代码拉取完成，页面将自动刷新
This is 3D convolution (C3D) and video reader implementation in the latest Caffe (Dec 2016). The original Facebook C3D implementation is branched out from Caffe on July 17, 2014 with git commit b80fc86, and has not been rebased with the original Caffe, hence missing out quite a few new features in the recent Caffe. I therefore pulled in C3D concept and an accompanying video reader and applied to the latest Caffe, and will try to rebase this repo with the upstream whenever there is a new important feature. This repo is rebased on 99bd997, on Aug`21 2018. Please reach me for any feedback or question.
Check out the original Caffe readme for Caffe-specific information.
refactor branch is a recent re-work, based on the original Caffe and Nd convolution and pooling with cuDNN PR. This is a cleaner, less-hacky implementation of 3D convolution/pooling than the master branch, and is supposed to more stable than the master branch. So, feel free to try this branch. One missing feature in the refactor branch (yet) is the python wrapper.
In addition to prerequisites for Caffe, video-caffe depends on cuDNN. It is known to work with CuDNN verson 4 and 5, but it may need some efforts to build with v3.
Makefile.config point to the right paths for CUDA and CuDNN.CUDNN_INCLUDE and CUDNN_LIBRARY are correct. If not, you may want something like cmake -DCUDNN_INCLUDE="/your/path/to/include" -DCUDNN_LIBRARY="/your/path/to/lib" ${video-caffe-root}.Key steps to build video-caffe are:
git clone git@github.com:chuckcho/video-caffe.gitcd video-caffemkdir build && cd buildcmake ..make all -j8make installmake runtest
Look at ${video-caffe-root}/examples/c3d_ucf101/c3d_ucf101_train_test.prototxt for how 3D convolution and pooling are used. In a nutshell, use NdConvolution or NdPooling layer with {kernel,stride,pad}_shape that specifies 3D shapes in (L x H x W) where L is the temporal length (usually 16).
...
# ----- video/label input -----
layer {
name: "data"
type: "VideoData"
top: "data"
top: "label"
video_data_param {
source: "examples/c3d_ucf101/c3d_ucf101_train_split1.txt"
batch_size: 50
new_height: 128
new_width: 171
new_length: 16
shuffle: true
}
include {
phase: TRAIN
}
transform_param {
crop_size: 112
mirror: true
mean_value: 90
mean_value: 98
mean_value: 102
}
}
...
# ----- 1st group -----
layer {
name: "conv1a"
type: "NdConvolution"
bottom: "data"
top: "conv1a"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 64
kernel_shape { dim: 3 dim: 3 dim: 3 }
stride_shape { dim: 1 dim: 1 dim: 1 }
pad_shape { dim: 1 dim: 1 dim: 1 }
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
...
layer {
name: "pool1"
type: "NdPooling"
bottom: "conv1a"
top: "pool1"
pooling_param {
pool: MAX
kernel_shape { dim: 1 dim: 2 dim: 2 }
stride_shape { dim: 1 dim: 2 dim: 2 }
}
}
...
Scripts and training files for C3D training on UCF-101 are located in examples/c3d_ucf101/. Steps to train C3D on UCF-101:
unrar x UCF101.rar
${video-caffe-root}/examples/c3d_ucf101/extract_UCF-101_frames.sh.${video-caffe-root}/examples/c3d_ucf101/c3d_ucf101_{train,test}_split1.txt to correctly point to UCF-101 videos or directories that contain extracted frames.${video-caffe-root}/examples/c3d_ucf101/c3d_ucf101_train_test.prototxt to your taste or HW specification. Especially batch_size may need to be adjusted for the GPU memory.cd ${video-caffe-root} && examples/c3d_ucf101/train_ucf101.sh (optionally use --gpu to use multiple GPU's)${video-caffe-root}/tools/extra/plot_training_loss.sh to get training loss / validation accuracy (top1/5) plot. It's pretty hacky, so look at the file to meet your need.A typical training will yield the following loss and top-1 accuracy: 
A pre-trained model is available (downloadable link) for UCF101 (trained from scratch), achieving top-1 accuracy of ~47%.
Caffe is released under the BSD 2-Clause license.
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。
马建仓 AI 助手
