_build_graph(inputs) -> build_graph(*inputs) (#318)

examples/A3C-Gym/train-atari.py

@@ -94,8 +94,7 @@ class Model(ModelDesc):
         value = FullyConnected('fc-v', l, 1)
         return logits, value
 
-    def _build_graph(self, inputs):
-        state, action, futurereward, action_prob = inputs
+    def build_graph(self, state, action, futurereward, action_prob):
         logits, value = self._get_NN_prediction(state)
         value = tf.squeeze(value, [1], name='pred_value')  # (B,)
         policy = tf.nn.softmax(logits, name='policy')

examples/CTC-TIMIT/train-timit.py

@@ -33,8 +33,7 @@ class Model(ModelDesc):
                 tf.placeholder(tf.int32, [None], 'seqlen'),   # b
                 ]
 
-    def _build_graph(self, inputs):
-        feat, labelidx, labelvalue, labelshape, seqlen = inputs
+    def build_graph(self, feat, labelidx, labelvalue, labelshape, seqlen):
         label = tf.SparseTensor(labelidx, labelvalue, labelshape)
 
         cell = rnn.MultiRNNCell([rnn.LSTMBlockCell(num_units=HIDDEN) for _ in range(NLAYER)])

examples/Char-RNN/char-rnn.py

@@ -74,9 +74,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.int32, (None, param.seq_len), 'input'),
                 tf.placeholder(tf.int32, (None, param.seq_len), 'nextinput')]
 
-    def _build_graph(self, inputs):
-        input, nextinput = inputs
-
+    def build_graph(self, input, nextinput):
         cell = rnn.MultiRNNCell([rnn.LSTMBlockCell(num_units=param.rnn_size)
                                 for _ in range(param.num_rnn_layer)])
 

examples/DeepQNetwork/DQNModel.py

@@ -43,8 +43,7 @@ class Model(ModelDesc):
     def get_DQN_prediction(self, image):
         return self._get_DQN_prediction(image)
 
-    def _build_graph(self, inputs):
-        comb_state, action, reward, isOver = inputs
+    def build_graph(self, comb_state, action, reward, isOver):
         comb_state = tf.cast(comb_state, tf.float32)
         state = tf.slice(comb_state, [0, 0, 0, 0], [-1, -1, -1, self.channel], name='state')
         self.predict_value = self.get_DQN_prediction(state)

examples/DisturbLabel/mnist-disturb.py

@@ -29,8 +29,7 @@ IMAGE_SIZE = 28
 
 
 class Model(mnist_example.Model):
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = tf.expand_dims(image, 3)
 
         logits = (LinearWrap(image)  # the starting brace is oactivationy for line-breaking

examples/DoReFa-Net/alexnet-dorefa.py

@@ -81,8 +81,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 224, 224, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = image / 255.0
 
         fw, fa, fg = get_dorefa(BITW, BITA, BITG)

examples/DoReFa-Net/resnet-dorefa.py

@@ -36,8 +36,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 224, 224, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = image / 256.0
 
         fw, fa, fg = get_dorefa(BITW, BITA, BITG)

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -47,8 +47,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 40, 40, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         is_training = get_current_tower_context().is_training
 
         fw, fa, fg = get_dorefa(BITW, BITA, BITG)

examples/DynamicFilterNetwork/steering-filter.py

@@ -120,9 +120,8 @@ class Model(ModelDesc):
         logger.info('Parameter net output: {}'.format(pred_filter.get_shape().as_list()))
         return pred_filter
 
-    def _build_graph(self, inputs):
+    def build_graph(self, theta, image, gt_image, gt_filter):
         kernel_size = 9
-        theta, image, gt_image, gt_filter = inputs
 
         image = image
         gt_image = gt_image

examples/FasterRCNN/train.py

@@ -90,7 +90,7 @@ class Model(ModelDesc):
                     -1, -1]), name='fm_anchors')
             return fm_anchors
 
-    def _build_graph(self, inputs):
+    def build_graph(self, *inputs):
         is_training = get_current_tower_context().is_training
         if config.MODE_MASK:
             image, anchor_labels, anchor_boxes, gt_boxes, gt_labels, gt_masks = inputs

examples/GAN/BEGAN.py

@@ -73,8 +73,7 @@ class Model(GANModelDesc):
                  .FullyConnected('fc', NH)())
         return l
 
-    def _build_graph(self, inputs):
-        image_pos = inputs[0]
+    def build_graph(self, image_pos):
         image_pos = image_pos / 128.0 - 1
 
         z = tf.random_uniform([args.batch, args.z_dim], minval=-1, maxval=1, name='z_train')

examples/GAN/ConditionalGAN-mnist.py

@@ -84,8 +84,7 @@ class Model(GANModelDesc):
                  .FullyConnected('fct', 1, activation=tf.identity)())
         return l
 
