vector-to-image using condition-gan【NUS_GAN】

Overview

This is a tensorflow implementation of generating face images using condition-GAN named NUS_GAN. We've gotten great progress on the basis of DCGAN after using our algorithm. At present, this is just a preliminary version and coupled with our lack of ability, maybe there are still some problems in the codes. Please fell free to correct us.

The following is the model architecture.

Requirements

• python3.6+
• tensorflow
• numpy
• scipy
• matplotlib
• others

Datasets

• The model is currently trained on the celeb-dataset. Considering the lack of computing power, we only selected 10,000 pictures for training, which number from 000001.jpg to 005000.jpg and 010001.jpg to 015000.jpg. Download the images from here and save them in data. Our trainging dataset includes the label.txt.

• Make empty directories, samples, checkpoint. They will be used for sampling the generated images and saving the trained models.

• You can also add your own dataset and transfer them to train the model, which is easy to modify.

Usage

• Training

  • Basic usage python3 train.py --train
  • Options
    • epoch: Epoch to train[300]
    • learning_rate: Learning rate of for adam [0.0002]
    • beta1: Momentum term of adam [0.5]
    • train_size: The size of train images [np.inf]
    • batch_size: 50, "The size of batch images [50]
    • input_height: The size of image to use (will be center cropped). [218]
    • input_width: The size of image to use (will be center cropped). If None, same value as input_height [178]
    • output_height: The size of the output images to produce [218]
    • output_width: The size of the output images to produce. If None, same value as output_height [178]
    • dataset: The name of dataset [celeb]
    • input_fname_pattern: Glob pattern of filename of input images [*.jpg]
    • checkpoint_dir: Directory name to save the checkpoints [checkpoint]
    • data_dir: Root directory of dataset [./data]
    • test_dir: Root directory of test dataset [./test]
    • test_dataset: The name of test_dataset[test_in]
    • test_out: Directory name of test outputs[test_out]
    • sample_dir: Directory name to save the image samples [samples]
    • train: True for training, False for testing [False]
    • crop: True for training, False for testing [True]
    • visualize: True for visualizing, False for nothing [False]
    • generate_test_images: Number of images to generate during test. [100]

• Generating Images from Vectors

  • Write the vectors in text file, and save it as test/test_in/test_label.txt. You can also download our test-dataset from here. Generate the Images for the vectors using:

  python3 main.py
  

  vectors specifies the number of images to be generated per caption. The generated images will be saved in test/test_out/.

  On our dataset celeb, the length of vectors is 40, each bit means yes or no from left to right. The concrete meaning from left to right list as follows:

5_o_Clock_Shadow Arched_Eyebrows Attractive Bags_Under_Eyes Bald Bangs Big_Lips Big_Nose Black_Hair Blond_Hair Blurry Brown_Hair Bushy_Eyebrows Chubby Double_Chin Eyeglasses Goatee Gray_Hair Heavy_Makeup High_Cheekbones Male Mouth_Slightly_Open Mustache Narrow_Eyes No_Beard Oval_Face Pale_Skin Pointy_Nose Receding_Hairline Rosy_Cheeks Sideburns Smiling Straight_Hair Wavy_Hair Wearing_Earrings Wearing_Hat Wearing_Lipstick Wearing_Necklace Wearing_Necktie Young

which 1 means yes and 0 means no. You can get more details about label from celeb-dataset

Training process

Pre-trained Models

• Download the pretrained model from here and save it in checkpoint/celeb_50_218_178/. You can choose one trained model to generate your images from checkpoint/celeb_50_218_178/ by modifying file checkpoint.

• Our pre-trained Models are trained with dataset: celeb, batch_size: 50, output_height: 218 and output_width: 178. You can use this model for generating the images as long as you change these parameters as above.

Sample Images Generated

Following are the images generated by the generative model from the vectors.

• Inputs: test_label.txt can be downloaded from here.

• Outputs:

From the test results, we can see that our model probably restores the corresponding features, but there are still details to be perfected. If we have enough computing power, we believe that we can generate enough realistic and qualified faces.

Implementation Details

References

• Text to image synthesis using thought vectors Code
• A tensorflow implementation of "Deep Convolutional Generative Adversarial Networks" Code
• celeb-dataset Dataset