In this repository we will try to build image classification prediction model based on Convolutional Neural Networks architecture using MXNet library for Deep Learning. The data set consist of 25 000 pictures of dogs and cats. The provided images has different sizes.
Make sure that you have Xcode command line tools installed on your macOS. Run following command to install if missed:
sudo xcode-select --install
Next we will install mxnet from source code as it seems the best way to do this now. The installation script will use Homebrew, so make sure that you have following:
echo 'import site; site.addsitedir("/usr/local/lib/python2.7/site-packages")' >> ~/Library/Python/2.7/lib/python/site-packages/homebrew.pth
Next download MXNet source code and run installation script
# Clone mxnet repository. In terminal, run the commands WITHOUT "sudo"
git clone https://github.com/dmlc/mxnet.git ~/mxnet --recursive
# Install MXNet for Python with all required dependencies
cd ~/mxnet/setup-utils
bash install-mxnet-osx-python.sh
Our goal is to generate two files, imgdata_train.rec for training and imgdata_val.rec for validation, and the former contains 95% images. This can be done with following steps.
First download train data set from here and unpack it.
Then create corresponding directories and make sure that images belonging to the same class are placed in the same directory:
- cats - into directory named
data/full_train_data/cat - dogs - into directory named
data/full_train_data/dog
Make sure to create output directories where generated data sets will be stored.
mkdir data/data_set_train
mkdir data/data_set_test
After that we will prepare two .lst files, which consist of the labels and image paths can be used for generating rec files.
python tools/im2rec.py --list True --recursive True --train-ratio 0.95 data/data_set_train/imgdata data/full_train_data
As result two files with image lists will be generated:
data/data_set_train/imgdata_train.lst- train data listdata/data_set_train/imgdata_val.lst- validation data list
The class labels will be generated: 0 - cat, 1 - dog
Finaly we will generate the .rec files. We will resize the images such that the short edge is at least 32px and save them with 95/100 quality. We also use 16 threads to accelerate the packing. I believe that bigger image sizes may improve prediction accuracy, but it will result in dramatic increase of learning time, which is which is prohibitive in my hardware setup.
python tools/im2rec.py --resize 32 --quality 95 --num-thread 16 data/data_set_train/imgdata data/full_train_data
The resulting records will be generated in directory: data/data_set/
Start training with following command
python src/dogs_vs_cats_train.py --data-train "data/data_set_train/imgdata_train.rec" --data-val "data/data_set_train/imgdata_val.rec" --image-shape 3,32,32 --network resnet --num-layers 18 --batch-size 128 --num-examples 25000
I believe that bigger number of layers may improve prediction accuracy, but it will result in dramatic increase of learning time, which is prohibitive in my hardware setup.
The sample output for first the four epoch's runs are following:
python src/dogs_vs_cats_train.py --image-shape 3,32,32 --network resnet --num-layers 18 --batch-size 128 --num-examples 25000
INFO:root:start with arguments Namespace(batch_size=128, benchmark=0, data_nthreads=4, data_train='data/data_set/imgdata_train.rec', data_val='data/data_set/imgdata_val.rec', disp_batches=20, gpus=None, image_shape='3,32,32', kv_store='device', load_epoch=None, lr=0.05, lr_factor=0.1, lr_step_epochs='200,250', max_random_aspect_ratio=0, max_random_h=36, max_random_l=50, max_random_rotate_angle=0, max_random_s=50, max_random_scale=1, max_random_shear_ratio=0, min_random_scale=1, model_prefix='./out/model_dogs_vs_cats', mom=0.9, monitor=0, network='resnet', num_classes=2, num_epochs=300, num_examples=25000, num_layers=18, optimizer='sgd', pad_size=4, random_crop=1, random_mirror=1, rgb_mean='123.68,116.779,103.939', test_io=0, top_k=0, wd=0.0001)
[20:50:28] src/io/iter_image_recordio.cc:221: ImageRecordIOParser: data/data_set/imgdata_train.rec, use 1 threads for decoding..
[20:50:29] src/io/iter_image_recordio.cc:221: ImageRecordIOParser: data/data_set/imgdata_val.rec, use 1 threads for decoding..
