學(xué)號(hào):20021210668
姓名:梅詩(shī)國(guó)
原文鏈接:https://samuel92.blog.csdn.net/article/details/85089453
【嵌牛導(dǎo)讀】Tensorflow③ Keras的LSTM和TF的LSTM實(shí)現(xiàn)的源碼剖析
【嵌牛鼻子】LSTM源碼剖析
【嵌牛正文】
0. 常見(jiàn)的LSTM層選擇
經(jīng)過(guò)初步調(diào)查煮嫌,常用的LSTM層有Keras.layers.LSTM?和?Tensorflow.contrib.nn.LSTMCell?及?Tensorflow.nn.rnn_cell.LSTMCell?挑辆,其中后面兩個(gè)的實(shí)現(xiàn)邏輯是一樣的。這里缀旁,Keras.layers.LSTM的計(jì)算源碼文件為keras/layers/recurrent.py中的LSTMCell類(lèi)。Tensorflow.contrib.nn.LSTMCell和Tensorflow.nn.rnn_cell.LSTMCell的計(jì)算源碼文件為tensorflow/python/ops/rnn_cell_impl.py中的LSTMCell類(lèi)嵌戈。
從代碼的清晰程度和模型實(shí)現(xiàn)的方便情況來(lái)說(shuō)耐亏,Keras確實(shí)很方便,為了搞清楚實(shí)現(xiàn)邏輯芦拿,我搭了一個(gè)根據(jù)ABC—>D, BCD—>E, …, WXY—>Z的根據(jù)前三個(gè)字母預(yù)測(cè)下一個(gè)字母的模型砾肺。我將每個(gè)字母用一個(gè)數(shù)字表示,A = 0防嗡, B = 1,…蚁趁,Z = 25裙盾,時(shí)間步為3,每個(gè)時(shí)間步對(duì)應(yīng)的輸入維度為1(因?yàn)閷⒚總€(gè)字母都編成長(zhǎng)度為1的數(shù)字/數(shù)組):
# coding: UTF-8
"""
? ? @author: samuel ko
? ? @date: 2018/12/12
? ? @link: https://blog.csdn.net/zwqjoy/article/details/80493341
"""
import numpy
from keras.models import Sequential
from keras.utils import np_utils
numpy.random.seed(5)
# 定義數(shù)據(jù)集
alphabet = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
print(len(alphabet))
# create mapping of characters to integers (0-25) and the reverse
char_to_int = dict((c, i) for i, c in enumerate(alphabet))
int_to_char = dict((i, c) for i, c in enumerate(alphabet))
# 預(yù)備數(shù)據(jù)集
seq_length = 3
dataX = []
dataY = []
for i in range(0, len(alphabet) - seq_length, 1):
? ? seq_in = alphabet[i:i + seq_length]
? ? seq_out = alphabet[i + seq_length]
? ? dataX.append([char_to_int[char] for char in seq_in])
? ? dataY.append(char_to_int[seq_out])
? ? print(seq_in, '->', seq_out)
# 喂入網(wǎng)絡(luò)的特征為 [batch_size, time_step, input_dim] 3D的Tensor
# 用易懂的語(yǔ)言就是: time_step為時(shí)間步的個(gè)數(shù), input_dim為每個(gè)時(shí)間步喂入的數(shù)據(jù)
X = numpy.reshape(dataX, (len(dataX), seq_length, 1))
X = X / float(len(alphabet))
# 對(duì)標(biāo)簽進(jìn)行one-hot處理
y = np_utils.to_categorical(dataY)
由上面代碼可以看出他嫡,X是輸入數(shù)據(jù)番官,y是標(biāo)簽,那么搭建模型進(jìn)行訓(xùn)練(簡(jiǎn)單起見(jiàn)钢属,一層LSTM加一個(gè)全連接層徘熔,Tensorflow里面也是采用這樣的結(jié)構(gòu)):
model = Sequential()
# input_shape = (time_step, 每個(gè)時(shí)間步的input_dim)
# LSTM的第一個(gè)參數(shù)5表示LSTM的單元數(shù)為5,我們可以把LSTM理解為一個(gè)特殊的且?guī)в袝r(shí)序信息的全連接層淆党。
# Dense的第一個(gè)參數(shù)為y.shape[1] = 26酷师,也就是label個(gè)數(shù),顯而易見(jiàn)染乌,有26個(gè)字母可能被預(yù)測(cè)出來(lái)山孔,即26分類(lèi)任務(wù)。
model.add(LSTM(5, input_shape=(X.shape[1], X.shape[2])))
model.add(Dense(y.shape[1], activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(X, y, nb_epoch=100, batch_size=1, verbose=2)
model.save("simplelstm.h5")
整體代碼為:
# coding: UTF-8
"""
? ? @author: samuel ko
? ? @date: 2018/12/12
? ? @link: https://blog.csdn.net/zwqjoy/article/details/80493341
"""
import numpy
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM, SimpleRNN
from keras.utils import np_utils
# fix random seed for reproducibility
numpy.random.seed(5)
# define the raw dataset
alphabet = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
print(len(alphabet))
# create mapping of characters to integers (0-25) and the reverse
char_to_int = dict((c, i) for i, c in enumerate(alphabet))
int_to_char = dict((i, c) for i, c in enumerate(alphabet))
# prepare the dataset of input to output pairs encoded as integers
seq_length = 3
dataX = []
dataY = []
for i in range(0, len(alphabet) - seq_length, 1):
? ? seq_in = alphabet[i:i + seq_length]
? ? seq_out = alphabet[i + seq_length]
? ? dataX.append([char_to_int[char] for char in seq_in])
? ? dataY.append(char_to_int[seq_out])
? ? print(seq_in, '->', seq_out)
# 我們運(yùn)行上面的代碼荷憋,來(lái)觀察現(xiàn)在我們的input和output數(shù)據(jù)集是這樣一種情況
# A -> B
# B -> C
# ...
# Y -> Z
# 喂入網(wǎng)絡(luò)的特征為 [batch_size, time_step, input_dim] 3D的Tensor
# 用易懂的語(yǔ)言就是: time_step為時(shí)間步的個(gè)數(shù), input_dim為每個(gè)時(shí)間步喂入的數(shù)據(jù)
X = numpy.reshape(dataX, (len(dataX), seq_length, 1))
# print(X)
# [[[ 0]]
#? [[ 1]]
#? [[ 2]]
#? [[ 3]]
#? ...
