def inference(images, hidden1_units, hidden2_units):
# 第一层隐藏层
with tf.compat.v1.name_scope('hidden1'):
weights = tf.Variable(
tf.random.truncated_normal([IMAGE_PIXELS, hidden1_units],
stddev=1.0 /
math.sqrt(float(IMAGE_PIXELS)),
name='weights'))
biases = tf.Variable(tf.zeros([hidden1_units]), name='biases')
hidden1 = tf.nn.relu(tf.matmul(images, weights) + biases)
# 第二层隐藏层
with tf.compat.v1.name_scope('hidden2'):
weights = tf.Variable(
tf.random.truncated_normal([hidden1_units, hidden2_units],
stddev=1.0 /
math.sqrt(float(hidden1_units)),
name='weights'))
biases = tf.Variable(tf.zeros([hidden2_units]), name='biases')
hidden2 = tf.nn.relu(tf.matmul(hidden1, weights) + biases)
# 线性层，softmax
with tf.compat.v1.name_scope('softmax_linear'):
weights = tf.Variable(tf.random.truncated_normal(
[hidden2_units, NUM_CLASSES],
stddev=1.0 / math.sqrt(float(hidden2_units))),
name='weights')
biases = tf.Variable(tf.zeros([NUM_CLASSES]), name='biases')
logits = tf.matmul(hidden2, weights) + biases
return logits
def forward(self,examples,labels):
"""建立前向传播图"""
opts = self._options
# 声明所有需要的变量
# embeddings ：[vocab-size,emb_size]
init_width = 0.5 / opts.emb_dim
emb = tf.Variable(
tf.random_uniform([opts.vocab_size,opts.emb_dim], -init_width,init_width),name = "emb")
self._emb = emb
# softmax_weights:[vocab_size,emb_dim]
sm_w_t = tf.Variable(
tf.zeros([opts.vocab_size,opts.emb_dim]),name="sm_w_t")
# softmax bias:[emd_dim]
sm_b = tf.Variable(
tf.zeros([opts.vocab_size]),name="sm_b")
# global step:scalar
self.global_step = tf.Variable(0,name="global_step")
# 候选采样计算nce loss的节点
labels_matrix = tf.reshape(
tf.cast(labels,dtype=tf.int64),[opts.batch_size,1])
# 负采样
sampled_ids, _,_ = (tf.nn.fixed_unigram_candidate_sampler(
true_classes=labels_matrix,
num_true=1,
num_sampled=opts.num_samples,
unique=True,
range_max=opts.vocab_size,
distortion=0.75,
unigrams=opts.vocab_counts.tolist()))
# 样本的嵌入:[batch_size,emb_dim]
example_emb = tf.nn.embedding_lookup(emb,examples)
# 标签的权重w：[batch_size,emb_dim]
true_w = tf.nn.embedding_lookup(sm_w_t,labels)
# 标签的偏差b ：[batch_size,1]
true_b = tf.nn.embedding_lookup(sm_b,labels)
# 采样样本的ids的权重(Weights for sampled ids):[num_sampled,emb_dim]
sampled_w = tf.nn.embedding_lookup(sm_w_t, sampled_ids)
# 采样样本的 bias :[num_sampled,1]
sampled_b = tf.nn.embedding_lookup(sm_b,sampled_ids)
# True logits:[batch_size,1]
true_logits = tf.reduce_sum(tf.multiply(example_emb,true_w),1) + true_b
# 采样样本预测值 sampled logits:[batch_size,num_sampled]
sampled_b_vec = tf.reshape(sampled_b,[opts.num_samples])
sampled_logits = tf.matmul(example_emb,
sampled_w,
transpose_b=True) + sampled_b_vec
return true_logits,sampled_logits
def positional_encoding(inputs,
num_units,
zero_pad=True,
scale=True,
scope="positional_encoding",
reuse=None):
N, T = inputs.get_shape().as_list()
with tf.variable_scope(scope, reuse=True):
postion_ind = tf.tile(tf.expand_dims(tf.range(T), 0), [N, 1])
postion_enc = np.array([[
pos / np.power(10000, 2. * i / num_units) for i in range(num_units)
] for pos in range(T)])
postion_enc[:, 0::2] = np.sin(postion_enc[:, 0::2])
postion_enc[:, 1::2] = np.cos(postion_enc[:, 1::2])
lookup_table = tf.convert_to_tensor(postion_enc)
if zero_pad:
lookup_table = tf.concat(
(tf.zeros(shape=[1, num_units]), lookup_table[1:, :]), 0)
outputs = tf.nn.embedding_lookup(lookup_table, postion_ind)
if scale:
outputs = outputs * num_units**0.5
return outputs
def normalize(inputs, epsilon=1e-8, scope="ln", reuse=None):
with tf.variable_scope(scope, reuse=reuse):
inputs_shape = inputs.get_shape()
param_shape = inputs_shape[-1:]
mean, variance = tf.nn.moments(inputs, [-1], keep_dims=True)
beta = tf.Variable(tf.zeros(param_shape))
gamma = tf.Variable(tf.ones(param_shape))
normalized = (inputs - mean) / ((variance + epsilon)**(.5))
outputs = gamma * normalized + beta
return outputs
def inference(images, hidden1_units, hidden2_units):
"""Build the MNIST model up to where it may be used for inference.
