def __init__(self, sess):
self.sess = sess
self.global_step = tf.Variable(0.0,
name='global_step',
dtype=tf.float32,
trainable=False)
self.G_para = []
self.D_para = []
#for debug
self.cnt_tep = 0
self.deb_kep = 0
self.deb_kep2 = 0
#for data input
self.pipline_data_train = cdata.get_pipline_data_train(
img_size, batchsize)
self.pipline_data_test = cdata.get_pipline_data_test(
img_size, batchsize_test)
#3个placeholder， img和noise,training
self.imgs_pla = tf.placeholder(
tf.float32,
[batchsize, img_size_h, img_size_w, G_group_img_num * img_channel],
name='imgs_in')
self.training = tf.placeholder(tf.bool, name='training_in')
print('placeholders:\n', 'img_placeholder:', self.imgs_pla,
'\ntraining:', self.training)
'''
placeholders:
img_placeholder: Tensor("imgs_in:0", shape=(12, 180, 320, 9), dtype=float32)
training: Tensor("training_in:0", dtype=bool)
'''
self.frame0 = self.imgs_pla[:, :, :, :img_channel]
self.frame1 = self.imgs_pla[:, :, :, img_channel:img_channel * 2]
self.frame2 = self.imgs_pla[:, :, :, img_channel * 2:]
#这里是为了看看第一帧和第3帧的差距，用来给evalte
self.frame_0_2_squareloss = tf.reduce_mean(tf.squared_difference(
self.frame0, self.frame2), [1, 2, 3],
name='frame_0_2_squareloss')
#frame0and2=tf.concat([self.frame0, self.frame2], -1) #在第三维度连接起来
#print ('after concat:',frame0and2)
#!!!!!!!!!!here is differs from v1,add to Generator output the ori img will reduce the generator difficulty
self.G_net = self.Generator_net(
self.frame0, self.frame2) #+self.frame0 #注意这里是直接作为生成结果
print(
'self.G_net:', self.G_net
) #self.G_net: Tensor("G_Net/G_tanh:0", shape=(12, 180, 320, 3), dtype=float32)
#D_1的输出
frame0_False_2 = tf.concat([self.frame0, self.G_net, self.frame2], -1)
self.D_linear_net_F, self.D_linear_net_F_logit = self.Discriminator_net_linear(
frame0_False_2)
self.D_linear_net_T, self.D_linear_net_T_logit = self.Discriminator_net_linear(
self.imgs_pla)
#下面是loss公式
self.D_linear_net_loss_sum, self.D_linear_net_loss_T, self.D_linear_net_loss_F=self.D_loss_TandF_logits(self.D_linear_net_T_logit, \
self.D_linear_net_F_logit, "D_linear_net")
print('D1 form finished..')
#D_2的输出
'''
self.D_clear_net_F, self.D_clear_net_F_logit=self.Discriminator_net_clear(self.G_net)
self.D_clear_net_T, self.D_clear_net_T_logit=self.Discriminator_net_clear(self.frame1)
#下面是loss公式
self.D_clear_net_loss_sum, self.D_clear_net_loss_T, self.D_clear_net_loss_F=self.D_loss_TandF_logits(self.D_clear_net_T_logit, \
self.D_clear_net_F_logit, "D_clear_net")
'''
self.G_loss_mean_Square = tf.reduce_mean(tf.squared_difference(
self.G_net, self.frame1),
name='G_clear_square_loss')
print('G_loss_mean_Square form finished..')
