yxl502
commented
1 year ago import tensorflow as tf import tensorflow.compat.v1 as tf tf.disable_v2_behavior() import tf_slim as slim from utils2.D_config import D_model_cfg from base.base_model import BaseModel import utils2.TensorflowUtils as utils3
定义权重变量
def weight_variable(shape): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial)
定义偏置向量
def bias_variable(shape): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial)
class D_Model(BaseModel):
def __init__(self, config):
super(D_Model, self).__init__(config)
self._scope = 'wm_model'
self.learing_rate = config.learning_rate
self.build_model()
self.init_saver()
def init_saver(self):
self.saver = tf.train.Saver(max_to_keep=self.config.saver_max_to_keep)
def build_model(self):
self.data = tf.placeholder(tf.float32, shape=[None, 36, 24, 1])
self.real = tf.placeholder(tf.float32, shape=[None, 10])
self.keep_prob = tf.placeholder(tf.float32)
with slim.arg_scope([slim.conv2d, slim.conv2d_in_plane, slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.01),
# weights_regularizer=tf.contrib.layers.l2_regularizer(self.config.weight_decay),
weights_regularizer=tf.keras.regularizers.l2(self.config.weight_decay),
biases_initializer=tf.constant_initializer(0.0)):
with tf.variable_scope(self._scope):
conv1 = slim.repeat(self.data, 2, slim.conv2d, 32, [5, 5], padding='SAME', scope='conv1')
print ('conv1:', conv1.get_shape())
pool1 = slim.max_pool2d(conv1, [2, 2], padding='VALID', scope='pool1')
print ('pool1:', pool1.get_shape())
conv2 = slim.repeat(pool1, 2, slim.conv2d, 64, [5, 5], padding='SAME', scope='conv2')
pool2 = slim.max_pool2d(conv2, [2, 2], padding='VALID', scope='pool2')
print ('pool2:', pool2.get_shape())
conv3 = tf.reshape(pool2, [-1, 3456])
W_1 = utils3.weight_variable([3456, 1024])
b_1 = utils3.bias_variable([1024])
conv4 = tf.nn.relu(tf.matmul(conv3, W_1) + b_1)
drop1 = tf.nn.dropout(conv4, keep_prob=self.keep_prob)
W_2 = utils3.weight_variable([1024, 10])
b_2 = utils3.bias_variable([10])
self.output = tf.nn.softmax(tf.matmul(drop1, W_2) + b_2)
print ('output:', self.output.get_shape())
self._build_loss()
return
def _build_loss(self):
with tf.variable_scope('loss') as (scope):
number_output = self.output.get_shape()[0]
print (number_output)
self.loss = -tf.reduce_sum(self.real * tf.log(self.output + 1e-10))
self.optimizer = tf.train.AdamOptimizer(
self.config.learning_rate).minimize(self.loss, global_step=self.global_step_tensor)
可以要一下d_model.py和wm_model.py这两个文件吗，谢谢