transfer-learning: A Python repository from ckhung

This is a collection of notes and codes about my successful attempt at transfer learning.

Start a floydhub deep learning docker
pip install opencv-python lmdb
Create lmdb's from pics: ./pic2lmdb.py wnid-apes.txt /path/to/ape/pics
Compute mean: $CAFFE_ROOT/build/tools/compute_image_mean -backend=lmdb /root/shared/imnet/training/ mean.binaryproto
Train: $CAFFE_ROOT/build/tools/caffe train --solver=/root/shared/imnet/ape/solver.prototxt --weights $CAFFE_ROOT/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel ; date) 2>&1 | tee train.log
Find the mean pixel: ./bpt2npy.py mean.binaryproto mean.npy This will print a 3-element list that looks like '[103.939, 116.779, 123.68]' Replace the --mean '...' values in the following command with these values. Why? Visit VGG's CNN page and find "mean pixel" within the two "information page" links -- we don't really need the mean files at classification time. Simply having the mean pixel values is good enough.
Test on unknown images: ./cnclassify.py -f csv -c /root/shared/imnet/ape/ --labels wnid-apes.txt --mean '[103.939, 116.779, 123.68]' --model deploy.prototxt --weights _iter_4000.caffemodel Note: By putting all the relevant files in /root/shared/imnet/ape/ you can leave out the paths and just specify the file names for --labels, --model, and --weights .

You may need to change file and directory paths on the command line as well as in the *.prototxt files. Try ./pic2lmdb.py -h and ./cnclassify.py -h to see more options.

Code is forked from Adil Moujahid's deeplearning-cats-dogs-tutorial and extensively modified to make it more generic. Also see his wonderful blog post: A Practical Introduction to Deep Learning with Caffe and Python explaining the code.

Use vimdiff or similar editors to study the difference between my *.prototxt files and the corresponding files from caffe or from Adil Moujahid. Also see these two tips for parameter setting in *.prototxt: Running Over Whole Sets/Computing Epochs Instead of Iterations, Choosing batch sizes and tuning sgd

[2018/5/24] See This job at floydhub for another demo: transfer learning for classifying a few kinds of fruits. All the code, config, and data are available for you to reproduce my experiment. Then, suppose you have downloaded fruit_iter_60.caffemodel from my output and have shared all the required files (via -v ...:/SH) into /SH of your local docker, you can create the validation statistics:

--model /SH/code/deploy.prototxt
--weights /SH/output/fruit_iter_2000.caffemodel
--labels /SH/code/fruit-wnid.txt
$(perl -pe 's#/fruit/#/SH/fruit/#'
/SH/fruit-lmdb/validation/index.txt)
> /SH/code/validation.csv```
Finally generate a table of correct/incorrect
counts by labels: ```grep , validation.csv | cut -d , -f 3- |
sed 's# ##g; s#/SH/fruit/##; s#-.*##' |
python wnidsubst.py -w fruit-wnid.txt |
python tabcc.py```
Note that wnidsubst.py and tabcc.py are newly written
scripts only present in this repo and not found in the floydhub job.

I hope the following illustration is helpful for other
deep learning newbies behind me.
![files needed in the process of caffe transfer learning](tlprocess.svg)

ckhung/transfer-learning