better bilinear & xentropy

examples/load-vgg16.py

 #!/usr/bin/env python2
 # -*- coding: UTF-8 -*-
-# File: load_vgg16.py
+# File: load-vgg16.py
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 import tensorflow as tf
@@ -41,30 +41,30 @@ class Model(ModelDesc):
             # 224
             l = Conv2D('conv1_1', image, 64)
             l = Conv2D('conv1_2', l, 64)
-            l = MaxPooling('pool1', l, 2, stride=2, padding='VALID')
+            l = MaxPooling('pool1', l, 2)
             # 112
 
             l = Conv2D('conv2_1', l, 128)
             l = Conv2D('conv2_2', l, 128)
-            l = MaxPooling('pool2', l, 2, stride=2, padding='VALID')
+            l = MaxPooling('pool2', l, 2)
             # 56
 
             l = Conv2D('conv3_1', l, 256)
             l = Conv2D('conv3_2', l, 256)
             l = Conv2D('conv3_3', l, 256)
-            l = MaxPooling('pool3', l, 2, stride=2, padding='VALID')
+            l = MaxPooling('pool3', l, 2)
             # 28
 
             l = Conv2D('conv4_1', l, 512)
             l = Conv2D('conv4_2', l, 512)
             l = Conv2D('conv4_3', l, 512)
-            l = MaxPooling('pool4', l, 2, stride=2, padding='VALID')
+            l = MaxPooling('pool4', l, 2)
             # 14
 
             l = Conv2D('conv5_1', l, 512)
             l = Conv2D('conv5_2', l, 512)
             l = Conv2D('conv5_3', l, 512)
-            l = MaxPooling('pool5', l, 2, stride=2, padding='VALID')
+            l = MaxPooling('pool5', l, 2)
         # 7
 
         l = FullyConnected('fc6', l, 4096)

tensorpack/models/conv2d.py

@@ -30,7 +30,6 @@ def Conv2D(x, out_channel, kernel_shape,
     :returns: a NHWC tensor
     """
     in_shape = x.get_shape().as_list()
-    num_in = np.prod(in_shape[1:])
     in_channel = in_shape[-1]
     assert in_channel % split == 0
     assert out_channel % split == 0

tensorpack/models/pool.py

@@ -121,6 +121,10 @@ def BilinearUpSample(x, shape):
 
     ch = x.get_shape().as_list()[3]
     shape = int(shape)
+    unpool_mat = np.zeros((shape, shape), dtype='float32')
+    unpool_mat[-1,-1] = 1
+    x = FixedUnPooling('unpool', x, shape, unpool_mat)
+
     filter_shape = 2 * shape
     w = bilinear_conv_filler(filter_shape)
     w = np.repeat(w, ch * ch).reshape((filter_shape, filter_shape, ch, ch))
@@ -128,9 +132,6 @@ def BilinearUpSample(x, shape):
                              tf.float32,
                              shape=(filter_shape, filter_shape, ch, ch))
 
-    unpool_mat = np.zeros((shape, shape), dtype='float32')
-    unpool_mat[-1,-1] = 1
-    x = FixedUnPooling('unpool', x, shape, unpool_mat)
     output = tf.nn.conv2d(x, weight_var, [1,1,1,1], padding='SAME')
     return output
 

tensorpack/tfutils/symbolic_functions.py

@@ -69,12 +69,11 @@ def class_balanced_binary_class_cross_entropy(pred, label, name='cross_entropy_l
 
     count_neg = tf.reduce_sum(1. - y)
     count_pos = tf.reduce_sum(y)
-    total = tf.add(count_neg, count_pos)
-    beta = tf.truediv(count_neg, total)
+    beta = count_neg / (count_neg + count_pos)
 
     eps = 1e-8
-    loss_pos = tf.mul(-beta, tf.reduce_sum(tf.mul(tf.log(tf.abs(z) + eps), y), 1))
-    loss_neg = tf.mul(1. - beta, tf.reduce_sum(tf.mul(tf.log(tf.abs(1. - z) + eps), 1. - y), 1))
+    loss_pos = -beta * tf.reduce_mean(y * tf.log(tf.abs(z) + eps), 1)
+    loss_neg = (1. - beta) * tf.reduce_mean((1. - y) * tf.log(tf.abs(1. - z) + eps), 1)
     cost = tf.sub(loss_pos, loss_neg)
     cost = tf.reduce_mean(cost, name=name)
     return cost