使用Keras編寫GAN的入門
Time: 2017-5-31
前言
主要參考了網頁[1]的教程，同時主要算法來自Ian J. Goodfellow 的論文，算法如下：
gan
代碼
%matplotlib inline
import numpy as np
import pandas as pd
from keras.models import Model
from keras.layers import Dense, Activation, Input, Reshape
from keras.layers import Conv1D, Flatten, Dropout
from keras.optimizers import SGD, Adam
from tqdm import tqdm_notebook as tqdm # 進度條
# 生成隨機正弦曲線的數據
def sample_data(n_samples=10000, x_vals=np.arange(0, 5, .1), max_offset=1000, mul_range=[1, 2]):
vectors = []
for i in range(n_samples):
offset = np.random.random() * max_offset
mul = mul_range[0] + np.random.random() * (mul_range[1] - mul_range[0])
vectors.append(np.sin(offset + x_vals * mul) / 2 + .5)
return np.array(vectors)
# 創建生成模型
def get_generative(G_in, dense_dim=200, out_dim=50, lr=1e-3):
x = Dense(dense_dim)(G_in)
x = Activation('tanh')(x)
G_out = Dense(out_dim, activation='tanh')(x)
G = Model(G_in, G_out)
opt = SGD(lr=lr)
G.compile(loss='binary_crossentropy', optimizer=opt)
return G, G_out
# 創建判別模型
def get_discriminative(D_in, lr=1e-3, drate = .25, n_channels=50, conv_sz=5, leak=.2):
x = Reshape((-1, 1))(D_in)
x = Conv1D(n_channels, conv_sz, activation='relu')(x)
x = Dropout(drate)(x)
x = Flatten()(x)
x = Dense(n_channels)(x)
D_out = Dense(2, activation='sigmoid')(x)
D = Model(D_in, D_out)
dopt = Adam(lr=lr)
D.compile(loss='binary_crossentropy', optimizer=dopt)
return D, D_out
def set_trainability(model, trainable=False):
model.trainable = trainable
for layer in model.layers:
layer.trainable = trainable
def make_gan(GAN_in, G, D):
set_trainability(D, False)
x = G(GAN_in)
GAN_out = D(x)
GAN = Model(GAN_in, GAN_out)
GAN.compile(loss='binary_crossentropy', optimizer=G.optimizer)
return GAN, GAN_out
# 通過生成數據 預訓練判別模型
def sample_data_and_gen(G, noise_dim=10, n_samples=10000):
XT = sample_data(n_samples=n_samples)
XN_noise = np.random.uniform(0, 1, size=[n_samples, noise_dim])
XN = G.predict(XN_noise)
X = np.concatenate((XT, XN))
y = np.zeros((2*n_samples, 2))
y[:n_samples, 1] = 1
y[n_samples:, 0] = 1
return X, y
def pretrain(G, D, noise_dim=10, n_samples=10000, batch_size=32):
X, y = sample_data_and_gen(G, noise_dim=noise_dim, n_samples=n_samples)
set_trainability(D, True)
D.fit(X, y, epochs=1, batch_size=batch_size)
# 開始交叉訓練步驟
def sample_noise(G, noise_dim=10, n_samples=10000):
X = np.random.uniform(0, 1, size=[n_samples, noise_dim])
y = np.zeros((n_samples, 2))
y[:, 1] = 1
return X, y
def train(GAN, G, D, epochs=500, n_samples=10000, noise_dim=10, batch_size=32, verbose=False, v_freq=50):
d_loss = []
g_loss = []
e_range = range(epochs)
if verbose:
e_range = tqdm(e_range)
for epoch in e_range:
X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim) # 對D進行訓練
set_trainability(D, True)
d_loss.append(D.train_on_batch(X, y))
X, y = sample_noise(G, n_samples=n_samples, noise_dim=noise_dim) # 對G訓練
set_trainability(D, False)
g_loss.append(GAN.train_on_batch(X, y))
if verbose and (epoch + 1) % v_freq == 0:
print("Epoch #{}: Generative Loss: {}, Discriminative Loss: {}".format(epoch + 1, g_loss[-1], d_loss[-1]))
return d_loss, g_loss
ax = pd.DataFrame(np.transpose(sample_data(5))).plot()
G_in = Input(shape=[10])
G, G_out = get_generative(G_in)
G.summary()
D_in = Input(shape=[50])
D, D_out = get_discriminative(D_in)
D.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_9 (InputLayer) (None, 10) 0