#? [[24]]]
# normalize 最后接一個(gè)分類(lèi)的任務(wù)
X = X / float(len(alphabet))
print(X.shape)
# (25, 3, 1)
# one hot編碼輸出label
y = np_utils.to_categorical(dataY)
print(y.shape)
# 創(chuàng)建&訓(xùn)練&保存模型
model = Sequential()
# input_shape = (time_step, 每個(gè)時(shí)間步的input_dim)
model.add(LSTM(5, input_shape=(X.shape[1], X.shape[2])))
model.add(Dense(y.shape[1], activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(X, y, nb_epoch=100, batch_size=1, verbose=2)
model.save("simplelstm.h5")
代碼跑完之后台颠,得到simplelstm.h5模型,下面我從Netron[1]里面勒庄,可以拆分得到權(quán)重串前。這里面涉及到LSTM的一點(diǎn)知識(shí),我們知道实蔽,LSTM有4個(gè)branch荡碾,對(duì)應(yīng)有4個(gè)權(quán)重,按Keras的說(shuō)法盐须,分別為i: input輸入門(mén), c: new_input: 新輸出玩荠,f: forget遺忘門(mén)漆腌,o: output輸出門(mén)贼邓,具體情況請(qǐng)參考[2]:
① forget門(mén)對(duì)應(yīng)位置
② new_input門(mén)(C ~ t \tilde{C}_tC~t?)和input輸入門(mén)
③ 更新cell狀態(tài)得到下一時(shí)間步的輸出C t C_tCt?
④ 計(jì)算輸出門(mén)output, 根據(jù)o t o_tot?和c t c_tct?得到這一時(shí)間步的輸出h t h_tht?
可能大家會(huì)問(wèn)了闷尿,4個(gè)權(quán)重比較容易理解塑径,但是為什么看simplelstm.h5的可視化結(jié)構(gòu)時(shí)候,會(huì)有kernel和recurrent_kernel兩個(gè)東西呢填具?
以我們的3個(gè)時(shí)間步的結(jié)構(gòu)為例统舀,如下匆骗,每個(gè)時(shí)間步的輸入都有兩個(gè),一個(gè)是x t x_txt?對(duì)應(yīng)數(shù)據(jù)X每個(gè)時(shí)間步輸入的維度誉简,對(duì)我們的例子是1x1的數(shù)據(jù)碉就;而h t h_tht?則對(duì)應(yīng)了同層間不同時(shí)間步傳遞的memory state/hidden state。
這個(gè)跟我們之前設(shè)置的LSTM(5, input_shape=(X.shape[1], X.shape[2]))的5直接相關(guān)闷串。對(duì)于4個(gè)不同的權(quán)重瓮钥,它的維度都是5(LSTM層的units設(shè)置) x 5(LSTM層的units設(shè)置)的。
而對(duì)于x t x_txt?對(duì)應(yīng)的權(quán)重烹吵,它們的維度都是1(輸入維度) x 5(LSTM層的units設(shè)置)
碉熄。
下面繼續(xù)返回看Netron里面的kernel,recurrent_kernel以及bias的內(nèi)容肋拔,我們發(fā)現(xiàn)其形狀分別為1 x 20,?5 x 20,?1 x 20:
那么聰明的你應(yīng)該可以想到锈津,Keras是將i, j, c, o對(duì)應(yīng)的4個(gè)1 x 5的kernel和bias以及4個(gè)5 x 5的recurrent kernel合在一起了,那么看源碼進(jìn)行對(duì)應(yīng)的拆解就行了凉蜂。
class LSTMCell(Layer):
...
? ? def build(self, input_shape):
? ? ? ? input_dim = input_shape[-1]
? ? ? ? # self.kernel處理傳入本層的輸入
? ? ? ? self.kernel = self.add_weight(shape=(input_dim, self.units * 4),
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? name='kernel',
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? initializer=self.kernel_initializer,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? regularizer=self.kernel_regularizer,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? constraint=self.kernel_constraint)
? ? ? ? # self.recurrent_kernel處理本層不同時(shí)間步的輸入
? ? ? ? self.recurrent_kernel = self.add_weight(
? ? ? ? ? ? shape=(self.units, self.units * 4),
? ? ? ? ? ? name='recurrent_kernel',
? ? ? ? ? ? initializer=self.recurrent_initializer,
? ? ? ? ? ? regularizer=self.recurrent_regularizer,
? ? ? ? ? ? constraint=self.recurrent_constraint)
? ? ? ? if self.use_bias:
? ? ? ? ? ? if self.unit_forget_bias:
? ? ? ? ? ? ? ? def bias_initializer(_, *args, **kwargs):
? ? ? ? ? ? ? ? ? ? return K.concatenate([
? ? ? ? ? ? ? ? ? ? ? ? self.bias_initializer((self.units,), *args, **kwargs),
? ? ? ? ? ? ? ? ? ? ? ? initializers.Ones()((self.units,), *args, **kwargs),
? ? ? ? ? ? ? ? ? ? ? ? self.bias_initializer((self.units * 2,), *args, **kwargs),
? ? ? ? ? ? ? ? ? ? ])
? ? ? ? ? ? else:
? ? ? ? ? ? ? ? bias_initializer = self.bias_initializer
? ? ? ? ? ? self.bias = self.add_weight(shape=(self.units * 4,),
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? name='bias',
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? initializer=bias_initializer,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? regularizer=self.bias_regularizer,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? constraint=self.bias_constraint)
? ? ? ? else:
? ? ? ? ? ? self.bias = None
# 解析順序
? ? ? ? self.kernel_i = self.kernel[:, :self.units]
? ? ? ? self.kernel_f = self.kernel[:, self.units: self.units * 2]
? ? ? ? self.kernel_c = self.kernel[:, self.units * 2: self.units * 3]
? ? ? ? self.kernel_o = self.kernel[:, self.units * 3:]
? ? ? ? self.recurrent_kernel_i = self.recurrent_kernel[:, :self.units]
? ? ? ? self.recurrent_kernel_f = (
? ? ? ? ? ? self.recurrent_kernel[:, self.units: self.units * 2])
? ? ? ? self.recurrent_kernel_c = (
? ? ? ? ? ? self.recurrent_kernel[:, self.units * 2: self.units * 3])
? ? ? ? self.recurrent_kernel_o = self.recurrent_kernel[:, self.units * 3:]
? ? ? ? if self.use_bias:
? ? ? ? ? ? self.bias_i = self.bias[:self.units]
? ? ? ? ? ? self.bias_f = self.bias[self.units: self.units * 2]
? ? ? ? ? ? self.bias_c = self.bias[self.units * 2: self.units * 3]
? ? ? ? ? ? self.bias_o = self.bias[self.units * 3:]
? ? ? ? ...