INFO:root:Epoch[0] Batch [20] Speed: 2.54 samples/sec Train-accuracy=0.524182
INFO:root:Epoch[0] Batch [40] Speed: 2.59 samples/sec Train-accuracy=0.538281
INFO:root:Epoch[0] Batch [60] Speed: 2.63 samples/sec Train-accuracy=0.536328
INFO:root:Epoch[0] Batch [80] Speed: 2.53 samples/sec Train-accuracy=0.600000
INFO:root:Epoch[0] Batch [100] Speed: 2.50 samples/sec Train-accuracy=0.625391
INFO:root:Epoch[0] Batch [120] Speed: 2.55 samples/sec Train-accuracy=0.626172
INFO:root:Epoch[0] Batch [140] Speed: 2.50 samples/sec Train-accuracy=0.561328
INFO:root:Epoch[0] Batch [160] Speed: 2.55 samples/sec Train-accuracy=0.637500
INFO:root:Epoch[0] Batch [180] Speed: 2.46 samples/sec Train-accuracy=0.619141
INFO:root:Epoch[0] Train-accuracy=0.670312
INFO:root:Epoch[0] Time cost=9360.959
INFO:root:Saved checkpoint to "./out/model_dogs_vs_cats-0001.params"
INFO:root:Epoch[0] Validation-accuracy=0.706250
INFO:root:Epoch[1] Batch [20] Speed: 2.53 samples/sec Train-accuracy=0.593378
INFO:root:Epoch[1] Batch [40] Speed: 2.65 samples/sec Train-accuracy=0.657813
INFO:root:Epoch[1] Batch [60] Speed: 2.65 samples/sec Train-accuracy=0.663281
INFO:root:Epoch[1] Batch [80] Speed: 2.67 samples/sec Train-accuracy=0.674219
INFO:root:Epoch[1] Batch [100] Speed: 2.67 samples/sec Train-accuracy=0.688281
INFO:root:Epoch[1] Batch [120] Speed: 2.67 samples/sec Train-accuracy=0.699219
INFO:root:Epoch[1] Batch [140] Speed: 2.67 samples/sec Train-accuracy=0.717578
INFO:root:Epoch[1] Batch [160] Speed: 2.67 samples/sec Train-accuracy=0.692187
INFO:root:Epoch[1] Batch [180] Speed: 2.67 samples/sec Train-accuracy=0.712109
INFO:root:Epoch[1] Train-accuracy=0.707812
INFO:root:Epoch[1] Time cost=8958.407
INFO:root:Saved checkpoint to "./out/model_dogs_vs_cats-0002.params"
INFO:root:Epoch[1] Validation-accuracy=0.738281
INFO:root:Epoch[2] Batch [20] Speed: 2.67 samples/sec Train-accuracy=0.681920
INFO:root:Epoch[2] Batch [40] Speed: 2.67 samples/sec Train-accuracy=0.730859
INFO:root:Epoch[2] Batch [60] Speed: 2.66 samples/sec Train-accuracy=0.739453
INFO:root:Epoch[2] Batch [80] Speed: 2.67 samples/sec Train-accuracy=0.742188
INFO:root:Epoch[2] Batch [100] Speed: 2.67 samples/sec Train-accuracy=0.745703
INFO:root:Epoch[2] Batch [120] Speed: 2.67 samples/sec Train-accuracy=0.739453
INFO:root:Epoch[2] Batch [140] Speed: 2.67 samples/sec Train-accuracy=0.735156
INFO:root:Epoch[2] Batch [160] Speed: 2.67 samples/sec Train-accuracy=0.755078
INFO:root:Epoch[2] Batch [180] Speed: 2.67 samples/sec Train-accuracy=0.758984
INFO:root:Epoch[2] Train-accuracy=0.767578
INFO:root:Epoch[2] Time cost=8847.880
INFO:root:Saved checkpoint to "./out/model_dogs_vs_cats-0003.params"
INFO:root:Epoch[2] Validation-accuracy=0.592969
INFO:root:Epoch[3] Batch [20] Speed: 2.67 samples/sec Train-accuracy=0.766369
INFO:root:Epoch[3] Batch [40] Speed: 2.67 samples/sec Train-accuracy=0.780859
INFO:root:Epoch[3] Batch [60] Speed: 2.67 samples/sec Train-accuracy=0.756641
INFO:root:Epoch[3] Batch [80] Speed: 2.67 samples/sec Train-accuracy=0.772656
INFO:root:Epoch[3] Batch [100] Speed: 2.67 samples/sec Train-accuracy=0.774219
INFO:root:Epoch[3] Batch [120] Speed: 2.67 samples/sec Train-accuracy=0.757812
INFO:root:Epoch[3] Batch [140] Speed: 2.67 samples/sec Train-accuracy=0.776953
INFO:root:Epoch[3] Batch [160] Speed: 2.67 samples/sec Train-accuracy=0.780078
INFO:root:Epoch[3] Batch [180] Speed: 2.66 samples/sec Train-accuracy=0.796875
INFO:root:Epoch[3] Train-accuracy=0.782813
INFO:root:Epoch[3] Time cost=8901.720
INFO:root:Saved checkpoint to "./out/model_dogs_vs_cats-0004.params"
INFO:root:Epoch[3] Validation-accuracy=0.823785
As it can be seen validation accuracy increase, but prediction model should converge in order to gurantee that our CNN is learning something. Unfortunatelly due to hardware limitations I was unable to test over all 300 epochs in order to check if model converge, but interested parties may try this with better hardware setup. With my hardware setup (CPU only, 2.5 GHz Intel Core i3, 8Gb RAM, SSD) training over 300 epochs would take about 740 hours (or one month). I believe that within CUDA-enabled setup it will take rather hours.
The trained models for first four epochs available for download here
After model training complete we will prepare test data set. And then do predictions over it.
First download test data files from here and unpack it.
Then create list of images in the data set
python tools/im2rec.py --list True --recursive False --shuffle False --train-ratio 0.0 --test-ratio 1.0 data/data_set_test/imgdata data/test
After run records generation
python tools/im2rec.py --resize 32 --quality 95 --num-thread 16 --recursive False --shuffle False --train-ratio 0.0 --test-ratio 1.0 data/data_set_test/imgdata data/test
Finally run prediction over test data set and trained model
python src/dogs_vs_cats_predict.py --image-shape 3,32,32 --rec-prefix imgdata_test --model out/model_dogs_vs_cats --model-checkpoint 4 out/results data/data_set_test
The predictions output will be stored as coma separated file into out/results directory.
The sample results output file can be found here