diff options
Diffstat (limited to 'cgan.py')
| -rwxr-xr-x | cgan.py | 41 |
1 files changed, 26 insertions, 15 deletions
@@ -15,7 +15,7 @@ from tqdm import tqdm import numpy as np class CGAN(): - def __init__(self, dense_layers = 3): + def __init__(self, dense_layers = 3, dropout=0.4): # Input shape self.img_rows = 28 self.img_cols = 28 @@ -24,6 +24,7 @@ class CGAN(): self.num_classes = 10 self.latent_dim = 100 self.dense_layers = dense_layers + self.dropout = dropout optimizer = Adam(0.0002, 0.5) @@ -87,10 +88,10 @@ class CGAN(): model.add(LeakyReLU(alpha=0.2)) model.add(Dense(512)) model.add(LeakyReLU(alpha=0.2)) - model.add(Dropout(0.4)) + model.add(Dropout(self.dropout)) model.add(Dense(512)) model.add(LeakyReLU(alpha=0.2)) - model.add(Dropout(0.4)) + model.add(Dropout(self.dropout)) model.add(Dense(1, activation='sigmoid')) #model.summary() @@ -107,7 +108,7 @@ class CGAN(): return Model([img, label], validity) - def train(self, epochs, batch_size=128, sample_interval=50, graph=False, smooth_real=1, smooth_fake=0): + def train(self, epochs, batch_size=128, sample_interval=50, graph=False, smooth_real=1, smooth_fake=0, gdstep=1): # Load the dataset (X_train, y_train), (_, _) = mnist.load_data() @@ -140,6 +141,7 @@ class CGAN(): gen_imgs = self.generator.predict([noise, labels]) # Train the discriminator + d_loss_real = self.discriminator.train_on_batch([imgs, labels], valid*smooth_real) d_loss_fake = self.discriminator.train_on_batch([gen_imgs, labels], valid*smooth_fake) d_loss = 0.5 * np.add(d_loss_real, d_loss_fake) @@ -151,7 +153,10 @@ class CGAN(): # Condition on labels sampled_labels = np.random.randint(0, 10, batch_size).reshape(-1, 1) # Train the generator - g_loss = self.combined.train_on_batch([noise, sampled_labels], valid) + if epoch % gdstep == 0: + g_loss = self.combined.train_on_batch([noise, sampled_labels], valid) + else: + g_loss = 0 # Plot the progress #print ("%d [D loss: %f, acc.: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100*d_loss[1], g_loss)) @@ -193,19 +198,25 @@ class CGAN(): fig.savefig("images/%d.png" % epoch) plt.close() - def generate_data(self): - noise_train = np.random.normal(0, 1, (55000, 100)) - noise_test = np.random.normal(0, 1, (10000, 100)) - noise_val = np.random.normal(0, 1, (5000, 100)) + def generate_data(self, output_train = 55000): + # with this output_train you specify how much training data you want. the other two variables produce validation + # and testing data in proportions equal to the ones of MNIST dataset + + val_size = int(output_train/11) + test_size = 2*val_size - labels_train = np.zeros(55000).reshape(-1, 1) - labels_test = np.zeros(10000).reshape(-1, 1) - labels_val = np.zeros(5000).reshape(-1, 1) + noise_train = np.random.normal(0, 1, (output_train, 100)) + noise_test = np.random.normal(0, 1, (test_size, 100)) + noise_val = np.random.normal(0, 1, (val_size, 100)) + labels_train = np.zeros(output_train).reshape(-1, 1) + labels_test = np.zeros(test_size).reshape(-1, 1) + labels_val = np.zeros(val_size).reshape(-1, 1) + for i in range(10): - labels_train[i*5500:] = i - labels_test[i*1000:] = i - labels_val[i*500:] = i + labels_train[i*int(output_train/10):-1] = i + labels_test[i*int(test_size/10):-1] = i + labels_val[i*int(val_size/10):-1] = i train_data = self.generator.predict([noise_train, labels_train]) test_data = self.generator.predict([noise_test, labels_test]) |