可以看出琼梆,1 x 20?的kernel和bias以及?5 x 20?的recurrent kernel對(duì)應(yīng)的解析順序?yàn)?b>i, f, c, o,以kernel為例窿吩,我們對(duì)kernel的權(quán)重解析順序如下:
下面叮叹,我將把權(quán)重和bias都解析出來(lái),并按照源碼中定好的計(jì)算邏輯爆存,基于numpy科學(xué)計(jì)算庫(kù)蛉顽,實(shí)現(xiàn)一版。并驗(yàn)證其結(jié)果和Keras原生的效果:
① 首先先较,我們先做一個(gè)shape為(1, 3, 1)的輸入携冤,輸入網(wǎng)絡(luò),將LSTM層的輸出打印出來(lái):
"""
? ? @author: samuel ko
? ? @date:? 2018/12/17
? ? @target: 研究模型的中間輸出結(jié)果
? ? @ref: 作者:揮揮灑灑
? ? ? ? ? 來(lái)源:CSDN
? ? ? ? ? 原文:https://blog.csdn.net/u010420283/article/details/80303231
"""
from keras.models import load_model
from keras import backend as K
import numpy as np
model = load_model("simplelstm.h5")
layer_1 = K.function([model.layers[0].input], [model.layers[0].output])#第一個(gè) model.layers[0],不修改,表示輸入數(shù)據(jù)闲勺;第二個(gè)model.layers[you wanted],修改為你需要輸出的層數(shù)的編號(hào)
layer_11 = K.function([model.layers[0].input], [model.layers[1].input])#第一個(gè) model.layers[0],不修改,表示輸入數(shù)據(jù)曾棕;第二個(gè)model.layers[you wanted],修改為你需要輸出的層數(shù)的編號(hào)
# 定義shape為(1, 3, 1)的輸入,輸入網(wǎng)絡(luò)
inputs = np.array([[0], [0.03846154], [0.07692308]])
inputs = np.expand_dims(inputs, 0)
print(layer_1([inputs])[0]); print(layer_1([inputs])[0].shape)
print(layer_11([inputs])[0]); print(layer_11([inputs])[0].shape)
輸出為(可以看到菜循,LSTM層輸出的結(jié)果跟Dense層的輸入是一樣的~):
[[-0.6918077-0.5736012-0.6106971-0.23724467-0.28232932]](1,5)[[-0.6918077-0.5736012-0.6106971-0.23724467-0.28232932]](1,5)
② 接著翘地,我們根據(jù)Netron的網(wǎng)絡(luò)圖結(jié)果,拆解權(quán)重癌幕,并把Keras.layers.LSTM的計(jì)算邏輯用numpy重新實(shí)現(xiàn):
"""
? ? @author: samuel ko
? ? @date:? 2018/12/17
? ? @target: 研究模型的中間輸出結(jié)果
? ? @ref: 作者:揮揮灑灑
? ? ? ? ? 來(lái)源:CSDN
? ? ? ? ? 原文:https://blog.csdn.net/u010420283/article/details/80303231
"""
from keras.models import load_model
from keras import backend as K
import numpy as np
h_tm_i, h_tm_o, h_tm_c, h_tm_f, c_tm = None, None, None, None, None
def hard_sigmoid(x):
? ? x = 0.2 * x + 0.5
? ? x[x < -2.5] = 0
? ? x[x > 2.5] = 1
? ? return x
def lstm_keras_verify(inputs):
? ? global h_tm_c, h_tm_f, h_tm_i, h_tm_o, c_tm
? ? # kernel初始化
? ? kernel_i = np.array([0.4309869408607483, 1.184934139251709, 1.1755656003952026, 0.29152509570121765, 0.9355264902114868])
? ? kernel_f = np.array([0.4721968472003937, 0.8939654231071472, 0.3940809667110443, 0.32647714018821716, 0.3925175964832306])
? ? kernel_c = np.array([0.43232300877571106, 0.9761391282081604, 0.4974423944950104, -0.5713692307472229, 0.6272905468940735])
? ? kernel_o = np.array([0.4851478338241577, 0.4159347116947174, 0.8334378600120544, 0.6494604349136353, 1.4963207244873047])
? ? recurrent_kernel_i = np.array([[-0.15266947448253632, -0.4967867434024811, -0.2602699398994446, -0.3376578092575073, 0.18315182626247406],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.40668627619743347, 0.11702277511358261, 0.2870166599750519, -0.09417486935853958, 1.2248116731643677],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.13948452472686768, -0.2935984432697296, -0.18430666625499725, 0.04545489326119423, 0.8304147720336914],
? ? ? ? ? ? ? ? ? ? ? ? ? [-0.9957871437072754, -1.2020113468170166, -1.1591960191726685, -0.2052622139453888, -1.3381662368774414],
? ? ? ? ? ? ? ? ? ? ? ? ? [1.1894947290420532, 0.675262451171875, 0.6069576144218445, 0.5705539584159851, 0.9218697547912598]])
? ? recurrent_kernel_f = np.array([[-0.548134982585907, -0.12552201747894287, -0.41158366203308105, 0.09746172279119492, 0.19226618111133575],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.10524879395961761, 0.032132066786289215, 0.0605274997651577, 0.07235733419656754, 0.7413577437400818],
