, 256, 1, 1, 1)
i3d_inceptionv11_Mixed_3b_batchnorm0_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_batchnorm0_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_batchnorm0_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_batchnorm0_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_conv1_weight is done with shape:  (128, 256, 1, 1, 1)
i3d_inceptionv11_Mixed_3b_batchnorm1_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_batchnorm1_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_batchnorm1_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_batchnorm1_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_3b_conv2_weight is done with shape:  (192, 128, 3, 3, 3)
i3d_inceptionv11_Mixed_3b_batchnorm2_gamma is done with shape:  (192,)
i3d_inceptionv11_Mixed_3b_batchnorm2_beta is done with shape:  (192,)
i3d_inceptionv11_Mixed_3b_batchnorm2_running_mean is done with shape:  (192,)
i3d_inceptionv11_Mixed_3b_batchnorm2_running_var is done with shape:  (192,)
i3d_inceptionv11_Mixed_3b_conv3_weight is done with shape:  (32, 256, 1, 1, 1)
i3d_inceptionv11_Mixed_3b_batchnorm3_gamma is done with shape:  (32,)
i3d_inceptionv11_Mixed_3b_batchnorm3_beta is done with shape:  (32,)
i3d_inceptionv11_Mixed_3b_batchnorm3_running_mean is done with shape:  (32,)
i3d_inceptionv11_Mixed_3b_batchnorm3_running_var is done with shape:  (32,)
i3d_inceptionv11_Mixed_3b_conv4_weight is done with shape:  (96, 32, 3, 3, 3)
i3d_inceptionv11_Mixed_3b_batchnorm4_gamma is done with shape:  (96,)
i3d_inceptionv11_Mixed_3b_batchnorm4_beta is done with shape:  (96,)
i3d_inceptionv11_Mixed_3b_batchnorm4_running_mean is done with shape:  (96,)
i3d_inceptionv11_Mixed_3b_batchnorm4_running_var is done with shape:  (96,)
i3d_inceptionv11_Mixed_3b_conv5_weight is done with shape:  (64, 256, 1, 1, 1)
i3d_inceptionv11_Mixed_3b_batchnorm5_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_3b_batchnorm5_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_3b_batchnorm5_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_3b_batchnorm5_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4a_conv0_weight is done with shape:  (192, 480, 1, 1, 1)
i3d_inceptionv11_Mixed_4a_batchnorm0_gamma is done with shape:  (192,)
i3d_inceptionv11_Mixed_4a_batchnorm0_beta is done with shape:  (192,)
i3d_inceptionv11_Mixed_4a_batchnorm0_running_mean is done with shape:  (192,)
i3d_inceptionv11_Mixed_4a_batchnorm0_running_var is done with shape:  (192,)
i3d_inceptionv11_Mixed_4a_conv1_weight is done with shape:  (96, 480, 1, 1, 1)
i3d_inceptionv11_Mixed_4a_batchnorm1_gamma is done with shape:  (96,)
i3d_inceptionv11_Mixed_4a_batchnorm1_beta is done with shape:  (96,)
i3d_inceptionv11_Mixed_4a_batchnorm1_running_mean is done with shape:  (96,)
i3d_inceptionv11_Mixed_4a_batchnorm1_running_var is done with shape:  (96,)
i3d_inceptionv11_Mixed_4a_conv2_weight is done with shape:  (208, 96, 3, 3, 3)
i3d_inceptionv11_Mixed_4a_batchnorm2_gamma is done with shape:  (208,)
i3d_inceptionv11_Mixed_4a_batchnorm2_beta is done with shape:  (208,)
i3d_inceptionv11_Mixed_4a_batchnorm2_running_mean is done with shape:  (208,)
i3d_inceptionv11_Mixed_4a_batchnorm2_running_var is done with shape:  (208,)
i3d_inceptionv11_Mixed_4a_conv3_weight is done with shape:  (16, 480, 1, 1, 1)
i3d_inceptionv11_Mixed_4a_batchnorm3_gamma is done with shape:  (16,)
i3d_inceptionv11_Mixed_4a_batchnorm3_beta is done with shape:  (16,)
i3d_inceptionv11_Mixed_4a_batchnorm3_running_mean is done with shape:  (16,)
i3d_inceptionv11_Mixed_4a_batchnorm3_running_var is done with shape:  (16,)
i3d_inceptionv11_Mixed_4a_conv4_weight is done with shape:  (48, 16, 3, 3, 3)
i3d_inceptionv11_Mixed_4a_batchnorm4_gamma is done with shape:  (48,)
i3d_inceptionv11_Mixed_4a_batchnorm4_beta is done with shape:  (48,)
i3d_inceptionv11_Mixed_4a_batchnorm4_running_mean is done with shape:  (48,)
i3d_inceptionv11_Mixed_4a_batchnorm4_running_var is done with shape:  (48,)
i3d_inceptionv11_Mixed_4a_conv5_weight is done with shape:  (64, 480, 1, 1, 1)
i3d_inceptionv11_Mixed_4a_batchnorm5_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4a_batchnorm5_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4a_batchnorm5_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4a_batchnorm5_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_conv0_weight is done with shape:  (160, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4b_batchnorm0_gamma is done with shape:  (160,)
i3d_inceptionv11_Mixed_4b_batchnorm0_beta is done with shape:  (160,)
i3d_inceptionv11_Mixed_4b_batchnorm0_running_mean is done with shape:  (160,)
i3d_inceptionv11_Mixed_4b_batchnorm0_running_var is done with shape:  (160,)
i3d_inceptionv11_Mixed_4b_conv1_weight is done with shape:  (112, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4b_batchnorm1_gamma is done with shape:  (112,)
i3d_inceptionv11_Mixed_4b_batchnorm1_beta is done with shape:  (112,)
i3d_inceptionv11_Mixed_4b_batchnorm1_running_mean is done with shape:  (112,)
i3d_inceptionv11_Mixed_4b_batchnorm1_running_var is done with shape:  (112,)
i3d_inceptionv11_Mixed_4b_conv2_weight is done with shape:  (224, 112, 3, 3, 3)
