纯python 实现lstm，进行点的跟踪。

纯python 实现lstm，进行点的跟踪。加载tensorflow训练得到的权重，实现追踪。c++版本的在另外一篇博客里。

// An highlighted block
import numpy as np

from pandas import read_excel
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler

# global h_tm_c, h_tm_f, h_tm_i, h_tm_o, c_tm
# global h_tm_c2, h_tm_f2, h_tm_i2, h_tm_o2, c_tm2

def hard_sigmoid(x):
    x = 0.2 * x + 0.5
    x[x < 0] = 0
    x[x > 1] = 1
    return x

def lstm_keras_verify(inputs,ten2,ten3):

    global h_tm_c, h_tm_f, h_tm_i, h_tm_o, c_tm
    global h_tm_c2, h_tm_f2, h_tm_i2, h_tm_o2, c_tm2

    if ten2 == 0:
        h_tm_c, h_tm_f, h_tm_i, h_tm_o, c_tm = None,None,None,None,None
        h_tm_c2, h_tm_f2, h_tm_i2, h_tm_o2, c_tm2 = None, None, None, None, None
    kernel_i = np.array([-0.34339943528175354,
        1.766756296157837,
        0.7450317740440369,
        -0.829913854598999,
        0.5050585269927979,
        0.5951665043830872,
        -0.12462729960680008,
        -0.057644475251436234,
        -0.195032998919487,
        0.9801527857780457])
    kernel_f = np.array([-0.2959004342556,
        0.1878744661808014,
        0.1974334567785263,
        -0.42833104729652405,
        0.655572772026062,
        0.5754011273384094,
        0.20625288784503937,
        0.368974894285202,
        -0.32602018117904663,
        0.276435434818267])
    kernel_c = np.array([0.09515336155891418,
        -0.5731109976768494,
        -0.4054526388645172,
        -0.314892441034317,
        0.3126268684864044,
        -0.5319279432296753,
        0.653582751750946,
        0.49748483300209045,
        0.20359081029891968,
        0.5867465138435364])
    kernel_o = np.array([0.20324330031871796,
        1.5852749347686768,
        0.6022287011146545,
        -1.1601568460464478,
        0.7637105584144592,
        0.7074271440505981,
        0.6657699942588806,
        -0.3173244595527649,
        -0.3717532157897949,
        1.0292516946792603])

    kernel_i2 = np.array([0.29539164900779724,
        -0.5408189296722412,
        0.31895196437835693,
        -0.3692389726638794,
        0.5499736070632935,
        0.0356333963572979,
        0.6757100820541382,
        0.930941641330719,
        0.4551529884338379,
        0.07838062196969986])
    kernel_f2 = np.array([0.1885366439819336,
        -0.5314029455184937,
        -0.25439608097076416,
        -0.32862451672554016,
        -0.026900850236415863,
        -0.6118189096450806,
        -0.8291504979133606,
        -0.10796643793582916,
        0.3633717894554138,
        -0.5390552878379822])
    kernel_c2 = np.array([ -0.042067818343639374,
        0.7635902166366577,
        -0.1499466896057129,
        0.7719531059265137,
        0.4475833773612976,
        -0.006783284712582827,
        -0.1789802610874176,
        -0.5963394045829773,
        -0.5599965453147888,
        -0.37145528197288513])
    kernel_o2 = np.array([ -0.18046832084655762,
        -0.17623309791088104,
        0.20015233755111694,
        -0.6863264441490173,
        0.0444098562002182,
        0.42466920614242554,
        0.31134355068206787,
        0.5258762240409851,
        0.6071751117706299,
        -0.7286037802696228])

