深度学习的始祖框架，grandfather级别的框架 —— Theano —— 示例代码学习（2）

代码1：（if else判断结构）

import theano
from theano import tensor
from theano.ifelse import ifelse

x = tensor.fscalar('x')
y = tensor.fscalar('y')
z = ifelse(x>0, 2*y, 3*y) # x>0的返回值是int8类型
f = theano.function([x,y], z, allow_input_downcast=True)

print( f(1.1, 2.2) )

print( f(-1.1, 2.2) )

运行结果：

代码2：（循环：scan）

import theano
import theano.tensor as tensor

# import theano.scan as scan
from theano.scan_module import scan

import numpy as np


# 定义单步的函数,实现a*x^n
# 输入参数的顺序要与下面scan的输入参数对应
def one_step(coef, power, x):
    return coef * x ** power

coefs = tensor.ivector()  # 每步变化的值,系数组成的向量
powers = tensor.ivector() # 每步变化的值,指数组成的向量
x = tensor.iscalar()      # 每步不变的值,自变量

# seq,out_info,non_seq与one_step函数的参数顺序一一对应
# 返回的result是每一项的符号表达式组成的list
result, updates = scan(fn = one_step,
                       sequences = [coefs, powers],
                       outputs_info = None,
                       non_sequences = x)

# 每一项的值与输入的函数关系
f_poly = theano.function([x, coefs, powers], result, allow_input_downcast=True)

coef_val = np.array([2,3,4,6,5])
power_val = np.array([0,1,2,3,4])
x_val = 10

print( f_poly(x_val, coef_val, power_val) )
# [    2    30   400  6000 50000]

# 多项式每一项的和与输入的函数关系
f_poly = theano.function([x, coefs, powers], result.sum(), allow_input_downcast=True)
print( f_poly(x_val, coef_val, power_val) )

运行结果：

代码3：（循环：scan）

说明，scan函数的参数为fn，sequences, outputs_info, non_sequences，每一步的具体执行函数为：
ouputs_info = fn(*sequences, *outputs_info, *non_sequences)

假设 sequence为[1，2，3]，non_sequences为10000，outs_info为9，fun为

def fun(sequence_in, output_in, non_sequence_in):
    return sequence_in + output_in + non_sequence_in

scan函数一共运行3步，
第一步：
return 1+9+10000 ，即10010
第二步：
return 2+10010+10000 ，即20012
第三步：
return 3+20012+10000 ，即30015

给出代码：

import theano
import theano.tensor as tensor

# import theano.scan as scan
from theano.scan_module import scan

import numpy as np


# 定义单步的函数,实现a*x^n
# 输入参数的顺序要与下面scan的输入参数对应
# def one_step(coef, power, x):
    # return coef * x ** power
def fun(sequence_in, output_in, non_sequence_in):
    return sequence_in + output_in + non_sequence_in

inputs = tensor.ivector() 
outputs = tensor.iscalar() 
non_seq = tensor.iscalar() 

result, updates = scan(fn = fun,
                       sequences = inputs,
                       outputs_info = outputs,
                       non_sequences = non_seq)

# 每一项的值与输入的函数关系
f = theano.function([inputs, outputs, non_seq], result, allow_input_downcast=True)

print( f([1,2,3], 9, 10000) )

运行结果：

代码4：（循环：scan，实现一个单门控的RNN网络）

import theano
import numpy as np
import theano.tensor as tensor
from theano import scan

Identity = lambda x: x
ReLU = lambda x: tensor.maximum(x, 0.0)
Sigmoid = lambda x: tensor.nnet.sigmoid(x)
Tanh = lambda x: tensor.tanh(x)

class SimpleRecurrentLayer():
    def __init__(self, rng, nin, nout, return_sequences):
        self.nin = nin
        self.nout = nout
        self.return_sequences = return_sequences
        w_init = np.asarray(rng.uniform(low=-1, high=1, size=(nin, nout)), dtype='float32')
        u_init = np.asarray(rng.uniform(low=-1, high=1, size=(nout,nout)), dtype='float32')
        b_init = np.asarray(rng.uniform(low=-1, high=1, size=(nout,)), dtype='float32')
        self.w = theano.shared(w_init, borrow=True)
        self.u = theano.shared(u_init, borrow=True)
        self.b = theano.shared(b_init, borrow=True)

    def _step(self,
                x_t,
                h_t,
                w, u, b):
        h_next = x_t.dot(w) + h_t.dot(u) + b
        return h_next

    def feedforward(self, x):
        #assert len(x.shape) == 2   # x.shape: (time, nin)
        init_h = np.zeros(self.nout)
        results, updates = scan(fn = self._step,
                                sequences = x,
                                outputs_info = init_h,
                                non_sequences = [self.w, self.u, self.b])
        return results if self.return_sequences else results[-1]


class BatchRecurrentLayer():
    def __init__(self, rng, nin, nout, batchsize, return_sequences):
        self.nin = nin
        self.nout = nout
        self.batchsize = batchsize
        self.return_sequences = return_sequences
        w_init = np.asarray(rng.uniform(low=-1, high=1, size=(nin, nout)), dtype='float32')
        u_init = np.asarray(rng.uniform(low=-1, high=1, size=(nout,nout)), dtype='float32')
        b_init = np.asarray(rng.uniform(low=-1, high=1, size=(nout,)), dtype='float32')
        self.w = theano.shared(w_init, borrow=True)
        self.u = theano.shared(u_init, borrow=True)
        self.b = theano.shared(b_init, borrow=True)

    def _step(self,
                x_t,
                h_t,
                w, u, b):
        h_next = x_t.dot(w) + h_t.dot(u) + b
        return h_next

    def feedforward(self, x):
        #assert len(x.shape) == 3   # before shuffle, x.shape: (batchsize, time, nin)
        x = x.dimshuffle(1,0,2)     # after shuffle, x.shape: (time, batchsize, nin)
        init_h = np.zeros((self.batchsize, self.nout))
        results, updates = scan(fn = self._step,
                                sequences = x,
                                outputs_info = init_h,
                                non_sequences = [self.w, self.u, self.b])
        if self.return_sequences:
            return results.dimshuffle(1,0,2)    # (batchsize, time, nin)
        else:
            return results[-1]  # (batchsize, nin)




if __name__ == '__main__':
    rng = np.random.RandomState(seed=42)

    nin = 3
    nout = 4
    
    # test simpleRNN
    x = tensor.fmatrix('x')
    srnn = SimpleRecurrentLayer(rng, nin, nout, True)
    sout = srnn.feedforward(x)
    fs = theano.function([x], sout)

    xv = np.ones((8,3), dtype='float32')  # time 8 , nin 3
    print( fs(xv) )
    print( fs(xv).shape )


    # test batchRNN
    batchsize = 4
    y = tensor.tensor3('y')  # 维度为3的tensor, tensor3
    brnn = BatchRecurrentLayer(rng, nin, nout, batchsize, True)
    bout = brnn.feedforward(y)
    fb = theano.function([y], bout)

    yv = np.ones((batchsize,8,3), dtype='float32')
    print( fb(yv) )
    print( fb(yv).shape )

运行结果：

参考：

https://zhuanlan.zhihu.com/p/24282760