i3d_inceptionv11_Mixed_4b_batchnorm2_gamma is done with shape:  (224,)
i3d_inceptionv11_Mixed_4b_batchnorm2_beta is done with shape:  (224,)
i3d_inceptionv11_Mixed_4b_batchnorm2_running_mean is done with shape:  (224,)
i3d_inceptionv11_Mixed_4b_batchnorm2_running_var is done with shape:  (224,)
i3d_inceptionv11_Mixed_4b_conv3_weight is done with shape:  (24, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4b_batchnorm3_gamma is done with shape:  (24,)
i3d_inceptionv11_Mixed_4b_batchnorm3_beta is done with shape:  (24,)
i3d_inceptionv11_Mixed_4b_batchnorm3_running_mean is done with shape:  (24,)
i3d_inceptionv11_Mixed_4b_batchnorm3_running_var is done with shape:  (24,)
i3d_inceptionv11_Mixed_4b_conv4_weight is done with shape:  (64, 24, 3, 3, 3)
i3d_inceptionv11_Mixed_4b_batchnorm4_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_batchnorm4_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_batchnorm4_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_batchnorm4_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_conv5_weight is done with shape:  (64, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4b_batchnorm5_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_batchnorm5_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_batchnorm5_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4b_batchnorm5_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_conv0_weight is done with shape:  (128, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4c_batchnorm0_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_batchnorm0_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_batchnorm0_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_batchnorm0_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_conv1_weight is done with shape:  (128, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4c_batchnorm1_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_batchnorm1_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_batchnorm1_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_batchnorm1_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_4c_conv2_weight is done with shape:  (256, 128, 3, 3, 3)
i3d_inceptionv11_Mixed_4c_batchnorm2_gamma is done with shape:  (256,)
i3d_inceptionv11_Mixed_4c_batchnorm2_beta is done with shape:  (256,)
i3d_inceptionv11_Mixed_4c_batchnorm2_running_mean is done with shape:  (256,)
i3d_inceptionv11_Mixed_4c_batchnorm2_running_var is done with shape:  (256,)
i3d_inceptionv11_Mixed_4c_conv3_weight is done with shape:  (24, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4c_batchnorm3_gamma is done with shape:  (24,)
i3d_inceptionv11_Mixed_4c_batchnorm3_beta is done with shape:  (24,)
i3d_inceptionv11_Mixed_4c_batchnorm3_running_mean is done with shape:  (24,)
i3d_inceptionv11_Mixed_4c_batchnorm3_running_var is done with shape:  (24,)
i3d_inceptionv11_Mixed_4c_conv4_weight is done with shape:  (64, 24, 3, 3, 3)
i3d_inceptionv11_Mixed_4c_batchnorm4_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_batchnorm4_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_batchnorm4_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_batchnorm4_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_conv5_weight is done with shape:  (64, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4c_batchnorm5_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_batchnorm5_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_batchnorm5_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4c_batchnorm5_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_conv0_weight is done with shape:  (112, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4d_batchnorm0_gamma is done with shape:  (112,)
i3d_inceptionv11_Mixed_4d_batchnorm0_beta is done with shape:  (112,)
i3d_inceptionv11_Mixed_4d_batchnorm0_running_mean is done with shape:  (112,)
i3d_inceptionv11_Mixed_4d_batchnorm0_running_var is done with shape:  (112,)
i3d_inceptionv11_Mixed_4d_conv1_weight is done with shape:  (144, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4d_batchnorm1_gamma is done with shape:  (144,)
i3d_inceptionv11_Mixed_4d_batchnorm1_beta is done with shape:  (144,)
i3d_inceptionv11_Mixed_4d_batchnorm1_running_mean is done with shape:  (144,)
i3d_inceptionv11_Mixed_4d_batchnorm1_running_var is done with shape:  (144,)
i3d_inceptionv11_Mixed_4d_conv2_weight is done with shape:  (288, 144, 3, 3, 3)
i3d_inceptionv11_Mixed_4d_batchnorm2_gamma is done with shape:  (288,)
i3d_inceptionv11_Mixed_4d_batchnorm2_beta is done with shape:  (288,)
i3d_inceptionv11_Mixed_4d_batchnorm2_running_mean is done with shape:  (288,)
i3d_inceptionv11_Mixed_4d_batchnorm2_running_var is done with shape:  (288,)
i3d_inceptionv11_Mixed_4d_conv3_weight is done with shape:  (32, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4d_batchnorm3_gamma is done with shape:  (32,)
i3d_inceptionv11_Mixed_4d_batchnorm3_beta is done with shape:  (32,)
i3d_inceptionv11_Mixed_4d_batchnorm3_running_mean is done with shape:  (32,)
i3d_inceptionv11_Mixed_4d_batchnorm3_running_var is done with shape:  (32,)