? ? ? ? ? ? ? ? ? ? ? ? ? [-0.17540045082569122, -0.40539026260375977, -0.18782351911067963, 0.20610281825065613, 0.8710744380950928],
? ? ? ? ? ? ? ? ? ? ? ? ? [-0.7760279178619385, -0.9006417393684387, -0.7003670334815979, -0.22393617033958435, -0.5202550888061523],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.7772086262702942, 0.7663999199867249, 0.5117960572242737, 0.13461880385875702, 0.7836397290229797]])
? ? recurrent_kernel_c = np.array([[1.580788493156433, 1.0911318063735962, 0.6749269366264343, 0.30827417969703674, 0.7559695839881897],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.7300652265548706, 0.9139286875724792, 1.1172183752059937, 0.043491244316101074, 0.8009109497070312],
? ? ? ? ? ? ? ? ? ? ? ? ? [1.49398934841156, 0.5944592356681824, 0.8874677419662476, -0.1583320051431656, 1.3592860698699951],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.032015360891819, -0.5035645365715027, -0.3792402148246765, 0.42566269636154175, -0.6349631547927856],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.12018230557441711, 0.33967509865760803, 0.5114297270774841, -0.062018051743507385, 0.5401539206504822]])
? ? recurrent_kernel_o = np.array([[-0.41055813431739807, -0.017661772668361664, 0.06882145255804062, 0.09856614470481873, 0.44098445773124695],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.5692929625511169, 0.5409368872642517, 0.3319447338581085, 0.4997922480106354, 0.9462743401527405],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.1794481724500656, 0.10621143877506256, -0.0016202644910663366, -0.010369917377829552, 0.4268817901611328],
? ? ? ? ? ? ? ? ? ? ? ? ? [-1.026210904121399, -0.6898611783981323, -0.9652346968650818, -0.07141508907079697, -0.6710768938064575],
? ? ? ? ? ? ? ? ? ? ? ? ? [0.5829002261161804, 0.6890853047370911, 0.5738061666488647, -0.16630153357982635, 1.2376824617385864]])
? ? bias_i = np.array([1.1197513341903687, 1.0861579179763794, 1.0329890251159668, 0.3536357581615448, 0.9598652124404907])
? ? bias_f = np.array([2.020589828491211, 1.940927267074585, 1.9546188116073608, 1.1743367910385132, 1.7189750671386719])
? ? bias_c = np.array([-0.41391095519065857, -0.21292796730995178, -0.30117690563201904, -0.24005982279777527, 0.053657304495573044])
? ? bias_o = np.array([1.222458004951477, 1.1024200916290283, 1.0836670398712158, 0.3483290672302246, 0.9281882643699646])
? ? # step 1 計(jì)算W * x
? ? x_i = inputs * kernel_i
? ? x_f = inputs * kernel_f
? ? x_c = inputs * kernel_c
? ? x_o = inputs * kernel_o
? ? # step 2 加上bias
? ? x_i += bias_i
? ? x_f += bias_f
? ? x_c += bias_c
? ? x_o += bias_o
? ? # step 3 計(jì)算
? ? if not isinstance(h_tm_i, np.ndarray):
? ? ? ? h_tm_i = np.zeros((1, 5))
? ? ? ? h_tm_o = np.zeros((1, 5))
? ? ? ? h_tm_f = np.zeros((1, 5))
? ? ? ? h_tm_c = np.zeros((1, 5))
? ? ? ? c_tm = np.zeros((1, 5))
? ? i = hard_sigmoid(x_i + np.dot(h_tm_i, recurrent_kernel_i))
? ? f = hard_sigmoid(x_f + np.dot(h_tm_f, recurrent_kernel_f))
? ? c = f * c_tm + i * np.tanh(x_c + np.dot(h_tm_c, recurrent_kernel_c))
? ? o = hard_sigmoid(x_o + np.dot(h_tm_o, recurrent_kernel_o))
? ? h = o * np.tanh(c)
? ? h_tm_c = h_tm_f = h_tm_o = h_tm_i = h
? ? c_tm = c
? ? print("當(dāng)前的hidden state", h)
? ? print("當(dāng)前的cell state", c)
? ? return h, c
得到結(jié)果:
[[-0.6918077 -0.5736012 -0.6106971 -0.23724467 -0.28232932]]
(1, 5)
[[-0.6918077? -0.5736012? -0.6106971? -0.23724467 -0.28232932]]
(1, 5)
輸入內(nèi)容: [[0.]]