Args:
images: Images placeholder, from inputs().图像占位符，输入
hidden1_units: Size of the first hidden layer.第一个隐藏层
hidden2_units: Size of the second hidden layer.
Returns:
softmax_linear: Output tensor with the computed logits.
"""
# Hidden 1 tf.name_scope
with tf.name_scope('hidden1'):
weights = tf.Variable(
# tf.truncated_normal(shape,mean,stddev)#shape表示生成Tensor的维度，mean是均值，stddev是标准差
# 这个函数产生正太分布，均值和标准差自己设定。这是一个截断的产生正太分布的函数，就是说产生正太分布的值如果与均值的差值大于两倍的标准差，那就重新生成
tf.truncated_normal([IMAGE_PIXELS, hidden1_units],
stddev=1.0 / math.sqrt(float(IMAGE_PIXELS))),
name='weights')
biases = tf.Variable(tf.zeros([hidden1_units]), name='biases')
hidden1 = tf.nn.relu(tf.matmul(images, weights) + biases)
# Hidden 2
with tf.name_scope('hidden2'):
weights = tf.Variable(tf.truncated_normal(
[hidden1_units, hidden2_units],
stddev=1.0 / math.sqrt(float(hidden1_units))),
name='weights')
biases = tf.Variable(tf.zeros([hidden2_units]), name='biases')
hidden2 = tf.nn.relu(tf.matmul(hidden1, weights) + biases)
# Linear
with tf.name_scope('softmax_linear'):
weights = tf.Variable(tf.truncated_normal(
[hidden2_units, NUM_CLASSES],
stddev=1.0 / math.sqrt(float(hidden2_units))),
name='weights')
biases = tf.Variable(tf.zeros([NUM_CLASSES]), name='biases')
logits = tf.matmul(hidden2, weights) + biases
return logits
def embedding(inputs,
vocab_size,
num_units,
zero_pad=True,
scale=True,
scope="embedding",
reuse=None):
with tf.variable_scope(scope, reuse=reuse):
lookup_table = tf.get_variable(
'lookup_table',
dtype=tf.float32,
shape=[vocab_size, num_units],
initializer=tf.contrib.layers.xavier_initializer())
if zero_pad:
lookup_table = tf.concat(
(tf.zeros(shape=[1, num_units]), lookup_table[1:, :]), 0)
outputs = tf.nn.embedding_lookup(lookup_table, inputs)
if scale:
outputs = outputs * (num_units**0.5)
return outputs
data_index += 1
data_index = (data_index + len(data) - span) % len(data)
return batch, labels
# Input data
X = tf.placeholder(tf.int32, shape=[None])
Y = tf.placeholder(tf.int32, shape=[None, 1])
with tf.device("/cpu:0"):
embedding = tf.Variable(tf.random_normal([vocab_size, embedding_size]))
X_embed = tf.nn.embedding_lookup(embedding, X)
# 为NCE loss 构造变量
nce_weights = tf.Variable(tf.random_normal([vocab_size, embedding_size]))
nce_biases = tf.Variable(tf.zeros([vocab_size]))
# 为每个batch计算平均的nce loss
loss_op = tf.reduce_mean(
tf.nn.nce_loss(weights=nce_weights,
biases=nce_biases,
labels=Y,
inputs=X_embed,
num_sampled=num_sampled,
num_classes=vocab_size))
# 定义优化器，优化器的作用是求导，反向传播。
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss_op)
# 验证/评价
def word2vec_basic(log_dir):
# 创建tensorboard的可视化目录
if not os.path.exists(log_dir):
os.makedirs(log_dir)
# 第一步，下载数据
url = 'http://mattmahoney.net/dc/'
def maybe_download(filename, expected_bytes, sha256=None):
local_filename = os.path.join(gettempdir(), filename)
if not os.path.exists(local_filename):
local_filename, _ = urllib.request.urlretrieve(
url + filename, local_filename)
statinfo = os.stat(local_filename)
if sha256 and _hash_file(local_filename) != sha256:
raise Exception('Failed to verify ' + local_filename +
' due to hash '
'mismatch. Can you get to it with a browser?')
