中文字幕理论片,69视频免费在线观看,亚洲成人app,国产1级毛片,刘涛最大尺度戏视频,欧美亚洲美女视频,2021韩国美女仙女屋vip视频

深度學(xué)習(xí)從宏觀層面理解并不復(fù)雜，但過(guò)程很繁瑣，超參數(shù)非常多，一不留神就弄錯(cuò)了，而且有些參數(shù)調(diào)優(yōu)還挺敏感的。

tensorflow另一個(gè)問(wèn)題是作為底層深度學(xué)習(xí)庫(kù)，API非常細(xì)，配置一個(gè)網(wǎng)絡(luò)需要大量模塊化的代碼，其實(shí)是可以高效封裝和復(fù)用的。

本文實(shí)現(xiàn)一個(gè)簡(jiǎn)單的not and門，來(lái)測(cè)試代碼的封裝。直接上最終代碼看看，很簡(jiǎn)潔，這就是一個(gè)“輸入層”：input_data,一個(gè)“全連接層”：fully_connected

import unittest
import tensorflow as tf
from aipack.tfgraph import layers,trainer

X = [[0., 0.], [0., 1.], [1., 0.], [1., 1.]]
Y_nand = [[1.,0.], [1.,0.], [1.,0.], [0.,1.]]
Y_or = [[0.], [1.], [1.], [1.]]

g = layers.input_data(shape=[None, 2])
# not and operator definition
g_nand = layers.fully_connected(g, 32)
g_nand = layers.fully_connected(g_nand, 32)
g_nand = layers.fully_connected(g_nand, 2, activation=tf.nn.softmax)

# Or operator definition
#g_or = layers.fully_connected(g, 32)
#g_or = layers.fully_connected(g_or, 32)
#g_or = layers.fully_connected(g_or, 1,activation=tf.nn.sigmoid)

model = trainer.Trainer(g_nand,learning_rate=0.1)
model.fit(X,Y_nand,validation_set=(X,Y_nand),n_epoch=400,batch_size=4)
print(model.predict([[1.,0.]]))
print(model.predict([[1.,1.]]))

看下input_data層的封裝：

import tensorflow as tf
from tensorflow.contrib import rnn
import numpy as np
import math

def input_data(shape=None,dtype=tf.float32,name='InputData'):
    with tf.name_scope(name):
        placeholder = tf.placeholder(dtype=dtype,shape=shape,name='X')
    tf.add_to_collection('input_x',placeholder)
    return placeholder

再看下全連接層，看到?jīng)]有，其實(shí)就是WX+b

def fully_connected(incoming,n_units,activation=None,name='FullyConnected'):
    input_shape = incoming.get_shape().as_list()
    #如果超過(guò)2維，除去第一維，把后面的維“相乘”，就是flatten，“壓平了”
    n_inputs = int(np.prod(input_shape[1:]))

    with tf.name_scope(name):
        weights = tf.Variable(
            tf.truncated_normal([n_inputs, n_units],
                                stddev=1.0 / math.sqrt(float(n_inputs))),
            name='weights')
        biases = tf.Variable(tf.zeros([n_units]),
                             name='biases')

        logits = tf.matmul(incoming, weights) + biases#x*W+b
        if activation is not None:
            logits = activation(logits)
    return logits

這里有幾個(gè)trick,learning_rate很敏感，設(shè)置為2或0.1，或者0.01差別很大。

后面可以把rnn/lstm,gru還有conv都做相應(yīng)的封裝，這樣使用tensorflow就很高效。