Caffe:ImageNetTutorial

lastmod 2025-02-14created 2015-10-1631 edits

Notes on how to train and test with Caffe.

The terms

• Training images: List of images for training. Stored in train.txt along with label information.
• Validation images: List of images and labels for validation. Stored in val.txt along with label information.
• deploy:
• solver:

Preparing the data

Get the image data from ImageNet. The obtained TEXT file containing the list of URLs can be downloaded as follows.

#!/bin/sh

WORKING=$PWD
URLS_FILE=$WORKING/fall11_urls.txt
DEST=$WORKING/train

if [[ ! -d $DEST ]]; then
    mkdir $DEST
fi

cd $DEST

COUNT=`wc -l $URLS_FILE | awk '{ print $1 }'`
INDEX=0

echo "Total count: $COUNT"
echo "Download start!"

for path in $(cat $URLS_FILE); do
    let "INDEX = INDEX + 1"
    echo "Download: $INDEX/$COUNT"
    wget --timeout=10 --tries=2 $path
done

echo "Download end."

Place the training images (train) and validation images (val) on disk as follows.

/path/to/imagenet/train/n01440764/n01440764_10026.JPEG
/path/to/imagenet/val/ILSVRC2012_val_00000001.JPEG

Also, download the auxiliary data for training using the script below. (For reference, it is about 17M in size)

./data/ilsvrc12/get_ilsvrc_aux.sh

When it finishes, the following data is created in the Caffe root directory. ¹

det_synset_words.txt
imagenet_mean.binaryproto
synset_words.txt
synsets.txt
test.txt
train.txt
val.txt
imagenet.bet.pickle

The training list (train.txt) and validation list (val.txt) files contain the list of files along with each file's label. For the definitions of the labels, see the synset_words.txt file.

Optionally, at the user's choice, you can resize the images in advance.²

for name in /path/to/imagenet/val/*.JPEG; do
    convert -resize 256x256\! $name $name
done

The reason this is not done explicitly is that you can benefit from resizing the images in parallel using mapreduce.³

Open the examples/imagenet/create_imagenet.sh file, set the paths of the training folder and validation folder (TRAIN_DATA_ROOT and VAL_DATA_ROOT respectively), and apply RESIZE=true if you did not resize the images in advance.

You can easily create the Level DB by running the examples/imagenet/create_imagenet.sh file. Note that when using this script, the locations of the train.txt and val.txt files may differ, so double-check the script.

Computing the image mean

The model requires the mean image from each image.⁴ (Why?; this needs to be looked into.) ⁵ The mean can be computed as follows.

$ ./examples/imagenet/make_imagenet_mean.sh

As a result, the data/ilsvrc12/imagenet_mean.binaryproto file is created.

What is it about?

• Stackoverflow: Is it necessary to apply mean-subtraction on FC7 features of an CNN before feeding into L-SVM?
• Caffe Users group: What exactly is the meanfile do?

The following answer was posted.

It computes the mean pixel by pixel across the whole data-set (look at the compute_image_mean script). So say you had a constant background in the top half of all your images, that would all become zero'd out from mean subtraction.

Mean subtraction (mean-subtraction) needs to be looked into.

Defining the model

First, an explanation of the reference implementation is needed. ⁶

The network definition follows models/bvlc_reference_caffenet/train_val.prototxt. If you need to modify it, open that file. Looking at the include sections in that file, you will find the following.

• include { phase: TRAIN }
• include { phase: TEST }

These differ in their input and output layers.

Input layer differences

Data for the training (TRAIN) network is randomly mirrored (Randomly mirrors) from the examples/imagenet/ilsvrc12_train_leveldb file.

Data for the test (TEST) network is obtained from the examples/imagenet/ilsvrc12_val_leveldb file, and random mirroring is not performed (does not perform random mirroring).

Output layer differences

Both networks output a softmax_loss layer.

During training, it computes the loss function and initializes the backpropagation (Initialize the backpropagation).⁷

The test network has a second output layer that outputs the accuracy on the data (Test set).

In the process of training, the test network will occasionally be instantiated and tested on the test set, producing lines like Test score #0: xxx and Test score #1: xxx. In this case score 0 is the accuracy (which will start around 1/1000 = 0.001 for an untrained network) and score 1 is the loss (which will start around 7 for an untrained network).

Plans

Set up the protocol buffer (Protocol buffer) for the Solver.

• We will run in batches of 256, and run a total of 450,000 iterations (about 90 epochs).
• For every 1,000 iterations, we test the learned net on the validation data.
• We set the initial learning rate to 0.01, and decrease it every 100,000 iterations (about 20 epochs).
• Information will be displayed every 20 iterations.
• The network will be trained with momentum 0.9 and a weight decay of 0.0005.
• For every 10,000 iterations, we will take a snapshot of the current status.

The above is implemented in models/bvlc_reference_caffenet/solver.prototxt.

Starting ImageNet training

You can start training with the command below.

$ ./build/tools/caffe train --solver=models/bvlc_reference_caffenet/solver.prototxt

To change the number of Forward & Backward Iterations, change the max_iter property in the models/bvlc_reference_caffenet/solver.prototxt file.

Tip

If you want to check the computation time, run the following.

