def __init__(self):
# ..........................................................................
print('data pre-processing is done')
self.x = tf.placeholder(shape=[self.batch_size, self.word_size],
dtype=tf.int32)
self.y_ = tf.placeholder(shape=[self.batch_size], dtype=tf.float32)
embeddings = tf.Variable(
tf.random_uniform([566 + 2, self.embedding_size], -1.0, 1.0))
embed = tf.nn.embedding_lookup(embeddings, self.x)
X = tf.reshape(
embed, [self.batch_size, self.word_size, self.embedding_size, 1])
c1 = ml.conv2d(X,
conv_filter=[10, self.embedding_size, 1, 2],
padding='VALID',
ksize=[1, 10, 1, 1],
pool_stride=[1, 4, 1, 1],
pool_padding='SAME')
c2 = ml.conv2d(c1,
conv_filter=[4, 1, 2, 4],
padding='SAME',
ksize=[1, 10, 1, 1],
pool_stride=[1, 5, 1, 1],
pool_padding='SAME')
c3 = ml.conv2d(c2,
conv_filter=[5, 1, 4, 8],
padding='VALID',
ksize=[1, 1, 1, 1],
pool_stride=[1, 1, 1, 1],
pool_padding='VALID')
out = tf.reshape(c3, shape=[self.batch_size, 8])
self.y = tf.nn.sigmoid(ml.layer_basic(out, 1))[:, 0]
# ...................................................................
self.sess = tf.Session()
saver = tf.train.Saver()
saver.restore(
self.sess,
'/home/liangoy/Desktop/project/xingqiao_model/msgTfidf566')
train_data_t = np.array([i for i in data[:-1 * batch_size] if i[0] == 1])
def next(batch_size=batch_size):
r_t = np.random.randint(0, len(train_data_t), int(batch_size / 2))
r_f = np.random.randint(0, len(train_data_f), int(batch_size / 2))
data = np.concatenate([train_data_t[r_t], train_data_f[r_f]])
return data[:, 1:], data[:, 0]
x = tf.placeholder(shape=[batch_size, score_size], dtype=tf.float32)
y_ = tf.placeholder(shape=[batch_size], dtype=tf.float32)
X = tf.reshape(x, [batch_size, score_size, 1, 1])
c1 = ml.conv2d(X,
conv_filter=[4, 1, 1, 2],
ksize=[1, 11, 1, 1],
pool_stride=[1, 10, 1, 1])
c2 = ml.conv2d(c1,
conv_filter=[4, 1, 2, 4],
ksize=[1, 30, 1, 1],
pool_stride=[1, 20, 1, 1])
c3 = ml.conv2d(c2,
conv_filter=[5, 1, 4, 8],
padding='VALID',
ksize=[1, 1, 1, 1],
pool_stride=[1, 1, 1, 1])
c_out = tf.reshape(c3, [batch_size, 8])
lay2 = ml.layer_basic(c_out, 1)
y = tf.nn.sigmoid(lay2[:, 0])
sample = data[i:i + long]
a.append(
np.concatenate([sample[:-1, :10], [[sample[-1][0]]] * (long - 1)],
axis=-1))
b.append(sample[-1][otype])
return a, b
x = tf.placeholder(shape=[batch_size, long - 1, 11], dtype=tf.float16)
y_ = tf.placeholder(shape=[batch_size], dtype=tf.float16)
X = tf.reshape(tf.nn.tanh(x), [batch_size, long - 1, x.shape[-1], 1])
c1 = ml.conv2d(X,
conv_filter=[1, x.shape[-1], 1, 4],
padding='VALID',
ksize=[1, 1, 1, 1],
pool_padding='VALID',
nn=tf.nn.tanh)
c2 = ml.conv2d(c1,
conv_filter=[4, 1, 4, 6],
padding='SAME',
ksize=[1, 6, 1, 1],
pool_stride=[1, 5, 1, 1],
pool_padding='SAME',
nn=tf.nn.tanh)
c3 = ml.conv2d(c2,
conv_filter=[3, 1, 6, 8],
padding='SAME',
ksize=[1, 6, 1, 1],
pool_stride=[1, 6, 1, 1],
pool_padding='VALID',
x_test, y_test = np.array(test_data.drop(word_size, axis=1)), np.array(
test_data[word_size])
x = tf.placeholder(shape=[batch_size, word_size], dtype=tf.int32)
y_ = tf.placeholder(shape=[batch_size], dtype=tf.float32)
embeddings = tf.Variable(
tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
embed = tf.nn.embedding_lookup(embeddings, x + 1)
X = tf.reshape(embed, [batch_size, word_size, embedding_size, 1])
c1 = ml.conv2d(X,
conv_filter=[3, embedding_size, 1, 2],
padding='VALID',
ksize=[1, 20, 1, 1],
pool_stride=[1, 10, 1, 1],
pool_padding='SAME')
c2 = ml.conv2d(c1,
conv_filter=[4, 1, 2, 4],
padding='SAME',
ksize=[1, 10, 1, 1],
pool_stride=[1, 10, 1, 1],
pool_padding='SAME')
c3 = ml.conv2d(c2,
conv_filter=[2, 1, 4, 8],
padding='SAME',
ksize=[1, 10, 1, 1],
pool_stride=[1, 10, 1, 1],
pool_padding='VALID')
# lay1 = tf.reshape(c2, [batch_size, -1])
import tensorflow as tf
