TF-slim-flowers

Fine-tunes an InceptionV3 model on the Flowers training set. InceptionV3 paper)

Contents

1. Technology
2. Structure
3. Running
4. Demo

Technology

1. Docker
2. Data parallel read
3. TF-Slim
4. tensorboard
5. TF-distribute

Structure

.
├── docker-compose.yml
├── Dockerfile // docker build file
├── log 
│   ├── train_dir
│   └── eval_dir
├── nvidia-docker-compose.yml
├── README.md
├── src
│   ├── datasets // generate tfrecord and slim Dataset object 
│   │   ├── __init__.py
│   │   ├── convert_flowers.py //  generate tfrecord
│   │   ├── dataset_utils.py 
│   │   └── flowers.py
│   ├── models // pre-trained model 
│   │   └── inception_v3.ckpt
│   ├── nets // network architecture
│   │   ├── __init__.py 
│   │   ├── inception_utils.py
│   │   ├── inception_v3.py
│   │   ├── inception_v3_test.py
│   │   ├── lenet.py
│   │   └── nets_factory.py
│   ├── preprocessing // data preprecessing
│   │   ├── __init__.py 
│   │   ├── inception_preprocessing.py
│   │   ├── lenet_preprocessing.py
│   │   └── preprocessing_factory.py
│   ├── scriptes
│   │   ├── download_uncompress_flows.sh
│   │   ├── eval_inception_v3_on_flowers.sh
│   │   ├── finetune_inception_v3_on_flowers.sh
│   │   └── unittest.sh
│   ├── common_flags.py // common flags for train/val
│   ├── __init__.py
│   ├── data_provider.py // slim DataProvider object 
│   ├── data_provider_test.py
│   ├── eval.py
│   ├── model.py
│   ├── train.py
│   └── utils.py
└──── screenshots

Running

Configure environment

1. install Docker-ce
2. install nvidia-docker(not nvidia-docker2)
3. install docker-compose
4. install nvidia-docker-compse
5. test everything Ok

nvidia-docker run --rm nvidia/cuda nvidia-smi

Entering enviroment

cd ${ROOT} // suppose ROOT is the root of the project
1. nvidia-docker-compose run --rm --service-ports tf-slim-demo bash
2. nvidia-smi // test OK

run scripts

cd src
1. bash ./scripts/unittest.sh
2. bash ./scripts/download_uncompress_flowers.sh

run scripts

# create dataset
1. python ./datasets/convert_flowers.py

Suppose you have two GPUs(0&1), if not, you must modify training and evaluation script to prevent them running parallelly.

run training/evaluation

# GPU 0 for training
cd ${ROOT}
bash ./src/scripts/finetune_inception_v3_on_flowers.sh
# GPU 1 for evaluation
bash ./src/scripts/eval_inception_v3_on_flowers.sh

show tensorboard

    attention: docker have not its own graphic X server, so it must connect to
    host X-server, there are various way to do this, but I recommend you 
    do this using X11 Socket, see docker-compose.yml for detail
    
    All ready, please open a new Terminal to entery the same container 
    which is running train/evaluation
    
    1. docker exec -it -e DISPLAY=$DISPLAY ${CONTAINER_DIR} bash
    2. tensorboard --logdir=train:./log/train_dir,eval:./log/eval_dir
    3. go to chrome : http://localhost:6006/
    You can get everything you want!

Demo

screenshots

1. tfrecord
2. train
3. eval
4. tensorboard for summary
5. tensorboard for graph
6. tensorboard for image
7. tensorboard for TotalLoss
8. tensorboard for eval_accuracy