refactoring / add lsgan (wip) / use basemodel / fix dcgan bugs

Author: khanrc

README.md

@@ -6,10 +6,15 @@
 * MNIST 에 대해서는 해 봤지만 사실 MNIST 는 데이터셋이 간단해서 구별이 잘 안 감
 * 따라서 CelebA 나 LSUN 등 좀 복잡한 데이터셋에 대해서 해 보면서 몇몇 코드들 리팩토링도 하고...
 
+## Checks
+
+* Data read 가 좀 이상함. 20 epoch 을 주면 10 epoch 만 돌고 끝남. epoch 계산을 내가 잘못하고 있나? 체크해보자.
+* LSGAN hyperparams
+
 ## GANs
 
-* [ ] DCGAN
-* [ ] LSGAN
+* [x] DCGAN
+* [ ] LSGAN - WIP
 * [ ] EBGAN
 * [ ] WGAN
 * [ ] WGAN-GP
@@ -26,4 +31,4 @@
 ## Resuable code
 
 * `utils.py`
-* `inputpipe.py`
+* `inputpipe.py`

basemodel.py

@@ -0,0 +1,45 @@
+# coding: utf-8
+
+'''
+BaseModel for Generative Adversarial Netowrks.
+이 모델을 상속받아서 _build_train_graph, _discriminator, _generator 세가지만 구현해주면 된다.
+'''
+
+import tensorflow as tf
+slim = tf.contrib.slim
+
+
+class BaseModel(object):
+    def __init__(self, input_pipe, z_dim=100, name='basemodel'):
+        '''
+        training mode: input_pipe = input pipeline
+        generation mode: input_pipe = None
+        '''
+        self.name = name
+        training = input_pipe is not None
+        # check: DCGAN specified BN-params?
+        self.bn_params = {
+            "decay": 0.99,
+            "epsilon": 1e-5,
+            "scale": True,
+            "is_training": training
+        }
+        self.z_dim = z_dim
+        if training == True:
+            self._build_train_graph(input_pipe)
+        else:
+            self._build_gen_graph()
+
+
+    def _build_gen_graph(self):
+        '''build computational graph for generation (evaluation)
+        '''
+        with tf.variable_scope(self.name):
+            self.z = tf.placeholder(tf.float32, [None, self.z_dim])
+            self.fake_sample = self._generator(self.z)
+
+
+    def _build_train_graph(self, X):
+        '''build computational graph for training
+        '''
+        pass

config.py

@@ -0,0 +1,30 @@
+import dcgan, lsgan
+
+
+model_zoo = ['DCGAN', 'LSGAN', 'WGAN', 'WGAN-GP', 'BEGAN']
+
+def get_model(mtype, name, input_pipe):
+    model = None
+    if mtype == 'DCGAN':
+        model = dcgan.DCGAN
+    elif mtype == 'LSGAN':
+        model = lsgan.LSGAN
+    elif mtype == 'WGAN':
+        pass
+    elif mtype == 'WGAN-GP':
+        pass
+    elif mtype == 'BEGAN':
+        pass
+    else:
+        assert False, mtype + ' is not in the model zoo'
+
+    assert model, mtype + ' is work in progress'
+
+    return model(input_pipe=input_pipe, name=name)
+
+
+def pprint_args(FLAGS):
+    print("\nParameters:")
+    for attr, value in sorted(vars(FLAGS).items()):
+        print("{}={}".format(attr.upper(), value))
+    print("")

dcgan.py

@@ -1,7 +1,9 @@
 # coding: utf-8
 import tensorflow as tf
 slim = tf.contrib.slim
-from utils import *
+from utils import expected_shape
+import ops
+from basemodel import BaseModel
 
 '''
 일단 MNIST 는 무시하자... 귀찮다.
@@ -19,44 +21,13 @@
 init - normal dist + stddev 0.02
 '''
 
-def lrelu(inputs, leak=0.2, scope="lrelu"):
-    """
-    https://github.com/tensorflow/tensorflow/issues/4079
-    """
-    with tf.variable_scope(scope):
-        f1 = 0.5 * (1 + leak)
-        f2 = 0.5 * (1 - leak)
-        return f1 * inputs + f2 * abs(inputs)
-
-
-class DCGAN(object):
-    def __init__(self, input_pipe, z_dim=100, name='dcgan'):
+class DCGAN(BaseModel):
+    def __init__(self, input_pipe, z_dim=100, name='dcgan2'):
         '''
         training mode: input_pipe = input pipeline
         generation mode: input_pipe = None
         '''
-        self.name = name
-        training = bool(input_pipe)
-        # check: DCGAN specified BN-params?
-        self.bn_params = {
-            "decay": 0.99,
-            "epsilon": 1e-5,
-            "scale": True,
-            "is_training": training
-        }
-        self.z_dim = z_dim
-        if training == True:
-            self._build_train_graph(input_pipe)
-        else:
-            self._build_gen_graph()
-
-
-    def _build_gen_graph(self):
-        '''build computational graph for generation (evaluation)
-        '''
-        with tf.variable_scope(self.name):
-            self.z = tf.placeholder(tf.float32, [None, self.z_dim])
-            self.fake_sample = self._generator(self.z)
+        super(DCGAN, self).__init__(input_pipe=input_pipe, z_dim=z_dim, name=name)
 