from util.image import generate_image
from util import ml
batch_size = 128
shape = [batch_size] + list(generate_image(1)[1][0].shape)
print(shape)
x = tf.placeholder(shape=shape, dtype=tf.float32)
y_ = tf.placeholder(shape=[batch_size], dtype=tf.int32)
training = tf.placeholder(dtype=tf.bool)
c1 = ml.conv2d(x, conv_filter=[3, 4, 3, 8], ksize=[1, 4, 5, 1], pool_stride=[1, 3, 4, 1]) # [20,40 ]
c2 = ml.conv2d(c1, conv_filter=[3, 4, 8, 16], ksize=[1, 4, 5, 1], pool_stride=[1, 2, 4, 1]) # [10,10]
c3 = ml.conv2d(c2, conv_filter=[3, 4, 16, 32], ksize=[1, 3, 3, 1], pool_stride=[1, 2, 2, 1]) # [5,5]
w = tf.Variable(tf.random_uniform([5, 5, 32, 128], -1.0, 1.0))
b = tf.Variable(tf.random_uniform([128], -1.0, 1.0))
c4 = tf.nn.conv2d(c3, filter=w, strides=[1, 1, 1, 1], padding='VALID')
out = tf.reshape(c4, shape=[batch_size, 128])
y = tf.nn.softmax(ml.layer_basic(tf.layers.batch_normalization(out, axis=-1, training=training), 36))
loss = -tf.reduce_mean(tf.one_hot(y_, depth=36) * tf.log(y + 0.0001)) / batch_size / tf.log(2.0)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer(learning_rate=0.01).minimize(loss)
y_out = tf.argmax(y, axis=1)
print(shape)
x = tf.placeholder(shape=shape, dtype=tf.float32)
y_ = tf.placeholder(shape=[batch_size, 4], dtype=tf.int32)
training = tf.placeholder(dtype=tf.bool)
X = tf.layers.batch_normalization(x,
training=True,
trainable=False,
scale=False,
center=False,
axis=0)
c1 = ml.conv2d(tf.expand_dims(X, axis=-1),
conv_filter=[5, 5, 1, 32],
ksize=[1, 3, 4, 1],
pool_stride=[1, 3, 4, 1],
nn=tf.nn.relu) # [20,40 ]
c2 = ml.conv2d(c1,
conv_filter=[3, 4, 32, 64],
ksize=[1, 2, 4, 1],
pool_stride=[1, 2, 4, 1],
nn=tf.nn.relu) # [10,10]
c3 = ml.conv2d(c2,
conv_filter=[3, 4, 64, 128],
ksize=[1, 2, 2, 1],
pool_stride=[1, 2, 2, 1],
nn=tf.nn.relu) # [5,5]
w = tf.Variable(tf.random_uniform([5, 5, 128, 256], -1.0, 1.0))
#b = tf.Variable(tf.random_uniform([512], -1.0, 1.0))
xx.append(train_data_x[i])
yy.append(train_data_y[i])
return xx, yy
x_test, y_test = test_data_x[:batch_size], test_data_y[:batch_size]
x = tf.placeholder(shape=[batch_size, msg_count, msg_size], dtype=tf.int32)
y_ = tf.placeholder(shape=[batch_size], dtype=tf.float32)
embeddings = tf.constant(embeddings)
embed = tf.nn.embedding_lookup(embeddings, x)
c1 = ml.conv2d(embed,
conv_filter=[1, 4, embedding_size, 1],
padding='VALID',
ksize=[1, 100, 10, 1],
pool_stride=[1, 100, 10, 1],
pool_padding='VALID')
# c2 = ml.conv2d(c1, conv_filter=[4, 4, 1, 1], padding='VALID', ksize=[1, 20, 5, 1],
# pool_stride=[1, 10, 2, 1],
# pool_padding='VALID')
c3 = ml.conv2d(c1,
conv_filter=[int(c1.shape[1]),
int(c1.shape[2]), 1, 1],
padding='VALID',
ksize=[1, 1, 1, 1],
pool_stride=[1, 1, 1, 1],
pool_padding='VALID')
y = tf.nn.sigmoid(ml.layer_basic(c3[:, 0, 0]))[:, 0]
#gv = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
shape = [batch_size] + list(generate_number_image(1)[1][0].shape)
print(shape)
x = tf.placeholder(shape=shape, dtype=tf.float32)
y_ = tf.placeholder(shape=[batch_size, 4], dtype=tf.int32)
training = tf.placeholder(dtype=tf.bool)
x_mean = tf.reshape(tf.tile(tf.reduce_mean(x, axis=[1, 2]), [30 * 80]),
shape=[batch_size, 30, 80])
X = tf.nn.relu(tf.abs(x - x_mean) - 30)
XX = X / (X + 0.0001)
c1 = ml.conv2d(tf.expand_dims(XX, axis=-1),
conv_filter=[5, 5, 1, 16],
ksize=[1, 3, 4, 1],
pool_stride=[1, 3, 4, 1],
nn=tf.nn.relu) # [10,20 ]
c2 = ml.conv2d(c1,
conv_filter=[3, 4, 16, 32],
ksize=[1, 2, 4, 1],
pool_stride=[1, 2, 4, 1],
nn=tf.nn.relu) # [5,5]
w = tf.Variable(tf.random_uniform([5, 5, 32, 128], -1.0, 1.0))
#b = tf.Variable(tf.random_uniform([512], -1.0, 1.0))
c3 = tf.nn.conv2d(c2, filter=w, strides=[1, 1, 1, 1], padding='VALID')
out = tf.nn.relu(
tf.layers.batch_normalization(tf.reshape(c3, shape=[batch_size, 128]),
training=training))