_________________________________________________________________
dense_13 (Dense) (None, 200) 2200
_________________________________________________________________
activation_4 (Activation) (None, 200) 0
_________________________________________________________________
dense_14 (Dense) (None, 50) 10050
=================================================================
Total params: 12,250
Trainable params: 12,250
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_10 (InputLayer) (None, 50) 0
_________________________________________________________________
reshape_4 (Reshape) (None, 50, 1) 0
_________________________________________________________________
conv1d_4 (Conv1D) (None, 46, 50) 300
_________________________________________________________________
dropout_4 (Dropout) (None, 46, 50) 0
_________________________________________________________________
flatten_4 (Flatten) (None, 2300) 0
_________________________________________________________________
dense_15 (Dense) (None, 50) 115050
_________________________________________________________________
dense_16 (Dense) (None, 2) 102
=================================================================
Total params: 115,452
Trainable params: 115,452
Non-trainable params: 0
_________________________________________________________________
png
GAN_in = Input([10])
GAN, GAN_out = make_gan(GAN_in, G, D)
GAN.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_11 (InputLayer) (None, 10) 0
_________________________________________________________________
model_9 (Model) (None, 50) 12250
_________________________________________________________________
model_10 (Model) (None, 2) 115452
=================================================================
Total params: 127,702
Trainable params: 12,250
Non-trainable params: 115,452
_________________________________________________________________
pretrain(G, D)
Epoch 1/1
20000/20000 [==============================] - 3s - loss: 0.0072
d_loss, g_loss = train(GAN, G, D, verbose=True)
Epoch #50: Generative Loss: 4.41527795791626, Discriminative Loss: 0.6733301877975464
Epoch #100: Generative Loss: 3.8898046016693115, Discriminative Loss: 0.09901376813650131
Epoch #150: Generative Loss: 6.2410054206848145, Discriminative Loss: 0.034074194729328156
Epoch #200: Generative Loss: 5.206066608428955, Discriminative Loss: 0.13078376650810242
Epoch #250: Generative Loss: 3.5144925117492676, Discriminative Loss: 0.07160962373018265
Epoch #300: Generative Loss: 3.705162525177002, Discriminative Loss: 0.05893774330615997
Epoch #350: Generative Loss: 3.511479616165161, Discriminative Loss: 0.09775738418102264
Epoch #400: Generative Loss: 4.141300678253174, Discriminative Loss: 0.03169865906238556
Epoch #450: Generative Loss: 3.500260829925537, Discriminative Loss: 0.05957922339439392
Epoch #500: Generative Loss: 2.9797921180725098, Discriminative Loss: 0.10566817969083786
ax = pd.DataFrame(
{
'Generative Loss': g_loss,
'Discriminative Loss': d_loss,
}
).plot(title='Training loss', logy=True)
ax.set_xlabel("Epochs")
ax.set_ylabel("Loss")
png
N_VIEWED_SAMPLES = 2
data_and_gen, _ = sample_data_and_gen(G, n_samples=N_VIEWED_SAMPLES)
pd.DataFrame(np.transpose(data_and_gen[N_VIEWED_SAMPLES:])).plot()
png
N_VIEWED_SAMPLES = 2
data_and_gen, _ = sample_data_and_gen(G, n_samples=N_VIEWED_SAMPLES)
pd.DataFrame(np.transpose(data_and_gen[N_VIEWED_SAMPLES:])).rolling(5).mean()[5:].plot()
png