    recurrent_kernel_i = np.array([[-0.41451895236968994,
        2.3002400398254395,
        -0.2820678651332855,
        -0.45961064100265503,
        0.505663275718689,
        0.7919554114341736,
        0.20269405841827393,
        0.8496864438056946,
        0.08628702908754349,
        0.7238653302192688], [-0.9789332151412964,
        -1.5582021474838257,
        -1.3366812467575073,
        -0.27708300948143005,
        0.4506531357765198,
        -1.467345952987671,
        -0.5549973249435425,
        1.5742740631103516,
        0.2451542168855667,
        -0.940199077129364], [0.4405961036682129,
        -1.5445444583892822,
        0.05049801990389824,
        -0.35451313853263855,
        -0.269658625125885,
        -0.638664722442627,
        -0.789127767086029,
        -0.7010314464569092,
        -0.13462814688682556,
        -0.6025670766830444], [-0.887526273727417,
        0.4103408753871918,
        -0.4497901499271393,
        -0.2633923292160034,
        -0.22475440800189972,
        1.236531376838684,
        0.6960248947143555,
        0.33017680048942566,
        -0.5554746985435486,
        1.7138985395431519], [ 0.12573093175888062,
        0.15226393938064575,
        0.35927459597587585,
        -0.2610260248184204,
        0.1653788834810257,
        0.8279508352279663,
        1.0212059020996094,
        0.820050060749054,
        0.5119935870170593,
        0.3915463089942932], [-0.001828429289162159,
        -3.9691922664642334,
        0.37881338596343994,
        0.10328131914138794,
        -0.296207994222641,
        -1.0051754713058472,
        -0.8751426339149475,
        -0.4393799901008606,
        -0.5824257135391235,
        -0.9112383723258972], [0.14419855177402496,
        0.07607676088809967,
        0.5626516938209534,
        -1.290927529335022,
        -0.0973503440618515,
        0.5256897807121277,
        0.30444854497909546,
        -0.9554680585861206,
        -0.2767266035079956,
        0.2846880555152893], [ 0.20677947998046875,
        2.7582507133483887,
        -0.2818203270435333,
        0.9349485635757446,
        -0.985670268535614,
        0.8378878831863403,
        -0.7977424263954163,
        -0.6354966163635254,
        -0.5764774084091187,
        0.5060479044914246], [-0.14697682857513428,
        1.3863033056259155,
        -0.14390596747398376,
        0.8854043483734131,
        -0.9272205829620361,
        -0.7263381481170654,
        -1.2763465642929077,
        -1.5187703371047974,
        -0.8858398795127869,
        0.05383571982383728], [0.5134013891220093,
        2.266584634780884,
        0.4238872528076172,
        -0.03581789508461952,
        0.1086997240781784,
        0.645852267742157,
        1.1289607286453247,
        0.10689905285835266,
        0.2679310739040375,
        0.67705899477005]])
    recurrent_kernel_f = np.array([[-0.16183863580226898,
        0.7472390532493591,
        -0.334164023399353,
        -0.1287859082221985,
        0.32973605394363403,
        -0.1532825231552124,
        -0.18556086719036102,
        1.174559235572815,
        0.315372109413147,
        -0.14612999558448792], [-1.0360535383224487,
        0.19038058817386627,
        -0.9956535696983337,
        -0.28531670570373535,
        0.7210497260093689,
        -0.02345925383269787,
        -1.0118887424468994,
        0.5574779510498047,
        0.3704586923122406,
        -0.6031738519668579], [ 0.3754199743270874,
        -0.014456510543823242,
        -0.06675058603286743,
        0.17784777283668518,
        -0.16492195427417755,
        0.2980993986129761,
        0.25578248500823975,
        -0.7040256261825562,
        -0.5165166258811951,
        0.1266956329345703], [ -0.6861728429794312,
        -0.16793103516101837,
        -0.271343469619751,
        -0.27307748794555664,
        -0.19215670228004456,
        0.27285122871398926,
        0.025628695264458656,
        -0.2498171329498291,
        -0.571029543876648,
        0.3341529667377472], [ -0.2412928342819214,
        0.13404031097888947,
        0.027914635837078094,
        -0.3998990058898926,
        0.14130471646785736,
        -0.23286490142345428,
        -0.5493660569190979,
        0.20289260149002075,
        0.6077360510826111,
        -0.41169750690460205], [0.10601703077554703,
        -0.4763187766075134,
        0.0049462090246379375,
        0.5886510014533997,
        -0.2985691428184509,
        0.17912009358406067,
        0.346873015165329,
        -0.650505542755127,
        -0.385824054479599,
        0.09396341443061829], [0.5001235008239746,
        1.339636206626892,
        0.6780133247375488,
        -1.1757398843765259,
        -0.20317663252353668,
        0.5733743906021118,
        -0.506499171257019,
        -0.3316236138343811,
        -0.23696570098400116,
        0.017315169796347618], [-0.008928589522838593,
        0.7925891280174255,
        0.7797060012817383,
        0.8483789563179016,
        -0.7849624156951904,
        1.0222151279449463,
        0.3750150799751282,
        0.6228334307670593,
        -1.3550268411636353,
        0.3627028465270996], [ -0.18932075798511505,
        0.5054185390472412,
        0.21752338111400604,
        0.5696843862533569,
        -1.0764824151992798,
        0.8123449087142944,
        0.3687470555305481,
        -0.10713357478380203,
        -0.8887754082679749,
        -0.02434460259974003], [ -0.17165663838386536,
        0.7209696769714355,
        0.6251261830329895,
        0.013831086456775665,
        0.06497514992952347,
        0.4622625410556793,
        0.09293515235185623,
        0.20599240064620972,
        -0.0917106419801712,
        -0.08447670191526413]])
    recurrent_kernel_c = np.array([[ -0.6100238561630249,
        -0.12698210775852203,
        0.4954182505607605,
