From Documentation
Jump to: navigation, search
(ImageNet example on Angel and Monk)
Line 212: Line 212:
== ImageNet example on Angel and Monk==
For training AlexNet on ImageNet dataset, you should download the images first and put them into "data" folder under caffe root. Then go to "examples/imagenet/" and modify by adding "--backend=leveldb" and change the name from lmdb to leveldb for both "train" and "val" data like this:
GLOG_logtostderr=1 $TOOLS/convert_imageset \
    --resize_height=$RESIZE_HEIGHT \
    --resize_width=$RESIZE_WIDTH \
    --shuffle \
    --backend=leveldb \
    $DATA/train.txt \
Then you should also modify the "train_val.prototxt" under "models/bvlc_alexnet". You have to change "LMDB" to "LEVELDB" and fix the path to the leveldb database.
data_param {
    source: "examples/imagenet/ilsvrc12_train_leveldb"
    backend: LEVELDB
    batch_size: 256
If running on Angel, you should set a smaller batch_size if you get "out of memory" error. It is caused by the small amount of RAM on the GTX 750Ti(2GB).
After changing the parameters in solver.prototxt, you should be able to run Caffe now.
To submit a job, you should setup a properaiat memory size which is large enough for you network size.
sqsub -q gpu --gpp=1 --mpp=24g -r 24h -o out.txt ./build/tools/caffe train --solver=models/bvlc_alexnet/solver.prototxt
Caffe can read images from the disk (png, jpg) directly using openCV. Use the ImageData layer:
  layers {
name: "data"
  type: IMAGE_DATA
top: "data"
top: "label" image list you want to process
  image_data_param {
    source: "file_list.txt"
    mean_file: "/some/where/to/imagenet_mean.binaryproto"
    crop_size: 227
    new_height: 256
    new_width: 256
} }
The "file_list.txt" file in this example is a text file formatted as follows:
/work/xxx/data/7.png 7
/work/xxx/data/2.png 2
== Bug on Monk ==
The GPU on Monk is relatively old and with only SM2.0 support. For large networks, it will be running out of blocks per grid dim (SM 2.0 had only 65535 blocks per dim, while 3.0 bumped it to 2^31-1). This will happen when using AlexNet. There is no official fix for this problem yet. Based on this webpage[], I modifed the code in "include/util/device_alternate.hpp" to
// CUDA: number of blocks for threads.
inline int CAFFE_GET_BLOCKS(const int N) {
if (num_blocks > 65535)
return num_blocks;//deal with sm20 devices
}  // namespace caffe
== References ===

Revision as of 15:50, 5 October 2016

Caffe is a deep learning framework developed with cleanliness, readability, and speed in mind.

Running Caffe (master branch Jan 29, 2016) on Copper and Mosaic

Sharcnet doesn't maintain Caffe as a module, but we provide all its dependencies and a precompiled Caffe as an example.

All the files that needed to build and run Caffe are under:


Set up modules and environment variables

There is a script ( that can help setup all the modules and environment variables:

module unload intel gcc mkl openmpi hdf python cuda
module load intel/15.0.3
module load hdf/serial/
module load cuda/7.5.18
module load python/intel/2.7.10
export CAFFE_ROOT=/opt/sharcnet/testing/caffe/caffe-master-160129
export PATH=/opt/sharcnet/testing/caffe/caffe-libs/bin:/opt/sharcnet/testing/caffe/caffe-libs/include:/opt/sharcnet/testing/cudnn/cudnn3:$PATH
export LD_LIBRARY_PATH=/opt/sharcnet/testing/caffe/caffe-libs/lib:/opt/sharcnet/testing/cudnn/cudnn3:$LD_LIBRARY_PATH
export PYTHONPATH=/opt/sharcnet/testing/caffe/caffe-libs/lib/python2.7/site-packages:/opt/sharcnet/testing/caffe/caffe-master-160129/python:$PYTHONPATH

In this file, CAFFE_ROOT is set to a precompiled Caffe which was built on Jan 29th, 2016. If you want to use this version of caffe, you don't have to run the commands above, you can simply do:

source /opt/sharcnet/testing/caffe/
  • You should run this "source" command every time you login to a sharcnet system, or put it into your .bashrc if you really know what these commands will do.

Prepare data set (e.g. Imagenet)

Although Caffe can take images directly as an input, for a better I/O performance, LMDB database is strongly recommended. Caffe provides a tool and a script to build lmdb database from images. You should copy the script to your place and modify it to meet your own needs. The script is


In this file, there are some variables you should modify:

  • EXAMPLE, should be a folder that you want to storage the lmdb database file.
  • DATA, should be a folder where you keep the label files, train.txt and val.txt
  • TOOLS, should be tools folder in Caffe, here we use the precompiled Caffe as an example
  • TRAIN_DATA_ROOT and VAL_DATA_ROOT, should be folders that contains your images (e.g. .jpg files)
  • RESIZE, should be set to true if you haven't resized the images to 256x256
  • In the train.txt and val.txt, you should specify the image names with locations and lables. For example:
n01440764/n01440764_10254.JPEG 0
n01440764/n01440764_10281.JPEG 0

Each line contains one image location and label. "n01440764" is the subfolder's name under $TRAIN_DATA_ROOT. The full path to this image should be /work/yourname/imagenet/train/n01440764/n01440764_10254.JPEG

  • You can use data/ilsvrc12/ script in Caffe to download the labels for imagenet.

After building the LMDB database, we should also compute an image mean file (imagenet_mean.binaryproto) based on the data you have. The script is $CAFFE_ROOT/examples/imagenet/ You should modify the variables inside to the same as we modified in If you have run the, an imagenet_mean.binaryproto which was computed on imagenet dataset will be downloaded.