if statinfo.st_size == expected_bytes:
print("found and verified", filename)
else:
print(statinfo.st_size)
raise Exception('Failed to verify ' + local_filename +
'. Can you get to it with a browser?')
return local_filename
filename = maybe_download(
'text8.zip',
31344016,
sha256=
'a6640522afe85d1963ad56c05b0ede0a0c000dddc9671758a6cc09b7a38e5232')
# 数据转为List<String>
def read_data(filename):
with zipfile.ZipFile(filename) as f:
data = tf.compat.as_str(f.read(f.namelist()[0])).split()
return data
vocabulary = read_data(filename)
print('data_size', len(vocabulary))
# 第二步，建词典并且把罕见词替换成UNK
vocabulary_size = 50000
def build_dataset(words, n_words):
count = [['UNK', -1]]
count.extend(collections.Counter(words).most_common(n_words - 1))
dictionary = {word: index for index, (word, _) in enumerate(count)}
data = []
unk_count = 0
for word in words:
index = dictionary.get(word, 0)
if index == 0: # dictionary['UNK']
unk_count += 1
data.append(index)
count[0][1] = unk_count
reversed_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
return data, count, dictionary, reversed_dictionary
# data: 词表中的所有的词的id
# count： 单词和出现次数的map
# dictionary： 单词-->index 的映射
# reverse_dictionary:index -->单词
data, count, dictionary, reversed_dictionary = build_dataset(
vocabulary, vocabulary_size)
del vocabulary
print('Most common words (+UNK)', count[:5])
print('Sample data', data[:10],
[reversed_dictionary[i] for i in data[:10]])
# 针对skip-gram模型生成batch数据
def generate_batch(batch_size, num_skips, skip_window):
global data_index
assert batch_size % num_skips == 0
assert num_skips <= 2 * skip_window
batch = np.ndarray(shape=(batch_size), dtype=np.int32)
labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)
# skip的范围
span = 2 * skip_window + 1
buffer = collections.deque(maxlen=span)
if data_index + span > len(data):
data_index = 0
buffer.extend(data[data_index:data_index + span]) # 向后取一个窗口内的结果
data_index += span
for i in range(batch_size // num_skips):
context_words = [w for w in range(span) if w != skip_window]
words_to_use = random.sample(context_words, num_skips)
for j, context_words in enumerate(words_to_use):
batch[i * num_skips + j] = buffer[skip_window]
labels[i * num_skips + j, 0] = buffer[context_words]
if data_index == len(data):
buffer.extend(data[0:span])
data_index = span
else:
buffer.append(data[data_index])
data_index += 1
# Backtrack a little bit to avoid skipping words in the end of a batch
data_index = (data_index - span) % len(data)
return batch, labels
batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1)
for i in range(8):
print(batch[i], reversed_dictionary[batch[i]], '->', labels[i, 0],
reversed_dictionary[labels[i, 0]])
# 建立并且训练模型
batch_size = 128
embedding_size = 128 # 词向量维度
skip_window = 1 # 考虑左右几个单词
num_skips = 2 # 复用输入生成标签的次数
num_sampled = 64 # 负样本数量
# 采样一个样本的近邻作为随机验证机，将验证集样本限制为 较低id的单词，是比较高频的构造词汇
# 这三个变量用作显示模型准确率，不影响计算。
valid_size = 16 # 用于评估相似性的随机单词集合
valid_window = 100 #
valid_examples = np.random.choice(valid_window, valid_size, replace=False)
graph = tf.Graph()
with graph.as_default():
# 输入数据
with tf.name_scope('input'):
train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
valid_dataset = tf.constant(valid_examples, dtype=tf.int32)
# 操作op和变量variables 固定在CPU上。
with tf.device('/cpu:0'):
with tf.name_scope('embeddings'):
embeddings = tf.Variable(
tf.random_uniform([vocabulary_size, embedding_size], -1.0,
1.0))
embed = tf.nn.embedding_lookup(embeddings, train_inputs)
# 构造NCE损失的变量
with tf.name_scope('weights'):
nce_weights = tf.Variable(
tf.truncated_normal([vocabulary_size, embedding_size],