 
     def _build_train_graph(self, X):
@@ -115,7 +86,7 @@ def _discriminator(self, X, reuse=False):
         with tf.variable_scope('D', reuse=reuse):
             net = X
 
-            with slim.arg_scope([slim.conv2d], kernel_size=[5,5], stride=2, padding='SAME', activation_fn=lrelu, 
+            with slim.arg_scope([slim.conv2d], kernel_size=[5,5], stride=2, padding='SAME', activation_fn=ops.lrelu, 
                 normalizer_fn=slim.batch_norm, normalizer_params=self.bn_params):
                 net = slim.conv2d(net, 64, normalizer_fn=None)
                 expected_shape(net, [32, 32, 64])
@@ -141,13 +112,13 @@ def _generator(self, z, reuse=False):
 
             with slim.arg_scope([slim.conv2d_transpose], kernel_size=[5,5], stride=2, padding='SAME', activation_fn=tf.nn.relu, 
                 normalizer_fn=slim.batch_norm, normalizer_params=self.bn_params):
-                net = slim.conv2d_transpose(net, 512, normalizer_fn=None)
+                net = slim.conv2d_transpose(net, 512)
                 expected_shape(net, [8, 8, 512])
                 net = slim.conv2d_transpose(net, 256)
                 expected_shape(net, [16, 16, 256])
                 net = slim.conv2d_transpose(net, 128)
                 expected_shape(net, [32, 32, 128])
-                net = slim.conv2d_transpose(net, 3, activation_fn=tf.nn.tanh)
+                net = slim.conv2d_transpose(net, 3, activation_fn=tf.nn.tanh, normalizer_fn=None)
                 expected_shape(net, [64, 64, 3])
 
                 return net

eval.py

@@ -3,6 +3,7 @@
 from dcgan import DCGAN
 import numpy as np
 import utils
+from config import get_model, pprint_args
 import os, glob
 import scipy.misc
 from argparse import ArgumentParser
@@ -12,6 +13,7 @@
 def build_parser():
     parser = ArgumentParser()
     parser.add_argument('--model', help='DCGAN / LSGAN / WGAN / WGAN-GP / BEGAN', required=True) # DRAGAN, CramerGAN
+    parser.add_argument('--name', help='default: model')
 
     return parser
 
@@ -39,17 +41,19 @@ def get_all_checkpoints(ckpt_dir, force=False):
     return ckpts
 
 
-def eval(model, sample_shape=[4,4], load_all_ckpt=True):
-    dir_name = 'eval_' + model.name
+def eval(model, name, sample_shape=[4,4], load_all_ckpt=True):
+    if name == None:
+        name = model.name
+    dir_name = 'eval_' + name
     if tf.gfile.Exists(dir_name):
         tf.gfile.DeleteRecursively(dir_name)
     tf.gfile.MkDir(dir_name)
 
     # training=False => generator 만 생성
     restorer = tf.train.Saver(slim.get_model_variables())
     with tf.Session() as sess:
-        # ckpt = tf.train.get_checkpoint_state('./checkpoints/' + model.name)
-        ckpts = get_all_checkpoints('./checkpoints/' + model.name, force=load_all_ckpt)
+        # ckpt = tf.train.get_checkpoint_state('./checkpoints/' + name)
+        ckpts = get_all_checkpoints('./checkpoints/' + name, force=load_all_ckpt)
         size = sample_shape[0] * sample_shape[1]
 
         z_ = sample_z([size, model.z_dim])
@@ -88,7 +92,7 @@ def to_gif(dir_name='eval'):
     parser = build_parser()
     FLAGS = parser.parse_args()
     FLAGS.model = FLAGS.model.upper()
-    utils.pprint_args(FLAGS)
+    pprint_args(FLAGS)
 
-    model = utils.get_model(FLAGS.model, training=False, X=None)
-    eval(model, sample_shape=[4,4], load_all_ckpt=True)
+    model = get_model(FLAGS.model, FLAGS.name, input_pipe=None)
+    eval(model, name=FLAGS.name, sample_shape=[4,4], load_all_ckpt=True)