#这里对两个D的loss没有特殊处理，只是简单相加
self.D_loss_all = self.D_linear_net_loss_sum #+ self.D_clear_net_loss_sum
#下面是G的loss
self.G_loss_mean_D1 = self.G_loss_F_logits(self.D_linear_net_F_logit,
'G_loss_D1')
#self.G_loss_mean_D2=self.G_loss_F_logits(self.D_clear_net_F_logit, 'G_loss_D2')
#这里对两个G的loss没有特殊处理，只是简单相加
self.G_loss_all = self.G_loss_mean_D1 + self.G_loss_mean_Square * (
1 + self.global_step / G_squareloss_rate_globalstep
) # self.G_loss_mean_D2
#还是应该以tf.trainable_variables()为主
t_vars = tf.trainable_variables()
print("trainable vars cnt:", len(t_vars))
self.G_para = [var for var in t_vars if var.name.startswith('G')]
self.D_para = [var for var in t_vars if var.name.startswith('D')]
# weight clipping
self.clip_D = [
p.assign(tf.clip_by_value(p, weightclip_min, weightclip_max))
for p in self.D_para
]
#训练使用
self.train_D = self.train_op_D(decay_steps, decay_rate)
self.train_G = self.train_op_G(decay_steps, decay_rate)
'''
print ('\nshow all trainable vars:',len(tf.trainable_variables()))
for i in tf.trainable_variables():
print (i)
'''
print('\nfirst show G params')
for ind, i in enumerate(self.G_para):
print(ind, i)
print('\nnext is D:\n')
for ind, i in enumerate(self.D_para):
print(ind, i)
print('\nnext is tf.GraphKeys.UPDATE_OPS:')
print(tf.get_collection(tf.GraphKeys.UPDATE_OPS))
self.summary_all = tf.summary.merge_all()
init = tf.global_variables_initializer() #初始化tf.Variable
self.sess.run(init)
def __init__(self, sess):
self.sess=sess
self.global_step = tf.Variable(0, name='global_step', trainable=False)
self.G_para=[]
self.D_para=[]
#for debug
self.cnt_tep=0
self.deb_kep=0
self.deb_kep2=0
#for data input
self.pipline_data_train=cdata.get_pipline_data_train(img_size, batchsize)
self.pipline_data_test=cdata.get_pipline_data_test(img_size, batchsize_test)
#3个placeholder， img和noise,training
self.imgs_pla = tf.placeholder(tf.float32, [batchsize, img_size_h, img_size_w, G_group_img_num*img_channel], name='imgs_in')
self.training=tf.placeholder(tf.bool, name='training_in')
self.frame0=self.imgs_pla[:,:,:,:img_channel]
self.frame1=self.imgs_pla[:,:,:,img_channel:img_channel*2]
self.frame2=self.imgs_pla[:,:,:,img_channel*2:]
frame0and2=tf.concat([self.frame0, self.frame2], 3) #在第三维度连接起来
#print ('after concat:',frame0and2)
self.G_net=self.Generator_net(frame0and2)
#D_1的输出
frame0_False_2=tf.concat([self.frame0, self.G_net,self.frame2], 3)
self.D_linear_net_F, self.D_linear_net_F_logit=self.Discriminator_net_linear(frame0_False_2)
self.D_linear_net_T, self.D_linear_net_T_logit=self.Discriminator_net_linear(self.imgs_pla)
self.D_linear_net_loss_sum, self.D_linear_net_loss_T, self.D_linear_net_loss_F=self.D_loss_TandF_logits(self.D_linear_net_T_logit, \
self.D_linear_net_F_logit, "D_linear_net")
print ('D1 form finished..')
#D_2的输出
self.D_clear_net_F, self.D_clear_net_F_logit=self.Discriminator_net_clear(self.G_net)
self.D_clear_net_T, self.D_clear_net_T_logit=self.Discriminator_net_clear(self.frame1)
self.D_clear_net_loss_sum, self.D_clear_net_loss_T, self.D_clear_net_loss_F=self.D_loss_TandF_logits(self.D_clear_net_T_logit, \
self.D_clear_net_F_logit, "D_clear_net")
print ('D2 form finished..')
#这里对两个D的loss没有特殊处理，只是简单相加
self.D_loss_all=self.D_clear_net_loss_sum + self.D_linear_net_loss_sum
#下面是G的loss
self.G_loss_mean_D1=self.G_loss_F_logits(self.D_linear_net_F_logit, 'G_loss_D1')
self.G_loss_mean_D2=self.G_loss_F_logits(self.D_clear_net_F_logit, 'G_loss_D2')
self.G_loss_all=self.G_loss_mean_D1 + self.G_loss_mean_D2
#还是应该以tf.trainable_variables()为主
t_vars=tf.trainable_variables()
print ("trainable vars cnt:",len(t_vars))
self.G_para=[var for var in t_vars if var.name.startswith('G')]
self.D_para=[var for var in t_vars if var.name.startswith('D')]
#训练使用
self.train_D=self.train_op_D(decay_steps, decay_rate)
self.train_G=self.train_op_G(decay_steps, decay_rate)
'''
print ('\nshow all trainable vars:',len(tf.trainable_variables()))
for i in tf.trainable_variables():
print (i)
'''
print ('\nfirst show G params')
for ind,i in enumerate(self.G_para): print (ind,i)
print('\nnext is D:\n')
for ind,i in enumerate(self.D_para): print (ind,i)
print ('\nnext is tf.GraphKeys.UPDATE_OPS:')
print (tf.get_collection(tf.GraphKeys.UPDATE_OPS))
self.summary_all=tf.summary.merge_all()
init = tf.global_variables_initializer()#初始化tf.Variable
self.sess.run(init)