A tensorflow project template

This is a Tensorflow project template based on VGG16 model.

Config

All parameters are in config file, please see project_root/configs/config.yaml.

Data

1. Download data

As example, here we use the Kaggle data: cats vs. dogs, you can download data here: cats vs. dogs

After downloading, put the train.zip into folder data, and extract it, rename it to cat_dog.

2. Data preprocessing

• Resize image

Note images in cat_dog have different sizes, so first we should resize the images. Save the resized images in project_root/data/cat_dog_resize.

• Compute mean

Compute the images mean, and write it to config file.

To do above operations, please go to project_root/data_loader/data_reader.py .

3. Generate data batch

Generally, you can store data in three ways.

1. store data in cpu/gpu memory.
2. store data using tfrecord.
3. store data in disk.

The io speed is 1>2>3. But cpu/gpu memory size is often limited, in this situation, way 2 or 3 may be a better choice.

Here, for convenience, we only store data in way 1.

Note, when training, data batch is generated continuously. We implement this in /data_loader/data_generator.py

We define the Parent class :

class DataGenerator(object):
    def __init__(self, config=None):
        self.config = config

    def next_batch(self):
        raise NotImplementedError

you can implement anyone of the 3 child classes Inherited from this parent, depending on your choice.

# store data in cpu memory.
class MemoryDataGenerator(DataGenerator):
    pass

# store all data in disk, and read batch data from disk to cpu memory.
class DiskDataGenerator(DataGenerator):
    pass

# store data in TFrecord file, and read batch data from TFRecord file.
class TFRecordDataGenerator(DataGenerator):
    pass

For more details, please go to have a look.

Models

All models are defined in folder project_root/models.

Note: when you define your model, you must define the model loss, accuray, reference.

Trainer

1. Download pretrained model.

you can download vgg16 pretrained model here:VGG16 trained on ImageNet。And put it in the project_root/data/pretrained_model

2. Trainer

The solver is defined to train the model in folder project_root/trainers

Mains

Train and test are defined in project_root/mains.

For either training or testing, you should modify the config file and command-line arguments correspondingly.

Enjoy your training time!

Tensorboard

Tensorboard is a powerful tool for debugging tensorflow program, please don't ignore it.

class Summary defined in project_root/utils/logger.py is used to get all kinds of visualization.

For more details, please go to have a look.

TODO list

• Data process

• Data loader

• VGG16 model

• Train

• Test

• validation

• Use factory pattern to call specified model.