Define models (Alexnet as an example)

Caffe provides an Alexnet model in $CAFFE_ROOT/models/bvlc_alexnet. Here we copy this folder to /home or /work so that we can modify it. First we should modify "train_val.prototxt" file. We should change the "mean_file" to where we hold the image mean. And the "source" for training and validation.

name: "AlexNet"
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  transform_param {
    mirror: true
    crop_size: 227
    mean_file: "/work/feimao/imagenet/imagenet_mean.binaryproto"
  data_param {
    source: "/work/feimao/imagenet/ilsvrc12_train_lmdb"
    batch_size: 256
    backend: LMDB
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  transform_param {
    mirror: false
    crop_size: 227
    mean_file: "/work/feimao/imagenet/imagenet_mean.binaryproto"
  data_param {
    source: "/work/feimao/imagenet/ilsvrc12_val_lmdb"
    batch_size: 50
    backend: LMDB

And then changing the solver.prototxt to locate the train_val.prototxt file and the output snapshot.

net: "/work/feimao/imagenet/bvlc_alexnet/train_val.prototxt"
test_iter: 1000
test_interval: 1000
base_lr: 0.01
lr_policy: "step"
gamma: 0.1
stepsize: 100000
display: 20
max_iter: 450000
momentum: 0.9
weight_decay: 0.0005
snapshot: 10000
snapshot_prefix: "/work/feimao/imagenet/bvlc_alexnet/caffe_alexnet_train"
solver_mode: GPU

Submit jobs

  • On copper and Mosaic, please export MKL_CBWR=AUTO before submitting the jobs. This will be set if you do "source /opt/sharcnet/testing/caffe/"

Once we have solver.prototxt and train_val.prototxt files ready, we can submit jobs using sqsub. The command is

sqsub -q gpu -f threaded -n 2 --gpp=1 --mpp=32g -r 4h -o test-submit-caffe.out $CAFFE_ROOT/build/tools/caffe.bin train \

Caffe need at least 2 cores, one for controlling GPU, one for loading data. The size of memory (mpp) depends on the model size. 32 GB is recommended as a smallest size to try. "gpp" should be one as Caffe (master branch) supports only single GPU for each process.

Building Caffe with cudnn v4(CUDA7.5) on Copper and Mosaic

How to build Caffe

If you'd like to build your own Caffe, you can simply get the newest Caffe from github (notice that you should stay in login node to have internet access):

[feimao@mos-login test-install-caffe]$ git clone

Then you go to mos1 (or cop1), load the modules:

module unload intel mkl openmpi cuda
module load intel/15.0.3
module load hdf/serial/ 
module load cuda/7.5.18 
module load python/intel/2.7.10

And export the environment paths to cudnn and caffe-libs folder:

export PATH=/opt/sharcnet/testing/caffe/caffe-libs/bin:/opt/sharcnet/testing/caffe/caffe-libs/include:/opt/sharcnet/testing/cudnn/cudnn4:$PATH
export LD_LIBRARY_PATH=/opt/sharcnet/testing/caffe/caffe-libs/lib:/opt/sharcnet/testing/cudnn/cudnn4:$LD_LIBRARY_PATH
export PYTHONPATH=/opt/sharcnet/testing/caffe/caffe-libs/lib/python2.7/site-packages:$PYTHONPATH

and go to the Caffe folder and copy the Makefile.config.example to Makefile.config:

[feimao@mos1 caffe]$ cp Makefile.config.example Makefile.config

Then modify the Makefile.config:

  • 1, uncomment
  • 2, uncomment
  • 3, change CUDA_DIR to
CUDA_DIR := /opt/sharcnet/cuda/7.5.18
  • 4, If on Copper, add a line "-gencode arch=compute_37,code=sm_37" to CUDA_ARCH
  • 5, change BLAS to
BLAS := mkl
  • 6, uncomment BLAS_INCLUDE, and change to
BLAS_INCLUDE := /opt/sharcnet/intel/15.0.3/mkl/include
  • 7, uncomment BLAS_LIB, and change to
BLAS_LIB := /opt/sharcnet/intel/15.0.3/mkl/lib/intel64
  • 8, change PYTHON_INCLUDE to
PYTHON_INCLUDE := /opt/sharcnet/python/2.7.10/intel/include \
                /opt/sharcnet/python/2.7.10/intel/include/python2.7 \
  • 9, change PYTHON_LIB to
PYTHON_LIB := /opt/sharcnet/python/2.7.10/intel/lib
  • 10, change INCLUDE_DIRS to
INCLUDE_DIRS := $(PYTHON_INCLUDE) /opt/sharcnet/hdf/ /opt/sharcnet/testing/caffe/caffe-libs/include /opt/sharcnet/testing/cudnn/cudnn4 /usr/local/include
  • 10, change LIBRARY_DIRS to
LIBRARY_DIRS := $(PYTHON_LIB) /opt/sharcnet/hdf/ /opt/sharcnet/testing/caffe/caffe-libs/lib /opt/sharcnet/testing/cudnn/cudnn4 /usr/local/lib /usr/lib
  • 11, run command to build Caffe and pycaffe:
make all -j16 && make test -j16 && make pycaffe

How to test the Caffe build

When the "make all" and "make test" finish, you can run "make runtest" to test if you build Caffe properly. You may get errors related to "test_gradient_based_solver.cpp" or other numerical inconsistency between CPU and GPU. Those errors was caused by MKL's Conditional Numerical Reproducibility setting. To avoid these errors, you should set MKL_CBWR: