Simple tensorflow classification example codes. Works on small image datasets.
Supporting network achitectures includes standard networks under tensorflow.contrib.slim.nets
@@alexnet_v2
@@inception_v1
@@inception_v1_base
@@inception_v2
@@inception_v2_base
@@inception_v3
@@inception_v3_base
@@overfeat
@@vgg_a
@@vgg_16
TFRecords: data are pre-processed into TFRecords format
-
convert_to_records.py: create TFRecords data file from raw images
-
example data: logo apperance binary classification training dataset: around 800 images validation dataset: around 200 images
-
train.py train a VGG network by default
-
test.py restore checkpoint and test network on validation dataset
-
train2.py train and do validation periodically.
Picpac: data are pre-processed into picpac format
- train_picpac.py
Keep a training log by specify --log_dir
before starting train.py. While training, visualise learning on TensorBoard by command
$ tensorboard --logdir=path/to/log-directory
Networks can be trained on GPU with tensorflow installed, or in a docker container on CPU. To build this docker image, use Dockerfile provided in tensorflow-docker repository.
$ docker build -f /path/to/a/Dockerfile .