We train generative-competitive network for drawing pictures on Azure ML

Deep learning sometimes looks like pure magic, especially when the computer learns to do something really creative, for example, to paint pictures! The technology used for this is called GAN, a competitive generative network, and in this article we will look at how such networks are organized and how to train them to generate pictures using Azure Machine Learning.

This post is part of the AI April initiative . Every April day, my Microsoft colleagues write interesting articles on AI and machine learning. Look at the calendar - suddenly you will find there other topics of interest to you. Articles are mainly in English.

Azure ML VS Code ), , Azure ML. , "" MNIST. Azure ML — , , :

, 2019, Art of the Artificial keragan WikiArt:	, 2019, keragan WikiArt:

WikiArt. , , , WikiArt Retriever, WikiArt Dataset GANGogh.

, dataset. , , :

, , ( , ), (, ) .

-

- (GAN).

GAN keragan Keras, .

GAN :

• ,
• , "" ( )

GAN :

1. :
   

   noise = np.random.normal(0, 1, (batch_size, latent_dim))
   gen_imgs = generator.predict(noise)
   imgs = get_batch(batch_size)

2. . ones zeros :
   

   d_loss_r = discriminator.train_on_batch(imgs, ones)
   d_loss_f = discriminator.train_on_batch(gen_imgs, zeros)
   d_loss = np.add(d_loss_r , d_loss_f)*0.5

3. , , :
   

   g_loss = combined.train_on_batch(noise, ones)

, — :

discriminator = create_discriminator()
generator = create_generator()
discriminator.compile(loss='binary_crossentropy',optimizer=optimizer, 
                      metrics=['accuracy'])
discriminator.trainable = False
z = keras.models.Input(shape=(latent_dim,))
img = generator(z)
valid = discriminator(img)
combined = keras.models.Model(z, valid)
combined.compile(loss='binary_crossentropy', optimizer=optimizer)

, (CNN). 64x64 :

discriminator = Sequential()

for x in [16,32,64]: # number of filters on next layer
    discriminator.add(Conv2D(x, (3,3), strides=1, padding="same"))
    discriminator.add(AveragePooling2D())
    discriminator.addBatchNormalization(momentum=0.8))
    discriminator.add(LeakyReLU(alpha=0.2))
    discriminator.add(Dropout(0.3))

discriminator.add(Flatten())
discriminator.add(Dense(1, activation='sigmoid'))

3 :

• 64x64x3 16- , ( AveragePooling2D ) 32x32x16.
• 32x32x16 16x16x32
• 8x8x64.

, ( — Dense ).

. , , — latent_dim=100. , , 100..

— 100 . . UpSampling2D , :

generator = Sequential()
generator.add(Dense(8 * 8 * 2 * size, activation="relu", 
                                      input_dim=latent_dim))
generator.add(Reshape((8, 8, 2 * size)))

for x in [64;32;16]:
    generator.add(UpSampling2D())
    generator.add(Conv2D(x, kernel_size=(3,3),strides=1,padding="same"))
    generator.add(BatchNormalization(momentum=0.8))
    generator.add(Activation("relu"))

generator.add(Conv2D(3, kernel_size=3, padding="same"))
generator.add(Activation("tanh"))

64x64x3, . tanh [-1;1] — , . , , ImageDataset, .

Azure ML

- , Azure ML !

, , . Azure ML , (accuracy) (loss). run.log, , Azure ML.

, , ( ) . , , .. - , .

, Azure ML , . log_image, numpy-, , matplotlib. , , . callbk, keragan :

def callbk(tr):
    if tr.gan.epoch % 20 == 0:
        res = tr.gan.sample_images(n=3)
        fig,ax = plt.subplots(1,len(res))
        for i,v in enumerate(res):
            ax[i].imshow(v[0])
        run.log_image("Sample",plot=plt)

:

gan = keragan.DCGAN(args)
imsrc = keragan.ImageDataset(args)
imsrc.load()
train = keragan.GANTrainer(image_dataset=imsrc,gan=gan,args=args)

train.train(callbk)

, keragan , args, , , , learning rate ..

Azure ML VS Code, , SDK, Azure ML. submit_gan.ipynb, :

• : ws = Workspace.from_config()
• : cluster = ComputeTarget(workspace=ws, name='My Cluster'). GPU, [NC6][AzureVMNC].
• : ds.upload(...).

, , :

exp = Experiment(workspace=ws, name='KeraGAN')
script_params = {
    '--path': ws.get_default_datastore(),
    '--dataset' : 'faces',
    '--model_path' : './outputs/models',
    '--samples_path' : './outputs/samples',
    '--batch_size' : 32,
    '--size' : 512,
    '--learning_rate': 0.0001,
    '--epochs' : 10000
}
est = TensorFlow(source_directory='.',
    script_params=script_params,
    compute_target=cluster,
    entry_script='train_gan.py',
    use_gpu = True,
    conda_packages=['keras','tensorflow','opencv','tqdm','matplotlib'],
    pip_packages=['git+https://github.com/shwars/keragan@v0.0.1']

run = exp.submit(est)

model_path=./outputs/models samples_path=./outputs/samples, ( ) outputs. , , Azure ML.

Estimator, GPU, Tensorflow. Estimator, " " . Estimator- .

— , keragan GitHub. PyPI pip-, , GitHub , commit ID. , PyPI.

, Azure ML Portal:

GAN , . -, learning rate: , — . .

:

• --size , . (64 128) , ( 1024) . 1024 , , progressive growing
• --learning_rate . , .
• --dateset. , , Azure ML datastore, .

, for, . , . .

, , . outputs/models, — outputs/samples. Azure ML Portal :

, , . run, , , ( ):

run.download_files(prefix='outputs/samples')

outputs/samples, .

run ( , ), , run id, :

run = Run(experiment=exp,run_id='KeraGAN_1584082108_356cf603')

. , , . , ( gen_):

fnames = run.get_file_names()
fnames = filter(lambda x : x.startswith('outputs/models/gen_'),fnames)

: outputs/models/gen_0.h5, outputs/models/gen_100.h5 .. :

no = max(map(lambda x: int(x[19:x.find('.')]), fnames))
fname = 'outputs/models/gen_{}.h5'.format(no)
fname_wout_path = fname[fname.rfind('/')+1:]
run.download_file(fname)

, fname_wout_path.

, Keras, , , , :

model = keras.models.load_model(fname_wout_path)
latent_dim=model.layers[0].input.shape[1].value
res = model.predict(np.random.normal(0,1,(10,latent_dim)))

, [-1,1], [0,1], matplotlib:

res = (res+1.0)/2
fig,ax = plt.subplots(1,10,figsize=(15,10))
for i in range(10):
    ax[i].imshow(res[i])

:

, :

Colourful Spring, 2020	Countryside, 2020
Through the Icy Glass, 2020	Summer Landscape, 2020

( ) , — @art_of_artificial, .

, - . , , . , :

- , Azure. , . , — — . ? ...

keragan, , DCGAN, Maxime Ellerbach, GANGogh. GAN Keras .

Azure ML

• Azure ML
• Azure ML
• - Azure ML ( )