nbatch = 128
x, y, stream = get_mnist(nbatch)
g = InfoGenerator(g_size=(1, 28, 28),
g_nb_filters=64,
g_nb_noise=100, # not coding but noise
g_scales=2,
g_FC=[1024],
g_info=[
CategoryDist(n=10, lmbd=1e-3),
# UniformDist(min=-1.0,max=+1.0,lmbd=1e-3,stddev_fix=True),
# UniformDist(min=-1.0,max=+1.0,lmbd=1e-3,stddev_fix=True),
# UniformDist(min=-1.0,max=+1.0,lmbd=1e-3,stddev_fix=True),
],
g_init=InitNormal(scale=0.02),
)
d = Discriminator(d_size=g.g_size, d_nb_filters=64, d_scales=2, d_FC=[1024], d_init=InitNormal(scale=0.02))
Q = Sequential([ Dense(200, batch_input_shape=d.layers[-2].output_shape) ,
BN(),
Activation('relu'),
Dense(g.g_info.Qdim),
])
gan = InfoGAN(generator=g, discriminator=d, Qdist=Q)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(stream, save_dir='./samples/mnist_info', k=2,
nmax=nbatch*100,
nbatch=nbatch,
# opt=RMSprop(lr=0.0005))
opt=Adam(lr=0.0001))
from GAN.utils.data import transform_skeleton, inverse_transform_skeleton
from GAN.utils.init import InitNormal
from keras.optimizers import Adam, SGD, RMSprop
if __name__ == '__main__':
nbatch = 128
nmax = nbatch * 100
npxw, npxh = 64, 128
from load import people, load_all
va_data, tr_stream, _ = people(pathfile='protocol/PPPS.txt', size=(npxw, npxh), batch_size=nbatch)
g = Generator(g_size=(8, npxh, npxw), g_nb_filters=128, g_nb_coding=500, g_scales=4, g_init=InitNormal(scale=0.002))#, g_FC=[5000])
d = Discriminator(d_size=g.g_size, d_nb_filters=128, d_scales=4, d_init=InitNormal(scale=0.002))#, d_FC=[5000])
gan = GAN(g, d)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(tr_stream,
save_dir='./samples/parsing_skeleton/',
k=1,
nbatch=nbatch,
nmax=nmax,
opt=Adam(lr=0.0002, beta_1=0.5, decay=1e-5),
transform=transform_skeleton, #opt=RMSprop(lr=0.01))
inverse_transform=inverse_transform_skeleton)
import numpy as np
import pandas as pd
from sklearn.neighbors import KNeighborsClassifier
from facenet_pytorch import InceptionResnetV1
from GAN.models import Discriminator
resnet = InceptionResnetV1(pretrained='cassia_webface')
gallery = pd.read_csv('path/to/gallery')
disc = Discriminator()
knn_clf = KNeighborsClassifier(weights='distance', metric='cosine')
knn_clf.fit(gallery.drop('names', axis=1), gallery['names'])
def get_prediction(image_bytes):
tensor = prepare_image(image_bytes=image_bytes)
embedding = resnet(tensor).detach().cpu().numpy()
name = knn_clf.predict(embedding)
score = disc(tensor).item()
return name, score
if __name__ == '__main__':
# init with ae and then run gan
nbatch = 128
x, y, stream = get_mnist(nbatch)
g = Generator(g_size=(1, 28, 28),
g_nb_filters=64,
g_nb_coding=200,
g_scales=2,
g_FC=[1024],
g_init=InitNormal(scale=0.05))
d = Discriminator(d_size=g.g_size,
d_nb_filters=64,
d_scales=2,
d_FC=[1024],
d_init=InitNormal(scale=0.05))
g.load_weights('models/mnist_ae_g.h5')
print g.get_weights()[0].sum()
d.load_weights('models/mnist_ae_d.h5')
print d.get_weights()[0].sum()
# gan = GAN(g, d)
gan = AEGAN(g, d)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(stream,
save_dir='./samples/mnist',
k=1,
nbatch=nbatch,
va_data, tr_stream, _ = people(pathfile='protocol/cuhk01-train.txt', size=(npxw, npxh), batch_size=nbatch)
g = InfoGenerator(g_size=(3, npxh, npxw),
g_nb_filters=128,
g_nb_noise=500, # not coding but noise
g_scales=3,
g_FC=None,
g_info=[
# CategoryDist(n=10, lmbd=1e-3),
UniformDist(min=-1, max=+1, lmbd=1e-3),
],
g_init=InitNormal(scale=0.002)
)
d = Discriminator(d_size=g.g_size,
d_nb_filters=128,
d_scales=3,
d_FC=None,
d_init=InitNormal(scale=0.002))
Q = Sequential([ Dense(500, batch_input_shape=d.layers[-2].output_shape) ,
BN(),
Activation('relu'),
Dense(g.g_info.Qdim),
])
gan = InfoGAN(generator=g, discriminator=d, Qdist=Q)
from keras.optimizers import Adam, SGD, RMSprop
gan.fit(tr_stream,
save_dir='./samples/cuhk01_info_small/',
k=1,
nbatch=nbatch,