i3d_inceptionv11_Mixed_4d_conv4_weight is done with shape:  (64, 32, 3, 3, 3)
i3d_inceptionv11_Mixed_4d_batchnorm4_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_batchnorm4_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_batchnorm4_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_batchnorm4_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_conv5_weight is done with shape:  (64, 512, 1, 1, 1)
i3d_inceptionv11_Mixed_4d_batchnorm5_gamma is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_batchnorm5_beta is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_batchnorm5_running_mean is done with shape:  (64,)
i3d_inceptionv11_Mixed_4d_batchnorm5_running_var is done with shape:  (64,)
i3d_inceptionv11_Mixed_4e_conv0_weight is done with shape:  (256, 528, 1, 1, 1)
i3d_inceptionv11_Mixed_4e_batchnorm0_gamma is done with shape:  (256,)
i3d_inceptionv11_Mixed_4e_batchnorm0_beta is done with shape:  (256,)
i3d_inceptionv11_Mixed_4e_batchnorm0_running_mean is done with shape:  (256,)
i3d_inceptionv11_Mixed_4e_batchnorm0_running_var is done with shape:  (256,)
i3d_inceptionv11_Mixed_4e_conv1_weight is done with shape:  (160, 528, 1, 1, 1)
i3d_inceptionv11_Mixed_4e_batchnorm1_gamma is done with shape:  (160,)
i3d_inceptionv11_Mixed_4e_batchnorm1_beta is done with shape:  (160,)
i3d_inceptionv11_Mixed_4e_batchnorm1_running_mean is done with shape:  (160,)
i3d_inceptionv11_Mixed_4e_batchnorm1_running_var is done with shape:  (160,)
i3d_inceptionv11_Mixed_4e_conv2_weight is done with shape:  (320, 160, 3, 3, 3)
i3d_inceptionv11_Mixed_4e_batchnorm2_gamma is done with shape:  (320,)
i3d_inceptionv11_Mixed_4e_batchnorm2_beta is done with shape:  (320,)
i3d_inceptionv11_Mixed_4e_batchnorm2_running_mean is done with shape:  (320,)
i3d_inceptionv11_Mixed_4e_batchnorm2_running_var is done with shape:  (320,)
i3d_inceptionv11_Mixed_4e_conv3_weight is done with shape:  (32, 528, 1, 1, 1)
i3d_inceptionv11_Mixed_4e_batchnorm3_gamma is done with shape:  (32,)
i3d_inceptionv11_Mixed_4e_batchnorm3_beta is done with shape:  (32,)
i3d_inceptionv11_Mixed_4e_batchnorm3_running_mean is done with shape:  (32,)
i3d_inceptionv11_Mixed_4e_batchnorm3_running_var is done with shape:  (32,)
i3d_inceptionv11_Mixed_4e_conv4_weight is done with shape:  (128, 32, 3, 3, 3)
i3d_inceptionv11_Mixed_4e_batchnorm4_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_batchnorm4_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_batchnorm4_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_batchnorm4_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_conv5_weight is done with shape:  (128, 528, 1, 1, 1)
i3d_inceptionv11_Mixed_4e_batchnorm5_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_batchnorm5_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_batchnorm5_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_4e_batchnorm5_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_conv0_weight is done with shape:  (256, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5a_batchnorm0_gamma is done with shape:  (256,)
i3d_inceptionv11_Mixed_5a_batchnorm0_beta is done with shape:  (256,)
i3d_inceptionv11_Mixed_5a_batchnorm0_running_mean is done with shape:  (256,)
i3d_inceptionv11_Mixed_5a_batchnorm0_running_var is done with shape:  (256,)
i3d_inceptionv11_Mixed_5a_conv1_weight is done with shape:  (160, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5a_batchnorm1_gamma is done with shape:  (160,)
i3d_inceptionv11_Mixed_5a_batchnorm1_beta is done with shape:  (160,)
i3d_inceptionv11_Mixed_5a_batchnorm1_running_mean is done with shape:  (160,)
i3d_inceptionv11_Mixed_5a_batchnorm1_running_var is done with shape:  (160,)
i3d_inceptionv11_Mixed_5a_conv2_weight is done with shape:  (320, 160, 3, 3, 3)
i3d_inceptionv11_Mixed_5a_batchnorm2_gamma is done with shape:  (320,)
i3d_inceptionv11_Mixed_5a_batchnorm2_beta is done with shape:  (320,)
i3d_inceptionv11_Mixed_5a_batchnorm2_running_mean is done with shape:  (320,)
i3d_inceptionv11_Mixed_5a_batchnorm2_running_var is done with shape:  (320,)
i3d_inceptionv11_Mixed_5a_conv3_weight is done with shape:  (32, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5a_batchnorm3_gamma is done with shape:  (32,)
i3d_inceptionv11_Mixed_5a_batchnorm3_beta is done with shape:  (32,)
i3d_inceptionv11_Mixed_5a_batchnorm3_running_mean is done with shape:  (32,)
i3d_inceptionv11_Mixed_5a_batchnorm3_running_var is done with shape:  (32,)
i3d_inceptionv11_Mixed_5a_conv4_weight is done with shape:  (128, 32, 3, 3, 3)
