swap back 0,1 labels of GAN (#107)

.travis.yml

@@ -16,18 +16,18 @@ matrix:
   include:
   - os: linux
     python: 2.7
-    env: TF_VERSION=1.0.0rc1 TF_TYPE=release
+    env: TF_VERSION=1.0.0rc2 TF_TYPE=release
   - os: linux
     python: 3.5
-    env: TF_VERSION=1.0.0rc1 TF_TYPE=release
+    env: TF_VERSION=1.0.0rc2 TF_TYPE=release
   - os: linux
     python: 2.7
-    env: TF_VERSION=1.0.0rc1 TF_TYPE=nightly
+    env: TF_VERSION=1.0.0rc2 TF_TYPE=nightly
   - os: linux
     python: 3.5
-    env: TF_VERSION=1.0.0rc1 TF_TYPE=nightly
+    env: TF_VERSION=1.0.0rc2 TF_TYPE=nightly
   allow_failures:
-  - env: TF_VERSION=1.0.0rc1 TF_TYPE=nightly
+  - env: TF_TYPE=nightly
 
 install:
 - pip install -U pip    # the pip version on travis is too old

examples/GAN/DCGAN-CelebA.py

@@ -22,6 +22,7 @@ from GAN import GANTrainer, RandomZData, GANModelDesc
 DCGAN on CelebA dataset.
 
 1. Download the 'aligned&cropped' version of CelebA dataset
+   from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
    (or just use any directory of jpg files).
 2. Start training:
     ./DCGAN-CelebA.py --data /path/to/image_align_celeba/
@@ -39,7 +40,7 @@ class Model(GANModelDesc):
         return [InputDesc(tf.float32, (None, SHAPE, SHAPE, 3), 'input')]
 
     def generator(self, z):
-        """ return a image generated from z"""
+        """ return an image generated from z"""
         nf = 64
         l = FullyConnected('fc0', z, nf * 8 * 4 * 4, nl=tf.identity)
         l = tf.reshape(l, [-1, 4, 4, nf * 8])
@@ -79,7 +80,7 @@ class Model(GANModelDesc):
                       W_init=tf.truncated_normal_initializer(stddev=0.02)):
             with tf.variable_scope('gen'):
                 image_gen = self.generator(z)
-            tf.summary.image('gen', image_gen, max_outputs=30)
+            tf.summary.image('generated-samples', image_gen, max_outputs=30)
             with tf.variable_scope('discrim'):
                 vecpos = self.discriminator(image_pos)
             with tf.variable_scope('discrim', reuse=True):
@@ -106,10 +107,8 @@ def get_data():
 
 
 def get_config():
-    logger.auto_set_dir()
-    dataset = get_data()
     return TrainConfig(
-        dataflow=dataset,
+        dataflow=get_data(),
         callbacks=[ModelSaver()],
         session_config=get_default_sess_config(0.5),
         model=Model(),
@@ -145,6 +144,7 @@ if __name__ == '__main__':
         sample(args.load)
     else:
         assert args.data
+        logger.auto_set_dir()
         config = get_config()
         if args.load:
             config.session_init = SaverRestore(args.load)

examples/GAN/GAN.py

@@ -31,14 +31,9 @@ class GANModelDesc(ModelDesc):
 
           min_G max _D V(D, G) = IE_{x ~ p_data} [log D(x)] + IE_{z ~ p_fake} [log (1 - D(G(z)))]
 
-        Note, we swap 0, 1 labels as suggested in "Improving GANs".
-
         Args:
             logits_real (tf.Tensor): discrim logits from real samples
             logits_fake (tf.Tensor): discrim logits from fake samples produced by generator
-
-        Returns:
-            tf.Tensor: Description
         """
         with tf.name_scope("GAN_loss"):
             score_real = tf.sigmoid(logits_real)
@@ -48,20 +43,20 @@ class GANModelDesc(ModelDesc):
 
             with tf.name_scope("discrim"):
                 d_loss_pos = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
-                    logits=logits_real, labels=tf.zeros_like(logits_real)), name='loss_real')
+                    logits=logits_real, labels=tf.ones_like(logits_real)), name='loss_real')
                 d_loss_neg = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
-                    logits=logits_fake, labels=tf.ones_like(logits_fake)), name='loss_fake')
+                    logits=logits_fake, labels=tf.zeros_like(logits_fake)), name='loss_fake')
 
-                d_pos_acc = tf.reduce_mean(tf.cast(score_real < 0.5, tf.float32), name='accuracy_real')
-                d_neg_acc = tf.reduce_mean(tf.cast(score_fake > 0.5, tf.float32), name='accuracy_fake')
+                d_pos_acc = tf.reduce_mean(tf.cast(score_real > 0.5, tf.float32), name='accuracy_real')
+                d_neg_acc = tf.reduce_mean(tf.cast(score_fake < 0.5, tf.float32), name='accuracy_fake')
 
                 self.d_accuracy = tf.add(.5 * d_pos_acc, .5 * d_neg_acc, name='accuracy')
                 self.d_loss = tf.add(.5 * d_loss_pos, .5 * d_loss_neg, name='loss')
 
             with tf.name_scope("gen"):
                 self.g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
-                    logits=logits_fake, labels=tf.zeros_like(logits_fake)), name='loss')
-                self.g_accuracy = tf.reduce_mean(tf.cast(score_fake < 0.5, tf.float32), name='accuracy')
+                    logits=logits_fake, labels=tf.ones_like(logits_fake)), name='loss')
+                self.g_accuracy = tf.reduce_mean(tf.cast(score_fake > 0.5, tf.float32), name='accuracy')
 
             add_moving_summary(self.g_loss, self.d_loss, self.d_accuracy, self.g_accuracy)
 
@@ -76,6 +71,7 @@ class GANTrainer(FeedfreeTrainerBase):
         self.build_train_tower()
         opt = self.model.get_optimizer()
 
+        # by default, run one d_min after one g_min
         self.g_min = opt.minimize(self.model.g_loss, var_list=self.model.g_vars, name='g_op')
         with tf.control_dependencies([self.g_min]):
             self.d_min = opt.minimize(self.model.d_loss, var_list=self.model.d_vars, name='d_op')

scripts/dump-model-params.py

@@ -35,6 +35,7 @@ with tf.Graph().as_default() as G:
     else:
         init = sessinit.SaverRestore(args.model)
     sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
+    sess.run(tf.global_variables_initializer())
     init.init(sess)
 
     # dump ...