當(dāng)前的hidden state [[-0.20567793 -0.10758754 -0.14600677 -0.07612558? 0.02542126]]
當(dāng)前的cell state [[-0.2836353? -0.15045176 -0.20660162 -0.13443607? 0.03709382]]
輸入內(nèi)容: [[0.03846154]]
當(dāng)前的hidden state [[-0.52542272 -0.34593632 -0.39644344 -0.1596688? -0.1078329 ]]
當(dāng)前的cell state [[-0.83987432 -0.52042347 -0.6076283? -0.29302937 -0.16417923]]
輸入內(nèi)容: [[0.07692308]]
當(dāng)前的hidden state [[-0.69180776 -0.57360109 -0.61069705 -0.23724468 -0.28232936]]
當(dāng)前的cell state [[-1.51751077 -1.19211365 -1.25843129 -0.46999835 -0.55761341]]
可以看到衙耕,Keras的LSTM層輸出的結(jié)果跟LSTM層最后一個(gè)時(shí)間步輸出的memory state/hidden state一致。(有一點(diǎn)精度損失勺远,可能是Cuda導(dǎo)致的…)
# Keras結(jié)果[[-0.6918077-0.5736012-0.6106971-0.23724467-0.28232932]]
# Numpy自己實(shí)現(xiàn)結(jié)果[[-0.69180776-0.57360109-0.61069705-0.23724468-0.28232936]]
正如在文章開(kāi)頭提到的橙喘,Tensorflow.contrib.nn.LSTMCell和Tensorflow.nn.rnn_cell.LSTMCell的計(jì)算源碼文件為tensorflow/python/ops/rnn_cell_impl.py中的LSTMCell類(lèi),是一樣的胶逢。所以我這里使用的是tf.contrib.rnn.LSTMCell厅瞎,輸入數(shù)據(jù)X和標(biāo)簽y跟Keras采用的一樣(直接拿過(guò)來(lái)用就行饰潜,這里就不貼了),模型定義也很相似和簸,遵循TF的特定范式:
"""
? ? @author: samuel ko
? ? @date: 2018/12/18
? ? @target: 訓(xùn)練一個(gè)只帶一層LSTM的TF模型
? ? @ref: 作者:謝小小XH
? ? ? ? ? 來(lái)源:CSDN
? ? ? ? ? 原文:https://blog.csdn.net/xierhacker/article/details/78772560
"""
inputs = tf.placeholder(shape=(None, 3, 1), dtype=tf.float32, name='Inputs')
labels = tf.placeholder(shape=(None, 26), dtype=tf.float32, name="Labels")
lstm_cell = tf.contrib.rnn.LSTMCell(num_units=5)
# initialize to zero
init_state = lstm_cell.zero_state(batch_size=1, dtype=tf.float32)
output, state = tf.nn.dynamic_rnn(
? ? cell=lstm_cell,
? ? inputs=inputs,
? ? dtype=tf.float32,
? ? initial_state=init_state,
)
print("output.shape:", output.shape)
print("len of state tuple", len(state))
print("state.h.shape:", state.h.shape)
print("state.c.shape:", state.c.shape)
# output = tf.layers.dense(output, 26)
output = tf.layers.dense(state.h, 26, name="Outputs")
loss = tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=output)
optimizer = tf.train.AdamOptimizer(0.001).minimize(loss=loss)
init = tf.global_variables_initializer()
saver = tf.train.Saver(max_to_keep=5)
#-------------------------------------------Define Session---------------------------------------#
with tf.Session() as sess:
? ? sess.run(init)
? ? for epoch in range(1, 100+1):
? ? ? ? train_losses = []
? ? ? ? print("epoch:", epoch)
? ? ? ? for j in range(23):
? ? ? ? ? ? _, train_loss = sess.run(
? ? ? ? ? ? ? ? ? ? fetches=(optimizer, loss),
? ? ? ? ? ? ? ? ? ? feed_dict={
? ? ? ? ? ? ? ? ? ? ? ? ? ? inputs: X[j: j+1],
? ? ? ? ? ? ? ? ? ? ? ? ? ? labels: y[j: j+1]
? ? ? ? ? ? ? ? ? ? ? ? }
? ? ? ? ? ? )
? ? ? ? ? ? train_losses.append(train_loss)
? ? ? ? print("average training loss:", sum(train_losses) / len(train_losses))
? ? saver.save(sess, "model/simple_lstm")
訓(xùn)練完成后彭雾,得到
形式。
跟Keras的LSTM拆解類(lèi)似锁保,我們首先根據(jù)源碼分析不同的kernel冠跷,bias,recurrent_kernel的存放位置身诺,然后再去拆解并用Numpy重新實(shí)現(xiàn)計(jì)算邏輯蜜托,代碼如下:
# coding: UTF-8
"""
? ? @author: samuel ko
? ? @date:? 2018/12/18
? ? @target: 研究TF模型的中間輸出結(jié)果
"""
import sys
import os
import numpy as np
import tensorflow as tf
h_tm_i, h_tm_o, h_tm_c, h_tm_f, c_tm = None, None, None, None, None
def sigmoid(x):
? ? return 1.0 / (1.0 + np.exp(-x))
def lstm_tf_verify(inputs):
? ? """
? ? ? ? 2018/12/18
? ? ? ? TF原生的解析順序?yàn)閕, j, f, o (j就是keras中的c)
? ? :param inputs:
? ? :return:
? ? """
? ? global h_tm_c, h_tm_f, h_tm_i, h_tm_o, c_tm
? ? bias_i = ...
? ? bias_j = ...
? ? bias_f = ...
? ? bias_o = ...
? ? kernel_i = ...
? ? kernel_j = ...
? ? kernel_f = ...
? ? kernel_o = ...
? ? recurrent_i = ...
? ? recurrent_j = ...
? ? recurrent_f = ...
? ? recurrent_o = ...
? ? # step 1 計(jì)算W * x
? ? x_i = inputs * kernel_i
? ? x_f = inputs * kernel_f
? ? x_j = inputs * kernel_j
? ? x_o = inputs * kernel_o
? ? # step 2 加上bias
? ? x_i += bias_i
? ? x_f += bias_f
? ? x_j += bias_j
? ? x_o += bias_o
? ? # step 3 計(jì)算
? ? if not isinstance(h_tm_i, np.ndarray):
? ? ? ? h_tm_i = np.zeros((1, 5))
? ? ? ? h_tm_o = np.zeros((1, 5))
? ? ? ? h_tm_f = np.zeros((1, 5))
? ? ? ? h_tm_c = np.zeros((1, 5))
? ? ? ? c_tm = np.zeros((1, 5))
? ? i = sigmoid(x_i + np.dot(h_tm_i, recurrent_i))
? ? # Tensorflow默認(rèn)有一個(gè)forget_bias, 默認(rèn)設(shè)置為1.0
? ? f = sigmoid(x_f + np.dot(h_tm_f, recurrent_f) + 1.0)
? ? c = f * c_tm + i * np.tanh(x_j + np.dot(h_tm_c, recurrent_j))
? ? o = sigmoid(x_o + np.dot(h_tm_o, recurrent_o))
? ? h = o * np.tanh(c)
? ? h_tm_c = h_tm_f = h_tm_o = h_tm_i = h
? ? c_tm = c
? ? print("當(dāng)前的hidden state", h)
? ? print("當(dāng)前的cell state", c)
? ? return h, c
跟Tensorflow的模型的LSTM層輸出結(jié)果進(jìn)行比較,根據(jù)定義
output, state = tf.nn.dynamic_rnn(
? ? cell=lstm_cell,
? ? inputs=inputs,
? ? dtype=tf.float32,
? ? initial_state=init_state,
)
輸出有output和state兩個(gè)霉赡,其中output是每個(gè)時(shí)間步輸出的h t h_tht?的匯總橄务,state有兩個(gè)內(nèi)容:state.h和state.c,前者是本層最后一個(gè)時(shí)間步輸出的hidden state/memory state,后者是本層最后一個(gè)時(shí)間步輸出的cell state(細(xì)胞狀態(tài))穴亏。
整體代碼如下:
# coding: UTF-8
"""
? ? @author: samuel ko
? ? @date:? 2018/12/18
? ? @target: 研究TF模型的中間輸出結(jié)果
"""
import sys
import os
import numpy as np
import tensorflow as tf
path_file = __file__
dir_name = os.path.dirname(path_file)
# 1. 準(zhǔn)備輸入
inputs = np.array([[0], [0.03846154], [0.07692308]])
inputs = np.expand_dims(inputs, 0)
labels = np.array([[0, 1, 0, 0, 0, 0, 0, 0, 0, 0,
? ? ? ? ? ? ? ? ? ? 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
? ? ? ? ? ? ? ? ? ? 0, 0, 0, 0, 0, 0]])
# 2. 加載模型, 輸出中間結(jié)果和最后結(jié)果
with tf.Session() as sess:
? ? graph = tf.get_default_graph()
? ? new_saver = tf.train.import_meta_graph(os.path.join(dir_name, 'model/simple_lstm.meta'))
? ? # 注: tf.train_get_checkpoint_state不允許接收中文, tf.train.latest_checkpoint就沒(méi)問(wèn)題...