$ ./build/tools/caffe time --model=models/bvlc_reference_caffenet/train_val.prototxt

Resume Training?

If a power outage occurs due to war or an earthquake or training is canceled, you can restart it.⁸ Run it as follows.

$ ./build/tools/caffe train --solver=models/bvlc_reference_caffenet/solver.prototxt --snapshot=models/bvlc_reference_caffenet/caffenet_train_iter_10000.solverstate

Next

Once training is complete, perform classification (Classifying) using the Python interface: Caffe:ImageNetClassifying

Sample

See the following files.

• ./examples/imagenet/create_imagenet.sh
• ./examples/imagenet/make_imagenet_mean.sh
• ./examples/imagenet/train_imagenet.sh

All in one script

A script that does everything in one go is as follows.

#!/bin/sh

DATE_FORMAT=`date +%Y%m%d_%H-%M-%S`
LOG_FILE=brewing_imagenet-$DATE_FORMAT.log

echo "LOGFILE: $LOG_FILE"

echo "Create ImageNet."
./examples/imagenet/create_imagenet.sh >& ${LOG_FILE}.1

if [[ "$?" != "0" ]]; then; exit 1; fi

echo "Make ImageNet mean."
./examples/imagenet/make_imagenet_mean.sh >& ${LOG_FILE}.2

if [[ "$?" != "0" ]]; then; exit 1; fi

echo "Train CaffeNet."
./examples/imagenet/train_caffenet.sh >& ${LOG_FILE}.3

if [[ "$?" != "0" ]]; then; exit 1; fi

echo "Done."
exit 0

Use files

The list of files needed to follow the tutorial above is as follows. If possible, copy them to a separate directory before proceeding.

• data/ilsvrc12/ directory.
  • Text files such as train.txt, val.txt.
  • imagenet_mean.binaryproto: a file generated along the way.
• examples/imagenet/ directory.
  • Script files such as create_imagenet.sh.
  • ilsvrc12_train_lmdb/, ilsvrc12_val_lmdb/ directories (DB files storing the image data)
• models/bvlc_reference_caffenet/ directory.
  • *.solverstate and *.caffemodel files. (intermediate saved data and model files)
  • Text files such as solver.prototxt, train_val.prototxt.
• tools/ directory. (contains the executables)

Custom all in one script

A sample file containing the above is as follows. (Compiled on CentOS6.)

• Caffe_copy.tar.gz (OLD VERSION)

An all-in-one script using run.sh is uploaded below. (Compiled on Ubuntu.)

Troubleshooting

std::length_error

• Error in make_imagenet_mean.sh : 'std::length_error'

Using make_imagenet_mean.sh produces the following error.

[your@server caffe]$ ./examples/imagenet/make_imagenet_mean.sh 
terminate called after throwing an instance of 'std::length_error'
  what():  basic_string::_S_create
./examples/imagenet/make_imagenet_mean.sh: line 10: 17802 Aborted                 (core dumped) $TOOLS/compute_image_mean $EXAMPLE/ilsvrc12_train_lmdb $DATA/imagenet_mean.binaryproto
Done.

In this case, the images may not have been read properly when running the previous step, the create_imagenet.sh script.

[your@server caffe]$ ./examples/imagenet/create_imagenet.sh 
Creating train lmdb...
I1020 13:51:01.576401 17461 convert_imageset.cpp:82] Shuffling data
I1020 13:51:01.744885 17461 convert_imageset.cpp:85] A total of 2 images.
I1020 13:51:01.745170 17461 db_lmdb.cpp:23] Opened lmdb examples/imagenet/ilsvrc12_train_lmdb
E1020 13:51:01.745205 17461 io.cpp:80] Could not open or find file /home/your/Project/caffe/examples/imagenet/testnofire.jpeg
E1020 13:51:01.745221 17461 io.cpp:80] Could not open or find file /home/your/Project/caffe/examples/imagenet/testfire.jpg
Creating val lmdb...
I1020 13:51:01.811265 17463 convert_imageset.cpp:82] Shuffling data
I1020 13:51:01.954166 17463 convert_imageset.cpp:85] A total of 2 images.
I1020 13:51:01.954432 17463 db_lmdb.cpp:23] Opened lmdb examples/imagenet/ilsvrc12_val_lmdb
E1020 13:51:01.954473 17463 io.cpp:80] Could not open or find file /home/your/Project/caffe/examples/imagenet/valfire.jpg
E1020 13:51:01.954481 17463 io.cpp:80] Could not open or find file /home/your/Project/caffe/examples/imagenet/valnofire.jpeg
Done.

As shown, re-check the output of the create_imagenet.sh script.

The case above is one where the file paths were wrong.

See also

• Caffe
• Caffe:ImageNetClassifying

Favorite site

• Caffe: Brewing ImageNet
• How to Setup Caffe to Use Other Datasets by Modifying imagenet Example
• Blog.naver.com/sssmate1 Caffe_Tutorial

Guide

• Any simple example? #550
• CS231n Caffe Tutorial ⁹
• Train and Test LeNet on your own dataset
• Caffe tutorial ¹⁰
• Stackoverflow: How to train a caffe model?
• Caffe: Decaf, but Better (explanation of Caffe Layers)

References