Pix2Pix in Tensorflow

Study Friendly Implementation of Pix2Pix in Tensorflow

[Paper | Post(in Korean) | Pytorch Version]

• GAN: [Pytorch][Tensorflow]
• DCGAN: [Pytorch][Tensorflow]
• InfoGAN: [Pytorch][Tensorflow]
• Pix2Pix: [Pytorch][Tensorflow]
• DiscoGAN: [Pytorch][Tensorflow]

1. Environments

• Windows 10
• Python 3.5.3 (Anaconda)
• Tensorflow 1.4.0
• Numpy 1.13.1

2. Code Description

• train.py: Main Code
• test.py: Test Code after Training
• model.py: Generator and Discriminator
• dbread.py: My Own Code for Reading Database

3. Networks and Parameters

3.1 Hyper-Parameters

• Image Size = 256x256 (Resized)
• Batch Size = 1 or 4
• Learning Rate = 0.0002
• Adam_beta1 = 0.5
• Lambda_A = 100 (Weight of L1-Loss)

Detail Recommandations for Each Dataset are on the Last Page of Original Paper

3.2 Generator Networks (network.py)

3.3 Discriminator Networks (network.py)

4. Database

4.1 DB Download

download-dataset.py is from this github link

• facades: python database/download-dataset.py facades
• cityscapes: python database/download-dataset.py cityscapes
• maps: python database/download-dataset.py maps
• edges2shoes: python database/download-dataset.py edges2shoes
• edges2handbags: python database/download-dataset.py edges2handbags

4.2 DB Setting

• You just have to get a text file that contains all directories of your image data(filelist.txt).
• You can make filelist with following command.

cd <database_dir>
dir /b /s > filelist.txt

• Then you must open 'filelist.txt' and delete line 'your_db_dir/filelist.txt'.

5. Train

python train.py --train filelist.txt

• --train: Directory of the Text File of Train Filelist
• --out_dir: Directory to Save your Train Result
• --epochs: Num of Epochs You Want
• --batch_size: Batch Size You Want
• --direction: 'AtoB' or 'BtoA'

After finish training, saved models are in the ./output/checkpoint directory and the Train Results are saved in ./output/result.(default)

6. Test

python test.py --train filelist.txt

• --test: Directory of the Text File of Test Filelist
• --out_dir: Directory to Save your Train Result
• --ckpt_dir: Directory of Trained Model
• --visnum: Number of Visualization in an Image File
• --direction: 'AtoB' or 'BtoA'

Test results will be saved in ./output_test(default)

7. Results

[Input | Generated | Ground Truth]

Edges to Shoes (8 Epochs)

15 Epochs (which is in the Paper) will give better results

Maps to Aerials (200 Epochs)

Architectural labels to Photo (200 Epochs)