i3d_inceptionv11_Mixed_5a_batchnorm4_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_batchnorm4_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_batchnorm4_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_batchnorm4_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_conv5_weight is done with shape:  (128, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5a_batchnorm5_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_batchnorm5_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_batchnorm5_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_5a_batchnorm5_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_conv0_weight is done with shape:  (384, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5b_batchnorm0_gamma is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_batchnorm0_beta is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_batchnorm0_running_mean is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_batchnorm0_running_var is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_conv1_weight is done with shape:  (192, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5b_batchnorm1_gamma is done with shape:  (192,)
i3d_inceptionv11_Mixed_5b_batchnorm1_beta is done with shape:  (192,)
i3d_inceptionv11_Mixed_5b_batchnorm1_running_mean is done with shape:  (192,)
i3d_inceptionv11_Mixed_5b_batchnorm1_running_var is done with shape:  (192,)
i3d_inceptionv11_Mixed_5b_conv2_weight is done with shape:  (384, 192, 3, 3, 3)
i3d_inceptionv11_Mixed_5b_batchnorm2_gamma is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_batchnorm2_beta is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_batchnorm2_running_mean is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_batchnorm2_running_var is done with shape:  (384,)
i3d_inceptionv11_Mixed_5b_conv3_weight is done with shape:  (48, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5b_batchnorm3_gamma is done with shape:  (48,)
i3d_inceptionv11_Mixed_5b_batchnorm3_beta is done with shape:  (48,)
i3d_inceptionv11_Mixed_5b_batchnorm3_running_mean is done with shape:  (48,)
i3d_inceptionv11_Mixed_5b_batchnorm3_running_var is done with shape:  (48,)
i3d_inceptionv11_Mixed_5b_conv4_weight is done with shape:  (128, 48, 3, 3, 3)
i3d_inceptionv11_Mixed_5b_batchnorm4_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_batchnorm4_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_batchnorm4_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_batchnorm4_running_var is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_conv5_weight is done with shape:  (128, 832, 1, 1, 1)
i3d_inceptionv11_Mixed_5b_batchnorm5_gamma is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_batchnorm5_beta is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_batchnorm5_running_mean is done with shape:  (128,)
i3d_inceptionv11_Mixed_5b_batchnorm5_running_var is done with shape:  (128,)
i3d_inceptionv11_dense0_weight is skipped with shape:  (400, 1024)
i3d_inceptionv11_dense0_bias is skipped with shape:  (400,)
I3D_InceptionV1(
  (features): HybridSequential(
    (0): HybridSequential(
      (0): Conv3D(3 -> 64, kernel_size=(7, 7, 7), stride=(2, 2, 2), padding=(3, 3, 3), bias=False)
      (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
      (2): Activation(relu)
    )
    (1): MaxPool3D(size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
    (2): HybridSequential(
      (0): Conv3D(64 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
      (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
      (2): Activation(relu)
    )
    (3): HybridSequential(
      (0): Conv3D(64 -> 192, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
      (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=192)
      (2): Activation(relu)
    )
    (4): MaxPool3D(size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
    (5): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(192 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(192 -> 96, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=96)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(96 -> 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(192 -> 16, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=16)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(16 -> 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=32)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(192 -> 32, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=32)
          (2): Activation(relu)
        )
      )
    )
    (6): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(256 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(256 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(128 -> 192, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=192)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(256 -> 32, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=32)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(32 -> 96, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=96)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(256 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
    )
    (7): MaxPool3D(size=(3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
    (8): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(480 -> 192, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=192)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(480 -> 96, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=96)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(96 -> 208, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=208)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(480 -> 16, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=16)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(16 -> 48, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=48)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(480 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
