.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "build/examples_action_recognition/dive_deep_i3d_kinetics400.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note Click :ref:`here ` to download the full example code .. rst-class:: sphx-glr-example-title .. _sphx_glr_build_examples_action_recognition_dive_deep_i3d_kinetics400.py: 4. Dive Deep into Training I3D mdoels on Kinetcis400 ======================================================= This is a video action recognition tutorial using Gluon CV toolkit, a step-by-step example. The readers should have basic knowledge of deep learning and should be familiar with Gluon API. New users may first go through `A 60-minute Gluon Crash Course `_. You can `Start Training Now`_ or `Dive into Deep`_. Start Training Now ~~~~~~~~~~~~~~~~~~ .. note:: Feel free to skip the tutorial because the training script is self-complete and ready to launch. :download:`Download Full Python Script: train_recognizer.py<../../../scripts/action-recognition/train_recognizer.py>` For more training command options, please run ``python train_recognizer.py -h`` Please checkout the `model_zoo <../model_zoo/index.html#action_recognition>`_ for training commands of reproducing the pretrained model. Network Structure ----------------- First, let's import the necessary libraries into python. .. GENERATED FROM PYTHON SOURCE LINES 28-46 .. code-block:: default 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 .. GENERATED FROM PYTHON SOURCE LINES 47-54 Here we pick a widely adopted model, ``I3D-InceptionV1``, for the tutorial. `I3D `_ (Inflated 3D Networks) is a widely adopted 3D video classification network. It uses 3D convolution to learn spatiotemporal information directly from videos. I3D is proposed to improve C3D model by inflating from 2D models. We can not only reuse the 2D models' architecture (e.g., ResNet, Inception), but also bootstrap the model weights from 2D pretrained models. In this manner, training 3D networks for video classification is feasible and getting much better results. .. GENERATED FROM PYTHON SOURCE LINES 55-65 .. code-block:: default # 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) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none 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 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) ) .. GENERATED FROM PYTHON SOURCE LINES 66-75 Data Augmentation and Data Loader --------------------------------- Data augmentation for video is different from image. For example, if you want to randomly crop a video sequence, you need to make sure all the video frames in this sequence undergo the same cropping process. We provide a new set of transformation functions, working with multiple images. Please checkout the `video.py <../../../gluoncv/data/transforms/video.py>`_ for more details. Most video data augmentation strategies used here are introduced in [Wang15]_. .. GENERATED FROM PYTHON SOURCE LINES 75-89 .. code-block:: default 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]) ]) .. GENERATED FROM PYTHON SOURCE LINES 90-92 With the transform functions, we can define data loaders for our training datasets. .. GENERATED FROM PYTHON SOURCE LINES 92-108 .. code-block:: default # 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) .. rst-class:: sphx-glr-script-out Out: .. code-block:: none Load 375 training samples. .. GENERATED FROM PYTHON SOURCE LINES 109-111 Optimizer, Loss and Metric -------------------------- .. GENERATED FROM PYTHON SOURCE LINES 111-125 .. code-block:: default # 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) .. GENERATED FROM PYTHON SOURCE LINES 126-129 In order to optimize our model, we need a loss function. For classification tasks, we usually use softmax cross entropy as the loss function. .. GENERATED FROM PYTHON SOURCE LINES 129-132 .. code-block:: default loss_fn = gluon.loss.SoftmaxCrossEntropyLoss() .. GENERATED FROM PYTHON SOURCE LINES 133-136 For simplicity, we use accuracy as the metric to monitor our training process. Besides, we record metric values, and will print them at the end of training. .. GENERATED FROM PYTHON SOURCE LINES 136-140 .. code-block:: default train_metric = mx.metric.Accuracy() train_history = TrainingHistory(['training-acc']) .. GENERATED FROM PYTHON SOURCE LINES 141-150 Training -------- After all the preparations, we can finally start training! Following is the script. .. note:: In order to finish the tutorial quickly, we only train for 0 epoch on a tiny subset of Kinetics400, and 100 iterations per epoch. In your experiments, we recommend setting ``epochs=100`` for the full Kinetics400 dataset. .. GENERATED FROM PYTHON SOURCE LINES 150-203 .. code-block:: default 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() .. image-sg:: /build/examples_action_recognition/images/sphx_glr_dive_deep_i3d_kinetics400_001.png :alt: dive deep i3d kinetics400 :srcset: /build/examples_action_recognition/images/sphx_glr_dive_deep_i3d_kinetics400_001.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 204-216 Due to the tiny subset, the accuracy number is quite low. You can `Start Training Now`_ on the full Kinetics400 dataset. If you would like to explore more recent models (e.g., SlowFast), feel free to read the next `tutorial on SlowFast `__. References ---------- .. [Wang15] Limin Wang, Yuanjun Xiong, Zhe Wang, and Yu Qiao. \ "Towards Good Practices for Very Deep Two-Stream ConvNets." \ arXiv preprint arXiv:1507.02159 (2015). .. rst-class:: sphx-glr-timing **Total running time of the script:** ( 0 minutes 1.419 seconds) .. _sphx_glr_download_build_examples_action_recognition_dive_deep_i3d_kinetics400.py: .. only :: html .. container:: sphx-glr-footer :class: sphx-glr-footer-example .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: dive_deep_i3d_kinetics400.py ` .. container:: sphx-glr-download sphx-glr-download-jupyter :download:`Download Jupyter notebook: dive_deep_i3d_kinetics400.ipynb ` .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_