def discriminator(self,image, df_dim, reuse=True, name="discriminator"):
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
# image = [1,H,W,3]
tf.get_variable_scope().reuse_variables()
image = tf.expand_dims(image, 0)
print(image.shape)
h0 = lrelu(conv2d(image, self.df_dim, name='d_h0_conv'))
print('tt',h0.shape)
# h0 is [,H/2,W/2,self.df_dim]
h1 = lrelu(pixel_norm(conv2d(h0, self.df_dim * 2, name='d_h1_conv')))
print(h1.shape)
# h1 is [,H/4,W/4,self.df_dim*2]
h2 = lrelu(pixel_norm(conv2d(h1, self.df_dim * 4, name='d_h2_conv')))
print(h2.shape)
# h2 is [,H/8,W/8,self.df_dim*4]
h3 = lrelu(pixel_norm(conv2d(h2, self.df_dim * 8, s=1, name='d_h3_conv')))
print(h3.shape)
# h3 is [,H/8,W/8,self.df_dim*8]
h4 = conv2d(h3, 1, s=1, name='d_h3_pred')
print(h4.shape)
# h4 is [,H/8,W/8,1]
return h4
total_regions = []
total_asfmap = []
bbox = [(45, 90), (90, 45), (64, 64), (90, 180), (180, 90), (128, 128),
(181, 362), (362, 181), (256, 256), (362, 724), (724, 362), (512, 512)]
conv_height = 14
conv_width = 14
height = 600
width = 800
k = 12
gt_num = 10
feat_input = tf.placeholder(tf.float32, [None, conv_height, conv_width, 512])
with tf.variable_scope('rcnn', reuse=None):
W_conv6 = weight_variable([3, 3, 512, 256], name="W_conv6")
b_conv6 = bias_variable([256], name="b_conv6")
feat = conv2d(feat_input, W_conv6) + b_conv6
W_offset = weight_variable([1, 1, 256, k * 4], name="W_offset")
b_offset = bias_variable([k * 4], name="b_offset")
offset = conv2d(feat, W_offset) + b_offset
offset = tf.reshape(offset, [k * conv_height * conv_width, 4])
W_score = weight_variable([1, 1, 256, k], name="W_score")
b_score = bias_variable([k], name="b_score")
score = conv2d(feat, W_score) + b_score
score = tf.reshape(score, [k * conv_height * conv_width])
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
sess = tf.Session(config=config)
white_max): # 0.95这个参数请多调整，对应下面的0.05
end_ = m
break
return end_
#########################前向网络##########################################
# define placeholder for inputs to network
xs = tf.placeholder(tf.float32, [None, 28, 28, 1]) # 28x28
x_image = tf.reshape(xs, [-1, 28, 28, 1])
## conv1 layer ##
W_conv1 = func.weight_variable([5, 5, 1,
32]) # patch 5x5, in size 1, out size 32
b_conv1 = func.bias_variable([32])
h_conv1 = tf.nn.relu(func.conv2d(x_image, W_conv1) +
b_conv1) # output size 28x28x32
h_pool1 = func.max_pool_2x2(h_conv1) # output size 14x14x32
## conv2 layer ##
W_conv2 = func.weight_variable([5, 5, 32,
64]) # patch 5x5, in size 32, out size 64
b_conv2 = func.bias_variable([64])
h_conv2 = tf.nn.relu(func.conv2d(h_pool1, W_conv2) +
b_conv2) # output size 14x14x64
h_pool2 = func.max_pool_2x2(h_conv2) # output size 7x7x64
## fc1 layer ##
W_fc1 = func.weight_variable([7 * 7 * 64, 1024])
b_fc1 = func.bias_variable([1024])
# [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64]