    )
    (9): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 160, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=160)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 112, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=112)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(112 -> 224, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=224)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 24, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=24)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(24 -> 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(512 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
    )
    (10): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(128 -> 256, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=256)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 24, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=24)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(24 -> 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(512 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
    )
    (11): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 112, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=112)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 144, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=144)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(144 -> 288, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=288)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(512 -> 32, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=32)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(32 -> 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(512 -> 64, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=64)
          (2): Activation(relu)
        )
      )
    )
    (12): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(528 -> 256, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=256)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(528 -> 160, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=160)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(160 -> 320, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=320)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(528 -> 32, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=32)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(32 -> 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(528 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
    )
    (13): MaxPool3D(size=(2, 2, 2), stride=(2, 2, 2), padding=(0, 0, 0), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
    (14): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(832 -> 256, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=256)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(832 -> 160, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=160)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(160 -> 320, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=320)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(832 -> 32, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=32)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(32 -> 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(832 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
    )
    (15): HybridConcurrent(
      (0): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(832 -> 384, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=384)
          (2): Activation(relu)
        )
      )
      (1): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(832 -> 192, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=192)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(192 -> 384, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=384)
          (2): Activation(relu)
        )
      )
      (2): HybridSequential(
        (0): HybridSequential(
          (0): Conv3D(832 -> 48, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=48)
          (2): Activation(relu)
        )
        (1): HybridSequential(
          (0): Conv3D(48 -> 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
      (3): HybridSequential(
        (0): MaxPool3D(size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1), ceil_mode=False, global_pool=False, pool_type=max, layout=NCDHW)
        (1): HybridSequential(
          (0): Conv3D(832 -> 128, kernel_size=(1, 1, 1), stride=(1, 1, 1), bias=False)
          (1): BatchNorm(axis=1, eps=0.001, momentum=0.9, fix_gamma=False, use_global_stats=False, in_channels=128)
          (2): Activation(relu)
        )
      )
    )
    (16): GlobalAvgPool3D(size=(1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0), ceil_mode=True, global_pool=True, pool_type=avg, layout=NCDHW)
  )
  (head): HybridSequential(
    (0): Dropout(p = 0.5, axes=())
    (1): Dense(1024 -> 400, linear)
  )
  (output): Dense(1024 -> 400, linear)
)