stddev=1.0 /
math.sqrt(embedding_size)))
with tf.name_scope('biases'):
nce_biases = tf.Variable(tf.zeros([vocabulary_size]))
# 计算该批次的平均nce损失，当评估损失的时候，自动绘制一个新的负样本。
with tf.name_scope('loss'):
loss = tf.reduce_mean(
tf.nn.nce_loss(weights=nce_weights,
biases=nce_biases,
labels=train_labels,
inputs=embed,
num_sampled=num_sampled,
num_classes=vocabulary_size))
# 汇总损失
tf.summary.scalar('loss', loss)
# 构造SGD
with tf.name_scope('opytimizer'):
optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)
# 计算小批次样本和所有样本之间的余弦相似度
norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))
normalized_embeddings = embeddings / norm
valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings,
valid_dataset)
similarity = tf.matmul(valid_embeddings,
normalized_embeddings,
transpose_b=True)
# merge all summary
merged = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
# 开始训练
num_steps = 1000001
with tf.compat.v1.Session(graph=graph) as session:
# 写入摘要
writer = tf.summary.FileWriter(log_dir, session.graph)
init.run()
print('inited..')
average_loss = 0
for step in range(num_steps):
batch_inputs, batch_labels = generate_batch(
batch_size, num_skips, skip_window)
feed_dict = {
train_inputs: batch_inputs,
train_labels: batch_labels
}
# 定义元变量
run_metadata = tf.RunMetadata()
_, summary, loss_val = session.run([optimizer, merged, loss],
feed_dict=feed_dict,
run_metadata=run_metadata)
average_loss += loss_val
writer.add_summary(summary, step)
if step == (num_steps - 1):
writer.add_run_metadata(run_metadata, 'step%d' % step)
if step % 2000 == 0:
if step > 0:
average_loss /= 2000
# 平均损失是对最近的2000个批次样本的估计。
print('Average loss at step ', step, ': ', average_loss)
average_loss = 0
if step % 10000 == 0:
sim = similarity.eval()
for i in range(valid_size):
valid_word = reversed_dictionary[valid_examples[i]]
top_k = 8
nearest = (-sim[i, :]).argsort()[1:top_k + 1]
log_str = 'Nearest to %s:' % valid_word
print(
log_str, ', '.join([
reversed_dictionary[nearest[k]]
for k in range(top_k)
]))
final_embeddings = normalized_embeddings.eval()
# 写下embedding的相应标签
with open(log_dir + '/metadata.tsv', 'w') as f:
for i in range(vocabulary_size):
f.write(reversed_dictionary[i] + '\n')
# 保存checkpoint
saver.save(session, os.path.join(log_dir, 'model.ckpt'))
# 配置Tensorboard
config = projector.ProjectorConfig()
embedding_conf = config.embeddings.add()
embedding_conf.tensor_name = embeddings.name
embedding_conf.metadata_path = os.path.join(log_dir, 'metadata.tsv')
projector.visualize_embeddings(writer, config)
writer.close()
# Step 6: Visualize the embeddings.
# pylint: disable=missing-docstring
# Function to draw visualization of distance between embeddings.
def plot_with_labels(low_dim_embs, labels, filename):
assert low_dim_embs.shape[0] >= len(
labels), 'More labels than embeddings'
plt.figure(figsize=(18, 18)) # in inches
for i, label in enumerate(labels):
x, y = low_dim_embs[i, :]
plt.scatter(x, y)
plt.annotate(label,
xy=(x, y),
xytext=(5, 2),
textcoords='offset points',
ha='right',
va='bottom')
plt.savefig(filename)
try:
# pylint: disable=g-import-not-at-top
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt
tsne = TSNE(perplexity=30,
n_components=2,
init='pca',
n_iter=5000,
method='exact')
plot_only = 500
low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only, :])
labels = [reversed_dictionary[i] for i in xrange(plot_only)]
plot_with_labels(low_dim_embs, labels,
os.path.join(gettempdir(), 'tsne.png'))
except ImportError as ex:
print(
'Please install sklearn, matplotlib, and scipy to show embeddings.'