-    def _build_graph(self, inputs):
-        image_pos, y = inputs
+    def build_graph(self, image_pos, y):
         image_pos = tf.expand_dims(image_pos * 2.0 - 1, -1)
         y = tf.one_hot(y, 10, name='label_onehot')
 

examples/GAN/CycleGAN.py

@@ -85,8 +85,7 @@ class Model(GANModelDesc):
                  .Conv2D('conv4', 1, strides=1, activation=tf.identity, use_bias=True)())
         return l
 
-    def _build_graph(self, inputs):
-        A, B = inputs
+    def build_graph(self, A, B):
         with tf.name_scope('preprocess'):
             A = tf.transpose(A / 128.0 - 1.0, [0, 3, 1, 2])
             B = tf.transpose(B / 128.0 - 1.0, [0, 3, 1, 2])

examples/GAN/DCGAN.py

@@ -76,8 +76,7 @@ class Model(GANModelDesc):
                  .FullyConnected('fct', 1)())
         return l
 
-    def _build_graph(self, inputs):
-        image_pos = inputs[0]
+    def build_graph(self, image_pos):
         image_pos = image_pos / 128.0 - 1
 
         z = tf.random_uniform([self.batch, self.zdim], -1, 1, name='z_train')

examples/GAN/DiscoGAN-CelebA.py

@@ -78,8 +78,7 @@ class Model(GANModelDesc):
         add_moving_summary(ret)
         return ret
 
-    def _build_graph(self, inputs):
-        A, B = inputs
+    def build_graph(self, A, B):
         A = tf.transpose(A / 255.0, [0, 3, 1, 2])
         B = tf.transpose(B / 255.0, [0, 3, 1, 2])
 

examples/GAN/GAN.py

@@ -61,7 +61,7 @@ class GANModelDesc(ModelDescBase):
 
             add_moving_summary(self.g_loss, self.d_loss, d_accuracy, g_accuracy)
 
-    def _build_graph(self, inputs):
+    def build_graph(self, *inputs):
         """
         Have to build one tower and set the following attributes:
         g_loss, d_loss, g_vars, d_vars.

examples/GAN/Image2Image.py

@@ -117,8 +117,7 @@ class Model(GANModelDesc):
                  .Conv2D('convlast', 1, strides=1, padding='VALID', activation=tf.identity)())
         return l
 
-    def _build_graph(self, inputs):
-        input, output = inputs
+    def build_graph(self, input, output):
         input, output = input / 128.0 - 1, output / 128.0 - 1
 
         with argscope([Conv2D, Conv2DTranspose], kernel_initializer=tf.truncated_normal_initializer(stddev=0.02)):

examples/GAN/Improved-WGAN.py

@@ -41,8 +41,7 @@ class Model(DCGAN.Model):
                  .FullyConnected('fct', 1, activation=tf.identity)())
         return tf.reshape(l, [-1])
 
-    def _build_graph(self, inputs):
-        image_pos = inputs[0]
+    def build_graph(self, image_pos):
         image_pos = image_pos / 128.0 - 1
 
         z = tf.random_normal([self.batch, self.zdim], name='z_train')

examples/GAN/InfoGAN-mnist.py

@@ -138,8 +138,7 @@ class Model(GANModelDesc):
                        .FullyConnected('fce-out', DIST_PARAM_DIM)())
         return logits, encoder
 
-    def _build_graph(self, inputs):
-        real_sample = inputs[0]
+    def build_graph(self, real_sample):
         real_sample = tf.expand_dims(real_sample, -1)
 
         # sample the latent code:

examples/HED/hed.py

@@ -48,8 +48,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, None, None, 3], 'image'),
                 tf.placeholder(tf.int32, [None, None, None], 'edgemap')]
 
-    def _build_graph(self, inputs):
-        image, edgemap = inputs
+    def build_graph(self, image, edgemap):
         image = image - tf.constant([104, 116, 122], dtype='float32')
         edgemap = tf.expand_dims(edgemap, 3, name='edgemap4d')
 

examples/ImageNetModels/imagenet_utils.py

@@ -152,8 +152,7 @@ class ImageNetModel(ModelDesc):
         return [tf.placeholder(self.image_dtype, [None, self.image_shape, self.image_shape, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = ImageNetModel.image_preprocess(image, bgr=True)
         if self.data_format == 'NCHW':
             image = tf.transpose(image, [0, 3, 1, 2])

examples/ImageNetModels/inception-bn.py

@@ -28,8 +28,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, INPUT_SHAPE, INPUT_SHAPE, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = image / 128.0
 
         def inception(name, x, nr1x1, nr3x3r, nr3x3, nr233r, nr233, nrpool, pooltype):

examples/PennTreebank/PTB-LSTM.py

@@ -50,9 +50,8 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.int32, (None, SEQ_LEN), 'input'),
                 tf.placeholder(tf.int32, (None, SEQ_LEN), 'nextinput')]
 