? ? # new_saver.restore(sess, tf.train.get_checkpoint_state(os.path.join(dir_name, "model/")))
? ? new_saver.restore(sess, tf.train.latest_checkpoint(os.path.join(dir_name, "model/")))
? ? input_x = graph.get_tensor_by_name("Inputs:0")
? ? label_x = graph.get_tensor_by_name("Labels:0")
? ? # out 是輸入到下一層的匯總 3 x 1 x 5
? ? out = graph.get_tensor_by_name('rnn/TensorArrayStack/TensorArrayGatherV3:0')
? ? # state_h 是LSTM層最后一個(gè)時(shí)間步的結(jié)果 1 x 5
? ? state_h = graph.get_tensor_by_name('rnn/while/Exit_4:0') # 最后一個(gè)時(shí)間步的memory state 和state_h = graph.get_tensor_by_name('rnn/while/Switch_4:0') 一樣蜂挪!
? ? # state_h = graph.get_tensor_by_name('rnn/while/Exit_3:0') # 最后一個(gè)時(shí)間步的cell state
? ? print(sess.run(out, feed_dict={input_x: inputs,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? label_x: labels,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? }))
? ? print(sess.run(state_h, feed_dict={input_x: inputs,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? label_x: labels,
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? }))
h_tm_i, h_tm_o, h_tm_c, h_tm_f, c_tm = None, None, None, None, None
def sigmoid(x):
? ? return 1.0 / (1.0 + np.exp(-x))
def lstm_tf_verify(inputs):
? ? """
? ? ? ? 2018/12/18
? ? ? ? TF原生的解析順序?yàn)閕, j, f, o (j就是keras中的c)
? ? :param inputs:
? ? :return:
? ? """
? ? global h_tm_c, h_tm_f, h_tm_i, h_tm_o, c_tm
? ? bias_i = np.array([0.9502341, 1.1212865, 0.5962041, 0.56686985, 0.65736747])
? ? bias_j = np.array([-0.28798968, 0.31724977, -0.08590735, -0.13165179, -0.05694159])
? ? bias_f = np.array([0.89209175, 1.0639387, 0.3089665, 0.42762548, 0.4232108])
? ? bias_o = np.array([1.0723785, 1.2605966, 0.5964751, 0.6030057, 0.6930808])
? ? kernel_i = np.array([0.96915483, 0.5620192, 0.5136176, 0.1521692, 0.96555483])
? ? kernel_j = np.array([0.6295774, -0.72134864, 0.64238673, 0.48595947, 0.570404])
? ? kernel_f = np.array([0.7884312, 0.56634164, 0.14510694, 0.19882877, 0.6444183])
? ? kernel_o = np.array([0.55998164, 0.5682311, 0.9390488, 0.8536483, 0.9704966])
? ? recurrent_i = np.array([[-0.30848396, -0.13132317, 0.6034289, 0.59028447, 0.09684605],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.28015903, -0.24312414, -0.42499176, -0.3367074, -0.06846467],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.7987564, 0.93413734, -0.15053841, 0.66372687, 0.06576955],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.24111897, 0.1684269, 0.5229809, 0.09525479, 0.28952646],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.70739645, 0.8474347, 0.19091478, 0.02707534, 0.52820826]])
? ? recurrent_j = np.array([[1.272224, -1.475185, 0.38326767, 0.64769256, 0.83099645],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [-0.5344824, 1.2404263, -0.88588023, -0.7727197, -1.167835],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.86383224, -0.8951096, 0.08373257, 0.89576524, 0.53091526],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.7915831, -0.93986595, -0.02958089, 0.82741463, 0.55338454],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.39262557, -0.86354613, 0.62125677, 0.82101977, 0.13056423]])
? ? recurrent_f = np.array([[0.17595771, 0.27790356, 0.6525466, 0.05647744, 0.06983535],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.26703873, 0.04883758, 0.0888641, -0.05813761, 0.0277635],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.6442748, 0.4176797, 0.5382307, 0.48299634, 0.7003999],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.19449034, 0.01752495, 0.13846086, 0.00932326, 0.4014144],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.6212245, 0.59203285, 0.05094814, 0.85539377, 0.6473349]])
? ? recurrent_o = np.array([[0.29326066, 0.50268304, 0.544091, 0.76660025, 0.29213676],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [-0.44291726, -0.338039, -0.17275955, -0.7254445, -0.7070001],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.13272414, 0.8238844, -0.09202695, 0.9273238, 0.15251717],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.06204496, 0.6531808, 0.00607, 0.33238858, 0.04696886],
