|
|
import tensorflow.compat.v1 as tf
|
|
|
import utils
|
|
|
tf.compat.v1.disable_eager_execution()
|
|
|
|
|
|
def get_model(is_train=False, keep_prob=0.8, alpha=0.8):
|
|
|
with tf.variable_scope("model", reuse=not is_train):
|
|
|
tf_x = tf.placeholder(tf.float32, shape=[None, utils.IMAGE_HEIGHT, utils.IMAGE_WIDTH, utils.IMAGE_CHANNEL])
|
|
|
tf_y = tf.placeholder(tf.int32, shape=[None])
|
|
|
tf_y_one_hot = tf.one_hot(tf_y, utils.IMAGE_CLASSIFY)
|
|
|
|
|
|
# conv 1
|
|
|
filter_1 = tf.Variable(tf.truncated_normal(shape=[5, 5, 3, 32], mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_1 = tf.Variable(tf.constant(0.1, shape=[32]))
|
|
|
conv_1 = tf.nn.conv2d(tf_x, filter=filter_1, strides=[1, 2, 2, 1], padding='SAME') + bias_1
|
|
|
leaky_relu_1 = tf.nn.leaky_relu(conv_1, alpha=alpha)
|
|
|
|
|
|
# conv 2
|
|
|
filter_2 = tf.Variable(tf.truncated_normal(shape=[3, 3, 32, 48], mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_2 = tf.Variable(tf.constant(0.1, shape=[48]))
|
|
|
conv_2 = tf.nn.conv2d(leaky_relu_1, filter=filter_2, strides=[1, 2, 2, 1], padding='SAME') + bias_2
|
|
|
leaky_relu_2 = tf.nn.leaky_relu(conv_2, alpha=alpha)
|
|
|
|
|
|
# conv 3
|
|
|
filter_3 = tf.Variable(tf.truncated_normal(shape=[3, 3, 48, 64], mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_3 = tf.Variable(tf.constant(0.1, shape=[64]))
|
|
|
conv_3 = tf.nn.conv2d(leaky_relu_2, filter=filter_3, strides=[1, 2, 2, 1], padding='SAME') + bias_3
|
|
|
leaky_relu_3 = tf.nn.leaky_relu(conv_3, alpha=alpha)
|
|
|
|
|
|
dropout = tf.nn.dropout(leaky_relu_3, keep_prob=keep_prob)
|
|
|
|
|
|
# flatten，Flatten层用来将输入“压平”，即把多维的输入一维化，常用在从卷积层到全连接层的过渡
|
|
|
shape = dropout.get_shape().as_list()
|
|
|
flatten_size = shape[1] * shape[2] * shape[3]
|
|
|
flatten = tf.reshape(dropout, [-1, flatten_size])
|
|
|
|
|
|
# fc 1
|
|
|
filter_4 = tf.Variable(tf.truncated_normal(shape=[flatten.get_shape().as_list()[1], 100],
|
|
|
mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_4 = tf.Variable(tf.constant(0.1, shape=[100]))
|
|
|
fc_1 = tf.matmul(flatten, filter_4) + bias_4
|
|
|
leaky_relu_4 = tf.nn.leaky_relu(fc_1, alpha=alpha)
|
|
|
|
|
|
# fc 2
|
|
|
filter_5 = tf.Variable(tf.truncated_normal(shape=[100, 50], mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_5 = tf.Variable(tf.constant(0.1, shape=[50]))
|
|
|
fc_2 = tf.matmul(leaky_relu_4, filter_5) + bias_5
|
|
|
leaky_relu_5 = tf.nn.leaky_relu(fc_2, alpha=alpha)
|
|
|
|
|
|
# fc 3
|
|
|
filter_6 = tf.Variable(tf.truncated_normal(shape=[50, 10], mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_6 = tf.Variable(tf.constant(0.1, shape=[10]))
|
|
|
fc_3 = tf.matmul(leaky_relu_5, filter_6) + bias_6
|
|
|
leaky_relu_6 = tf.nn.leaky_relu(fc_3, alpha=alpha)
|
|
|
|
|
|
# result
|
|
|
filter_7 = tf.Variable(tf.truncated_normal(shape=[10, utils.IMAGE_CLASSIFY],
|
|
|
mean=utils.mu, stddev=utils.sigma))
|
|
|
bias_7 = tf.Variable(tf.constant(0.1, shape=[utils.IMAGE_CLASSIFY]))
|
|
|
result = tf.matmul(leaky_relu_6, filter_7) + bias_7
|
|
|
|
|
|
last_layer = result if is_train else tf.nn.softmax(result)
|
|
|
return tf_x, tf_y, tf_y_one_hot, last_layer |