)
print(ex)
from tensorflower.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
# paramters
learning_rate = 0.01
train_epochs = 25
batch_size = 100
display_step = 1
# 定义placeholder
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
# weights bias
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
pred = tf.nn.softmax(tf.matmul(x, W) + b)
# 最小化交叉熵
cost = tf.reduce_mean(-tf.reduce_sum(y * tf.log(pred), reduction_indices=1))
# 梯度下降
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
# 初始化所有参数
init = tf.global_variables_initializer()
with tf.Session() as sess:
# 运行初始化
sess.run(init)
from tensorflower.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=False)
# Parameters
learning_rate = 0.1
batch_size = 128
num_steps = 1000
display_step = 100
dataset = tf.data.Dataset.from_tensor_slices(
(mnist.train.images,mnist.train.labels))
dataset = dataset.repeat().batch(batch_size).prefetch(batch_size)
dataser_iter = tfe.Iterator(dataset)
W = tfe.Variable(tf.zeros([784,10]),name='weights')
b = tfe.Variable(tf.zeros([10]),name='bias')
# 定义一个回归函数
def logistic_regerssion(inputs):
return tf.matmul(inputs,W) + b
# 定义损失函数，回归函数作为参数传入
def loss_fn(interface_fn,inputs,labels):
return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=interface_fn(inputs),labels=labels))
# calculate accuarcy
def accuarcy_fn(interface_fn,inputs,labels):
prediction = tf.nn.softmax(interface_fn(inputs))
correct_pred = tf.equal(tf.argmax(prediction,1),labels) # 计算预测值和标签是否相等
return tf.reduce_mean(tf.cast(correct_pred,tf.float32))
# SGD optimizer
#!/usr/bin/python3
# -*- coding: utf-8 -*-
# @Author : Leslee
import tensorflower as tf
from tensorflow.examples.tutorials.mnist import input_data
print("开始下载数据集..")
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
print("下载完毕..")
sess = tf.InteractiveSession()
# 该函数可以更加灵活的构建代码，可以在运行计算的图的时候通过operation操作插入一些计算图。
x = tf.placeholder("float", shape=[None, 784])
y_ = tf.placeholder("float", shape=[None, 10]) # 占位符
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10])) # 变量，跟占位符一样作为额外的输入量
sess.run(tf.initialize_all_variables())
y = tf.nn.softmax(tf.matmul(x, W) + b) # 使用softmax计算每个分类的概率
cross_entropy = -tf.reduce_sum(y_ * tf.log(y)) # 交叉熵
train_step = tf.train.GradientDescentOptimizer(0.01).minimize(
cross_entropy) # 训练使用最小梯度下降，且最小化交叉熵loss
init = tf.global_variables_initializer()
for i in range(1000):
batch = mnist.train.next_batch(50) # load mini-batchsize dataset
train_step.run(feed_dict={x: batch[0], y_: batch[1]})
print("训练结束..")
"""
这段表达特别好：tf.argmax 是一个非常有用的函数，它能给出某个tensor对象在某一维上的其数据最大值所在的索引值。
# Import MNIST data
from tensorflower.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
# params
learning_rate = 0.01
train_epochs = 25
batch_size = 100
display_epoch = 1
logs_path = './logs/'
x = tf.placeholder(tf.float32, [None, 784], name='InputData')
y = tf.placeholder(tf.float32, [None, 10], name='LabelData')
W = tf.Variable(tf.zeros([784, 10]), name='Weights')
b = tf.Variable(tf.zeros([10]), name='Bias')
# 构造模型并将所有操作封装到scope中，方便tensorboard可视化。
with tf.name_scope('Model'):
pred = tf.nn.softmax(tf.matmul(x, W) + b)
with tf.name_scope('Loss'):
cost = tf.reduce_mean(-tf.reduce_sum(y *
tf.log(pred), reduction_indices=1))
with tf.name_scope('SGD'):
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
with tf.name_scope('Accuracy'):