? ? ? ? ? ? ? ? ? ? ? ? ? ? [0.9217779, 0.6748385, 0.61127436, 0.5573597, 0.21182081]])
? ? # step 1 計(jì)算W * x
? ? x_i = inputs * kernel_i
? ? x_f = inputs * kernel_f
? ? x_j = inputs * kernel_j
? ? x_o = inputs * kernel_o
? ? # step 2 加上bias
? ? x_i += bias_i
? ? x_f += bias_f
? ? x_j += bias_j
? ? x_o += bias_o
? ? # step 3 計(jì)算
? ? if not isinstance(h_tm_i, np.ndarray):
? ? ? ? h_tm_i = np.zeros((1, 5))
? ? ? ? h_tm_o = np.zeros((1, 5))
? ? ? ? h_tm_f = np.zeros((1, 5))
? ? ? ? h_tm_c = np.zeros((1, 5))
? ? ? ? c_tm = np.zeros((1, 5))
? ? i = sigmoid(x_i + np.dot(h_tm_i, recurrent_i))
? ? # Tensorflow默認(rèn)有一個(gè)forget_bias, 默認(rèn)設(shè)置為1.0
? ? f = sigmoid(x_f + np.dot(h_tm_f, recurrent_f) + 1.0)
? ? c = f * c_tm + i * np.tanh(x_j + np.dot(h_tm_c, recurrent_j))
? ? o = sigmoid(x_o + np.dot(h_tm_o, recurrent_o))
? ? h = o * np.tanh(c)
? ? h_tm_c = h_tm_f = h_tm_o = h_tm_i = h
? ? c_tm = c
? ? print("當(dāng)前的hidden state", h)
? ? print("當(dāng)前的cell state", c)
? ? return h, c
if __name__ == "__main__":
? ? for i in range(3):
? ? ? ? print("輸入內(nèi)容:", inputs[:, i])
? ? ? ? # lstm_keras_verify(inputs[:, i])
? ? ? ? lstm_tf_verify(inputs[:, i])
輸出結(jié)果為:
# output 3 x 1 x 5 當(dāng)前層的每個(gè)時(shí)間步的hidden state匯總
[[[-0.14857864? 0.17725913 -0.03559565 -0.05385567 -0.02496454]]
[[-0.3793954? 0.45447606 -0.13174371 -0.17756298 -0.17771873]]
[[-0.5253717? 0.55423415 -0.25274208 -0.25586015 -0.34587777]]]
# state.h 最后一個(gè)時(shí)間步的hidden state
[[-0.5253717? 0.55423415 -0.25274208 -0.25586015 -0.34587777]]
輸入內(nèi)容: [[0.]]
當(dāng)前的hidden state [[-0.14857867? 0.17725915 -0.03559565 -0.05385567 -0.02496454]]
當(dāng)前的cell state [[-0.20212986? 0.23156138 -0.05525611 -0.08351723 -0.03746516]]
輸入內(nèi)容: [[0.03846154]]
當(dāng)前的hidden state [[-0.37939543? 0.45447602 -0.13174374 -0.17756298 -0.17771877]]
當(dāng)前的cell state [[-0.58665553? 0.71037671 -0.21416421 -0.31547094 -0.28813169]]
輸入內(nèi)容: [[0.07692308]]
當(dāng)前的hidden state [[-0.5253716? 0.55423418 -0.25274209 -0.25586014 -0.34587777]]
當(dāng)前的cell state [[-1.12897442? 1.26972863 -0.47543917 -0.66030582 -0.70899148]]
可以看出,我們的實(shí)現(xiàn)跟TF基本一樣(跟Keras一樣嗓化,都有一點(diǎn)點(diǎn)精度損失)棠涮。
# TF結(jié)果
[[-0.5253717? 0.55423415 -0.25274208 -0.25586015 -0.34587777]]
# Numpy自己實(shí)現(xiàn)結(jié)果
[[-0.5253716? 0.55423418 -0.25274209 -0.25586014 -0.34587777]]
這部分,我們將對(duì)Keras和Tensorflow的LSTM層的計(jì)算邏輯進(jìn)行細(xì)致的分析刺覆,源碼位置在文章一開(kāi)頭严肪,建議大家進(jìn)去看后再來(lái)看這部分,會(huì)更加明白谦屑。
實(shí)現(xiàn)的代碼主要對(duì)比lstm_keras_verify函數(shù)和lstm_tf_verify函數(shù):顧名思義驳糯,前面是Keras的LSTM實(shí)現(xiàn)邏輯,后面的是Tensorflow的LSTM實(shí)現(xiàn)邏輯氢橙,下面講到的異同點(diǎn)如果源碼里面不好理解酝枢,直接看這里的實(shí)現(xiàn)區(qū)別也行。
① TF的self._kernel包含了input_depth(本例為1)和h_depth(本例為num_units,為5)悍手,即把Keras里面的kernel和recurrent_kernel統(tǒng)一放到了self._kernel里面了帘睦。
所以,當(dāng)我打印simple_lstm的Tensorflow模型時(shí)發(fā)現(xiàn)坦康,rnn/lstm_cell/kernel的size為6 x 20竣付, 6是啥意思呢?6也很簡(jiǎn)單涝焙,其包含了一個(gè)1 x 20的(input_w_kernel)和 5 x 20的(recurrent_w_kernel)——解析順序也是這樣的卑笨。(即不像Keras分為kernel和recurrent_kernel兩個(gè)分別保存權(quán)重。)
Tensorflow中LSTM用于存儲(chǔ)權(quán)重的self._kernel代碼:
@tf_export("nn.rnn_cell.LSTMCell")
class LSTMCell(LayerRNNCell):
...
? @tf_utils.shape_type_conversion
? def build(self, inputs_shape):
? ? if inputs_shape[-1] is None:
? ? ? raise ValueError("Expected inputs.shape[-1] to be known, saw shape: %s"
? ? ? ? ? ? ? ? ? ? ? % str(inputs_shape))
? ? input_depth = inputs_shape[-1]
? ? h_depth = self._num_units if self._num_proj is None else self._num_proj
...
# self._kernel即包含Keras里面的kernel仑撞,也包含recurrent_kernel,是對(duì)Keras的LSTM層權(quán)重的2合1.
? ? self._kernel = self.add_variable(
? ? ? ? _WEIGHTS_VARIABLE_NAME,
? ? ? ? shape=[input_depth + h_depth, 4 * self._num_units],
? ? ? ? initializer=self._initializer,
? ? ? ? partitioner=maybe_partitioner)
? ? ...