-    def _build_graph(self, inputs):
+    def build_graph(self, input, nextinput):
         is_training = get_current_tower_context().is_training
-        input, nextinput = inputs
         initializer = tf.random_uniform_initializer(-0.05, 0.05)
 
         def get_basic_cell():

examples/ResNet/cifar10-preact18-mixup.py

@@ -43,9 +43,8 @@ class ResNet_Cifar(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 32, 32, 3], 'input'),
                 tf.placeholder(tf.float32, [None, CLASS_NUM], 'label')]
 
-    def _build_graph(self, inputs):
+    def build_graph(self, image, label):
         assert tf.test.is_gpu_available()
-        image, label = inputs
 
         MEAN_IMAGE = tf.constant([0.4914, 0.4822, 0.4465], dtype=tf.float32)
         STD_IMAGE = tf.constant([0.2023, 0.1994, 0.2010], dtype=tf.float32)

examples/ResNet/cifar10-resnet.py

@@ -44,8 +44,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 32, 32, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = image / 128.0
         assert tf.test.is_gpu_available()
         image = tf.transpose(image, [0, 3, 1, 2])

examples/ResNet/load-resnet.py

@@ -32,8 +32,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 224, 224, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         blocks = CFG[DEPTH]
 
         bottleneck = functools.partial(resnet_bottleneck, stride_first=True)

examples/Saliency/CAM-resnet.py

@@ -35,8 +35,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.uint8, [None, INPUT_SHAPE, INPUT_SHAPE, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         image = image_preprocess(image, bgr=True)
         image = tf.transpose(image, [0, 3, 1, 2])
 

examples/Saliency/saliency-maps.py

@@ -57,8 +57,7 @@ class Model(tp.ModelDescBase):
     def inputs(self):
         return [tf.placeholder(tf.float32, (IMAGE_SIZE, IMAGE_SIZE, 3), 'image')]
 
-    def _build_graph(self, inputs):
-        orig_image = inputs[0]
+    def build_graph(self, orig_image):
         mean = tf.get_variable('resnet_v1_50/mean_rgb', shape=[3])
         with guided_relu():
             with slim.arg_scope(resnet_v1.resnet_arg_scope(is_training=False)):

examples/SimilarityLearning/mnist-embeddings.py

@@ -241,9 +241,7 @@ class SiameseModel(EmbeddingModel):
                 tf.placeholder(tf.float32, (None, 28, 28), 'input_y'),
                 tf.placeholder(tf.int32, (None,), 'label')]
 
-    def _build_graph(self, inputs):
-        # get inputs
-        x, y, label = inputs
+    def build_graph(self, x, y, label):
         # embed them
         x, y = self.embed([x, y])
 
@@ -261,8 +259,7 @@ class SiameseModel(EmbeddingModel):
 
 
 class CosineModel(SiameseModel):
-    def _build_graph(self, inputs):
-        x, y, label = inputs
+    def build_graph(self, x, y, label):
         x, y = self.embed([x, y])
 
         with tf.variable_scope(tf.get_variable_scope(), reuse=True):
@@ -289,8 +286,7 @@ class TripletModel(EmbeddingModel):
     def loss(self, a, p, n):
         return triplet_loss(a, p, n, 5., extra=True, scope="loss")
 
-    def _build_graph(self, inputs):
-        a, p, n = inputs
+    def build_graph(self, a, p, n):
         a, p, n = self.embed([a, p, n])
 
         with tf.variable_scope(tf.get_variable_scope(), reuse=True):
@@ -319,9 +315,7 @@ class CenterModel(EmbeddingModel):
         return [tf.placeholder(tf.float32, (None, 28, 28), 'input'),
                 tf.placeholder(tf.int32, (None,), 'label')]
 
-    def _build_graph(self, inputs):
-        # get inputs
-        x, label = inputs
+    def build_graph(self, x, label):
         # embed them
         x = self.embed(x)
 

examples/SpatialTransformer/mnist-addition.py

@@ -24,13 +24,11 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, (None, IMAGE_SIZE, IMAGE_SIZE, 2), 'input'),
                 tf.placeholder(tf.int32, (None,), 'label')]
 