数据增强和数据加载器

视频的数据增强与图像不同。例如，如果要随机裁剪视频序列，需要确保序列中的所有视频帧都经过相同的裁剪过程。我们提供了一组新的转换函数，用于处理多个图像。详情请查阅 video.py。此处使用的大多数视频数据增强策略在 [Wang15] 中介绍。

transform_train = transforms.Compose([
    # Fix the input video frames size as 256×340 and randomly sample the cropping width and height from
    # {256,224,192,168}. After that, resize the cropped regions to 224 × 224.
    video.VideoMultiScaleCrop(size=(224, 224), scale_ratios=[1.0, 0.875, 0.75, 0.66]),
    # Randomly flip the video frames horizontally
    video.VideoRandomHorizontalFlip(),
    # Transpose the video frames from height*width*num_channels to num_channels*height*width
    # and map values from [0, 255] to [0,1]
    video.VideoToTensor(),
    # Normalize the video frames with mean and standard deviation calculated across all images
    video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

使用这些转换函数，我们可以为训练数据集定义数据加载器。

# Batch Size for Each GPU
per_device_batch_size = 5
# Number of data loader workers
num_workers = 0
# Calculate effective total batch size
batch_size = per_device_batch_size * num_gpus

# Set train=True for training the model.
# ``new_length`` indicates the number of frames we use as input.
# ``new_step`` indicates we skip one frame to sample the input data.
train_dataset = Kinetics400(train=True, new_length=32, new_step=2, transform=transform_train)
print('Load %d training samples.' % len(train_dataset))
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size,
                                   shuffle=True, num_workers=num_workers)

输出

Load 375 training samples.

优化器、损失函数和评估指标

# Learning rate decay factor
lr_decay = 0.1
# Epochs where learning rate decays
lr_decay_epoch = [40, 80, 100]

# Stochastic gradient descent
optimizer = 'sgd'
# Set parameters
optimizer_params = {'learning_rate': 0.01, 'wd': 0.0001, 'momentum': 0.9}

# Define our trainer for net
trainer = gluon.Trainer(net.collect_params(), optimizer, optimizer_params)

为了优化我们的模型，我们需要一个损失函数。对于分类任务，我们通常使用 softmax 交叉熵作为损失函数。

loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()

为简单起见，我们使用准确率作为评估指标来监控训练过程。此外，我们记录指标值，并在训练结束时打印出来。

train_metric = mx.metric.Accuracy()
train_history = TrainingHistory(['training-acc'])

训练

所有准备工作完成后，我们终于可以开始训练了！以下是脚本。

注意

为了快速完成本教程，我们仅在 Kinetics400 的一个微小子集上训练 0 个 epoch，每个 epoch 100 次迭代。在你的实验中，我们建议在完整的 Kinetics400 数据集上设置 epochs=100。

epochs = 0
lr_decay_count = 0

for epoch in range(epochs):
    tic = time.time()
    train_metric.reset()
    train_loss = 0

    # Learning rate decay
    if epoch == lr_decay_epoch[lr_decay_count]:
        trainer.set_learning_rate(trainer.learning_rate*lr_decay)
        lr_decay_count += 1

    # Loop through each batch of training data
    for i, batch in enumerate(train_data):
        # Extract data and label
        data = split_and_load(batch[0], ctx_list=ctx, batch_axis=0)
        label = split_and_load(batch[1], ctx_list=ctx, batch_axis=0)

        # AutoGrad
        with ag.record():
            output = []
            for _, X in enumerate(data):
                X = X.reshape((-1,) + X.shape[2:])
                pred = net(X)
                output.append(pred)
            loss = [loss_fn(yhat, y) for yhat, y in zip(output, label)]

        # Backpropagation
        for l in loss:
            l.backward()

        # Optimize
        trainer.step(batch_size)

        # Update metrics
        train_loss += sum([l.mean().asscalar() for l in loss])
        train_metric.update(label, output)

        if i == 100:
            break

    name, acc = train_metric.get()

    # Update history and print metrics
    train_history.update([acc])
    print('[Epoch %d] train=%f loss=%f time: %f' %
        (epoch, acc, train_loss / (i+1), time.time()-tic))

# We can plot the metric scores with:
train_history.plot()

由于使用了微小子集，准确率数值较低。您可以在完整的 Kinetics400 数据集上 立即开始训练。

如果您想探索更先进的模型（例如 SlowFast），请随时阅读关于 SlowFast 的下一篇教程。

参考文献

Wang15

Limin Wang, Yuanjun Xiong, Zhe Wang, and Yu Qiao. “Towards Good Practices for Very Deep Two-Stream ConvNets.” arXiv 预印本 arXiv:1507.02159 (2015)。