? ? self._bias = self.add_variable(
? ? ? ? _BIAS_VARIABLE_NAME,
? ? ? ? shape=[4 * self._num_units],
? ? ? ? initializer=initializer)
② TF里面的i, j, f, o分別對(duì)應(yīng)Keras的LSTM中的i, c, f, o赤兴。也就是說(shuō):Keras對(duì)應(yīng)的權(quán)重和Tensorflow的權(quán)重順序不一樣了!K硐桶良!
3.2.1 Tensorflow的LSTM權(quán)重拆解順序
@tf_export("nn.rnn_cell.LSTMCell")
class LSTMCell(LayerRNNCell):
...
def call(self, inputs, state):
# i, j, f, o其中,j為下面Keras對(duì)應(yīng)的c
i, j, f, o = array_ops.split(
? ? value=lstm_matrix, num_or_size_splits=4, axis=1)
# Diagonal connections
if self._use_peepholes:
? ? # 我們先不看peephole這個(gè)LSTM變種.
? ? ...
else:
? c = (sigmoid(f + self._forget_bias) * c_prev + sigmoid(i) *
? ? ? self._activation(j))
...
m = sigmoid(o) * self._activation(c)
class LSTMCell(Layer):
def build(self, input_shape):
...
# Keras的4個(gè)權(quán)重存儲(chǔ)順序i, f, c, o與Tensorflow的權(quán)重存儲(chǔ)順序i, j, f, o中間順序調(diào)了一下沮翔,
# 也就是Keras的權(quán)重順序是a, b, c, d那么Tensorflow對(duì)應(yīng)的權(quán)重存儲(chǔ)為a, c, b, d.
? ? ? ? self.kernel_i = self.kernel[:, :self.units]
? ? ? ? self.kernel_f = self.kernel[:, self.units: self.units * 2]
? ? ? ? self.kernel_c = self.kernel[:, self.units * 2: self.units * 3]
? ? ? ? self.kernel_o = self.kernel[:, self.units * 3:]
# recurrent_kernel與kernel的順序是一樣的.
? ? ? ? self.recurrent_kernel_i = self.recurrent_kernel[:, :self.units]
? ? ? ? self.recurrent_kernel_f = (
? ? ? ? ? ? self.recurrent_kernel[:, self.units: self.units * 2])
? ? ? ? self.recurrent_kernel_c = (
? ? ? ? ? ? self.recurrent_kernel[:, self.units * 2: self.units * 3])
? ? ? ? self.recurrent_kernel_o = self.recurrent_kernel[:, self.units * 3:]
? ? ? ? if self.use_bias:
? ? ? ? ? ? self.bias_i = self.bias[:self.units]
? ? ? ? ? ? self.bias_f = self.bias[self.units: self.units * 2]
? ? ? ? ? ? self.bias_c = self.bias[self.units * 2: self.units * 3]
? ? ? ? ? ? self.bias_o = self.bias[self.units * 3:]
? ? ? ? ...
③ Keras的LSTM中的recurrent_activation: (對(duì)應(yīng)Part1的Keras的LSTM計(jì)算邏輯梳理介紹里面的σ σσ)用的是一種叫做hard_sigmoid的實(shí)現(xiàn)陨帆,TF的兩個(gè)的實(shí)現(xiàn)都是一樣的,用的是正常的sigmoid采蚀。而無(wú)論是Keras還是Tensorflow疲牵,它們的activation都是tanh,這個(gè)是一樣的榆鼠。
# Tensorflow LSTM用的recurrent_activation.
def sigmoid(x):
? ? return 1.0 / (1.0 + np.exp(-x))
# Keras LSTM用的recurrent_activation.
def hard_sigmoid(x):
? ? x = 0.2 * x + 0.5
? ? x[x < -2.5] = 0
? ? x[x > 2.5] = 1
? ? return x
④ Tensorflow還有一個(gè)叫做forget_bias的東西纲爸,默認(rèn)為1.0,關(guān)于這個(gè)參數(shù)的介紹如下:
Biases of the forget gate are initialized by default to 1 in order to reduce the scale of forgetting at the beginning of the training. Must set it manually to?0.0?when restoring from CudnnLSTM trained checkpoints.
它用在遺忘門(mén)(forget gate)(上面的lstm_tf_verify函數(shù)),如下:
# Tensorflow默認(rèn)有一個(gè)forget_bias, 默認(rèn)設(shè)置為1.0
f = sigmoid(x_f + np.dot(h_tm_f, recurrent_f) + 1.0)
# 而Keras默認(rèn)不帶這個(gè)東西:
f = hard_sigmoid(x_f + np.dot(h_tm_f, recurrent_kernel_f))
⑤ Keras的LSTM實(shí)現(xiàn)起來(lái)很清爽妆够,沒(méi)有什么亂78糟的參數(shù)识啦;而Tensorflow可以直接在LSTM上面做變種——比如peephole connection[3], 就是說(shuō),我們讓門(mén)層也會(huì)接受細(xì)胞狀態(tài)(cell state)的輸入神妹。
還有就是TF和Keras的LSTM實(shí)現(xiàn)上有一些不一致的地方颓哮,需要大家小心對(duì)待,找出異同點(diǎn)鸵荠,根據(jù)自己的情況對(duì)層進(jìn)行拆解冕茅,方便的完成解耦工作。
關(guān)于Keras和Tensorflow的LSTM層分析基本也就到此結(jié)束了蛹找,如果想更加深入的理解它們的實(shí)現(xiàn)嵌赠,比如分析這種帶時(shí)間信息的層的反向傳播邏輯,建議深挖源碼熄赡,這塊我也不甚了解姜挺。希望能跟大家多多交流,謝謝~
[1]?Netron: a viewer for neural network, deep learning and machine learning models.
[2]?理解 LSTM(Long Short-Term Memory, LSTM) 網(wǎng)絡(luò)
[3]?Gers & Schmidhuber (2000) : Recurrent Nets that Time and Count