/DCGAN-Art-Tensorflow

Dcgan with a focus on paintings and graphic design

Primary LanguageJupyter Notebook

Setup

Tested with Python 3.6 via virtual environment:

$ python3.6 -m venv venv
$ source venv/bin/activate
$ pip install -r requirements.txt

Or use the yml to create a conda environment

Notes on using checkpoints

checkpoint directory (under checkpoint_dir) must be the same name as dataset, plus _batchsize_outputw_output_h

for instance, if dataset is insects-square-512-rgb, data_dir is ./data, checkpoint_dir is ./checkpoints, checkpoints must be in ./checkpoints/insects-square-512-rgb_6_512_512/ and the images must be in ./data/insects-square-512-rgb . but they don't need to be the same images, just same depth and dimensions

generate_test_images shows up as an error if different from training: "tensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [100,2097152] rhs shape= [10,2097152] use the last shape. 10 in the example above

python main.py --input_width=512 --input_height=512 --output_width=512 --output_height=512 --batch_size=4 --sample_num=6 --z_dim=10 --dataset=insects-grayscale-512 --checkpoint_dir=/home/art/apps/machinelearning/tensorflow/dcgan/DCGAN-Art-Tensorflow-option/checkpoint/ --data_dir=/home/art/apps/machinelearning/ --generate_test_images=10

Run on Abstract dataset

Use Abstract dataset(8145 images) as an example:

  1. download abstract.zip from the author's link https://drive.google.com/open?id=17Cm2352V9G1tR4kii5yHI_KUkevLC67_ and unzip into the ./data folder
  2. python uniform.py to convert the images to RGB and resize them
  3. if you don't want to load checkpoints, make sure to empty checkpoint folder
  4. python main.py --train --crop

the default parameter settings are specified in the following section in main.py file, which can be overwritten by providing values in the command line.

For example, the default epoch for training is 25 and dataset is in "abstract-resize" folder. To use a different value, you can do the following:

  • first, download the flower dataset and unzip to ./data/flower

  • revise the following paths in uniform.py

    # path is the folder containing your data
path = './data/flower/'

# target_path is the folder storing resized images
target_path = './data/flower-resize'

  • run python uniform.py to generate resized images in ./data/flower-resize

  • train the model with flower dataset with 5 epochs: python main.py --epoch=5 --dataset=flower-resize --train, the checkpoints will be generated and saved in a subfolder in checkpoint folder with prefix "flower-resize"

Screen Shot 2019-09-23 at 2 54 48 PM

flags = tf.app.flags
flags.DEFINE_integer("epoch", 25, "Epoch to train [25]")
flags.DEFINE_float("learning_rate", 0.0002, "Learning rate of for adam [0.0002]")
flags.DEFINE_float("beta1", 0.5, "Momentum term of adam [0.5]")
flags.DEFINE_float("train_size", np.inf, "The size of train images [np.inf]")
flags.DEFINE_integer("batch_size", 64, "The size of batch images [64]")
flags.DEFINE_integer("input_height", 256, "The size of image to use (will be center cropped). [108]")
flags.DEFINE_integer("input_width", None, "The size of image to use (will be center cropped). If None, same value as input_height [None]")
flags.DEFINE_integer("output_height", 128, "The size of the output images to produce [64]")
flags.DEFINE_integer("output_width", None, "The size of the output images to produce. If None, same value as output_height [None]")
flags.DEFINE_string("dataset", "abstract-resize", "The name of dataset [celebA, mnist, lsun]")
flags.DEFINE_string("input_fname_pattern", "*.jpg", "Glob pattern of filename of input images [*]")
flags.DEFINE_string("checkpoint_dir", "checkpoint", "Directory name to save the checkpoints [checkpoint]")
flags.DEFINE_string("data_dir", "./data", "Root directory of dataset [data]")
flags.DEFINE_string("sample_dir", "samples", "Directory name to save the image samples [samples]")
flags.DEFINE_boolean("train", True, "True for training, False for testing [False]")
flags.DEFINE_boolean("crop", True, "True for training, False for testing [False]")
flags.DEFINE_boolean("visualize", False, "True for visualizing, False for nothing [False]")
flags.DEFINE_integer("generate_test_images", 100, "Number of images to generate during test. [100]")
flags.DEFINE_integer("option", 1, "visualization option [1]")

Pay attention to parameter option. Different option can generate different kinds of visualization, this is very useful for testing.

  • option = 0: Cannot be used for MNIST. Generate a big map containing batch_size images generated randomly.
  • option = 1: Almost do the same thing as OPTION 0, but can be applied for MNIST.
  • option = 2: Generate a GIF lasting 2 seconds with batch_size frames. If failed, do the same thing as OPTION 1.
  • option = 3: Cannot be used for MNIST, and generate the same GIF as OPTION 2.
  • option = 4: Generate a big map containing z_dim GIFs. So if you want to generate a big map containing 100 GIFs on flower dataset:
python main.py --dataset flower-resize --option 4

Training on CPU

  • Training Equipment: Intel Xeon E3 4 cores
  • Training Time: About 1 hour per epoch

Parameters

  • learning rate: 0.2
  • beta1: 0.5
  • batch_size: 64
  • input size: 256 * 256
  • output size: 128 * 128

Result on Abstract.zip

  • after 1 epoch

  • after 10 epochs

  • after 20 epochs

  • Some test results after 25 epochs

Train use Google Cloud Platform (GCP)

  • make a copy of the author's GCP notebook: https://colab.research.google.com/drive/18RglimpA1JH7bRbTXtxx9fAbDl60sFVQ#scrollTo=YLBwMdxMW3PR

  • make sure runtime is GPU: Screen Shot 2019-09-23 at 9 26 46 PM Screen Shot 2019-09-23 at 9 27 03 PM

from the original author

Art and Design DCGAN in Tensorflow

Modified version of Taehoon Kim’s tensorflow implementation of DCGAN https://carpedm20.github.io/faces/ with a focus on generating paintings and soon graphic design.

It includes a script to scrape WikiArt, one to uniform images in a format that the Dcgan can work with and a Google Colab Notebook to train it on a free GPU.

Prerequisites

You can find a zip of my dataset in: https://drive.google.com/open?id=17Cm2352V9G1tR4kii5yHI_KUkevLC67_

Checkpoints here: https://drive.google.com/open?id=1yABe4LsWeDQz5p5IO2AYJPosGOgqtD2Z

Colab Notebook: https://colab.research.google.com/drive/18RglimpA1JH7bRbTXtxx9fAbDl60sFVQ#scrollTo=YLBwMdxMW3PR

You will have to convert the images to RGB and resize them with: uniform.py by changing path with your directory

Train your own dataset:

$ mkdir data/DATASET_NAME
... add images to data/DATASET_NAME ...
$ python main.py --dataset DATASET_NAME --train
$ python main.py --dataset DATASET_NAME
$ # example
$ python main.py --dataset=eyes --input_fname_pattern="*_cropped.png" --train

If your dataset is located in a different root directory:

$ python main.py --dataset DATASET_NAME --data_dir DATASET_ROOT_DIR --train
$ python main.py --dataset DATASET_NAME --data_dir DATASET_ROOT_DIR
$ # example
$ python main.py --dataset=eyes --data_dir ../datasets/ --input_fname_pattern="*_cropped.png" --train