-    def _build_graph(self, inputs):
+    def build_graph(self, image, label):
         xys = np.array([(y, x, 1) for y in range(WARP_TARGET_SIZE)
                         for x in range(WARP_TARGET_SIZE)], dtype='float32')
         xys = tf.constant(xys, dtype=tf.float32, name='xys')    # p x 3
 
-        image, label = inputs
-
         image = image / 255.0 - 0.5  # bhw2
 
         def get_stn(image):

examples/SuperResolution/enet-pat.py

@@ -52,9 +52,9 @@ class Model(GANModelDesc):
         return [tf.placeholder(tf.float32, (None, self.height * 1, self.width * 1, CHANNELS), 'Ilr'),
                 tf.placeholder(tf.float32, (None, self.height * 4, self.width * 4, CHANNELS), 'Ihr')]
 
-    def _build_graph(self, inputs):
+    def build_graph(self, Ilr, Ihr):
+        Ilr, Ihr = Ilr / 255.0, Ihr / 255.0
         ctx = get_current_tower_context()
-        Ilr, Ihr = inputs[0] / 255.0, inputs[1] / 255.0
         Ibicubic = tf.image.resize_bicubic(
             Ilr, [4 * self.height, 4 * self.width], align_corners=True,
             name='bicubic_baseline')    # (0,1)

examples/basics/cifar-convnet.py

@@ -30,8 +30,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, (None, 30, 30, 3), 'input'),
                 tf.placeholder(tf.int32, (None,), 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
+    def build_graph(self, image, label):
         is_training = get_current_tower_context().is_training
         keep_prob = tf.constant(0.5 if is_training else 1.0)
 

examples/basics/mnist-tflayers.py

@@ -31,10 +31,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, (None, IMAGE_SIZE, IMAGE_SIZE), 'input'),
                 tf.placeholder(tf.int32, (None,), 'label')]
 
-    def _build_graph(self, inputs):
-        # inputs contains a list of input variables defined above
-        image, label = inputs
-
+    def build_graph(self, image, label):
         # In tensorflow, inputs to convolution function are assumed to be
         # NHWC. Add a single channel here.
         image = tf.expand_dims(image, 3)

examples/basics/mnist-visualizations.py

@@ -72,9 +72,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, (None, IMAGE_SIZE, IMAGE_SIZE), 'input'),
                 tf.placeholder(tf.int32, (None,), 'label')]
 
-    def _build_graph(self, inputs):
-
-        image, label = inputs
+    def build_graph(self, image, label):
         image = tf.expand_dims(image * 2 - 1, 3)
 
         with argscope(Conv2D, kernel_shape=3, nl=tf.nn.relu, out_channel=32):
@@ -103,7 +101,7 @@ class Model(ModelDesc):
         cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
         cost = tf.reduce_mean(cost, name='cross_entropy_loss')
 
-        accuracy = tf.reduce_mean(tf.to_float(tf.nn.in_top_k(logits, label, 1)), name='accuracy')
+        tf.reduce_mean(tf.to_float(tf.nn.in_top_k(logits, label, 1)), name='accuracy')
 
         wd_cost = tf.multiply(1e-5,
                               regularize_cost('fc.*/W', tf.nn.l2_loss),

examples/basics/svhn-digit-convnet.py

@@ -26,9 +26,7 @@ class Model(ModelDesc):
         return [tf.placeholder(tf.float32, [None, 40, 40, 3], 'input'),
                 tf.placeholder(tf.int32, [None], 'label')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
-
+    def build_graph(self, image, label):
         image = image / 128.0 - 1
 
         with argscope(Conv2D, activation=BNReLU, use_bias=False):

examples/boilerplate.py

@@ -20,15 +20,14 @@ CHANNELS = 3
 
 class Model(ModelDesc):
     def inputs(self):
-        return [tf.placeholder(tf.float32, (None, SHAPE, SHAPE, CHANNELS), 'input'),
-                tf.placeholder(tf.int32, (None,), 'label')]
+        return [tf.placeholder(tf.float32, (None, SHAPE, SHAPE, CHANNELS), 'input1'),
+                tf.placeholder(tf.int32, (None,), 'input2')]
 
-    def _build_graph(self, inputs):
-        image, label = inputs
-        image = image * 2 - 1
+    def build_graph(self, input1, input2):
 
-        self.cost = tf.identity(0., name='total_costs')
-        summary.add_moving_summary(self.cost)
+        cost = tf.identity(input1 - input2, name='total_costs')
+        summary.add_moving_summary(cost)
+        return cost
 
     def _get_optimizer(self):
         lr = tf.get_variable('learning_rate', initializer=5e-3, trainable=False)