        0.15199264883995056,
        0.35257288813591003,
        -0.12464402616024017,
        -0.44923147559165955,
        0.040395427495241165,
        0.7752857208251953,
        -0.05500243604183197], [0.4752281606197357,
        -0.2517543435096741,
        0.16683556139469147,
        0.2313675582408905,
        0.7824580669403076,
        -0.07187685370445251,
        0.19583779573440552,
        -0.15934820473194122,
        -0.09073260426521301,
        0.1494336724281311], [-0.07115137577056885,
        -0.27783897519111633,
        -0.1414336860179901,
        -0.5170375108718872,
        -0.054542649537324905,
        -0.320780485868454,
        0.44443681836128235,
        0.16018202900886536,
        0.16488415002822876,
        0.26717546582221985], [0.23076167702674866,
        -0.558992326259613,
        -0.28554731607437134,
        0.40353721380233765,
        0.07054921239614487,
        -0.28317591547966003,
        0.17094585299491882,
        0.252930223941803,
        0.3688906729221344,
        0.2711980938911438], [ 0.14932553470134735,
        0.056401338428258896,
        -0.25233572721481323,
        0.4610925614833832,
        0.22263102233409882,
        0.02338065765798092,
        -0.040763139724731445,
        -0.29332226514816284,
        -0.2517569959163666,
        -0.07778314501047134], [-0.07450161874294281,
        0.17594310641288757,
        -0.09124835580587387,
        -0.9320527911186218,
        0.06922359019517899,
        -0.1493844985961914,
        0.593464732170105,
        0.19774796068668365,
        0.058623287826776505,
        0.11784616112709045], [ 0.3252342641353607,
        -0.05548950284719467,
        -0.08142340183258057,
        -0.12902313470840454,
        -0.9895947575569153,
        -0.4129822254180908,
        -0.3162073493003845,
        0.5147839188575745,
        0.3044321835041046,
        -0.10074173659086227], [ 0.1463569700717926,
        0.006250813137739897,
        0.39591124653816223,
        0.5136163830757141,
        -0.5497452616691589,
        -0.16999530792236328,
        -0.43260541558265686,
        0.07086426019668579,
        -0.40135759115219116,
        -0.4523780345916748], [0.2282596379518509,
        -0.142124742269516,
        0.22567713260650635,
        0.13072924315929413,
        -0.5392906069755554,
        -0.18509604036808014,
        -0.02913133054971695,
        0.206122487783432,
        -0.11300422251224518,
        -0.2983923852443695], [ 0.27763694524765015,
        -0.08438488095998764,
        -0.2187069207429886,
        0.2885969281196594,
        -0.20823447406291962,
        -0.020448746159672737,
        -0.40587231516838074,
        0.271494060754776,
        0.09722976386547089,
        -0.4638107120990753]])
    recurrent_kernel_o = np.array([[0.08256174623966217,
        1.0839580297470093,
        -0.3347654640674591,
        -0.17489440739154816,
        -0.03872399777173996,
        0.1976941078901291,
        0.3668980002403259,
        0.6379971504211426,
        0.025410614907741547,
        0.03770061582326889], [-0.8977024555206299,
        -0.510123610496521,
        -1.548133373260498,
        -0.21091605722904205,
        -0.12713493406772614,
        -1.005964756011963,
        -0.3703922629356384,
        1.4118517637252808,
        0.20945164561271667,
        -1.00154447555542], [-0.005945556331425905,
        -0.7805576920509338,
        0.016335049644112587,
        -0.011125431396067142,
        0.2580377757549286,
        -0.290608286857605,
        -0.8096177577972412,
        -0.5068511366844177,
        -0.08132117241621017,
        -0.3691478967666626], [ -1.0412406921386719,
        1.1914699077606201,
        -0.8255971670150757,
        -0.15991421043872833,
        0.04529210552573204,
        0.21786856651306152,
        0.797028660774231,
        0.5845527648925781,
        -0.6345934867858887,
        0.8253769278526306], [ 0.3824268877506256,
        0.367237389087677,
        -0.03931231051683426,
        -0.2734659016132355,
        0.06321419030427933,
        0.7792273759841919,
        0.872334361076355,
        0.8914680480957031,
        0.5063027739524841,
        -0.02294449880719185], [0.17051495611667633,
        -3.346440315246582,
        -0.1220833957195282,
        0.16170988976955414,
        0.23057465255260468,
        -0.5420761108398438,
        -0.3005651533603668,
        -0.20853359997272491,
        -0.27472618222236633,
        -0.6494842767715454], [0.3111186623573303,
        -0.24375362694263458,
        1.1926289796829224,
        -1.8478845357894897,
        0.37887048721313477,
        0.7220742106437683,
        0.51273512840271,
        -0.3570004403591156,
        -1.0175435543060303,
        0.3988095819950104], [ -0.0040112873539328575,
        1.8080674409866333,
        0.5294325947761536,
        0.5869970917701721,
        -0.2983153164386749,
        0.87757807970047,
        -0.2792051136493683,
        -1.231807827949524,
        -1.0860964059829712,
        1.0274618864059448], [-0.06216314807534218,
        0.7327183485031128,
        0.11892963945865631,
        0.6781867742538452,
        -0.6871398687362671,
        0.319871187210083,
        -1.1046788692474365,
        -1.0522805452346802,
        -1.1571115255355835,
        0.4382123649120331], [-0.16790947318077087,
        1.1358662843704224,
        0.5037592649459839,
        0.175055593252182,
        -0.013465126045048237,
        0.46570730209350586,
        0.8169551491737366,
        0.3778005540370941,
        0.11658983677625656,
        0.3484398424625397]])

    bias_i = np.array([-0.446609765291214,
    -0.11702976375818253,
    -0.3065964877605438,
    -0.025546252727508545,
    -0.049140945076942444,
    -0.001070813275873661,
    0.612964928150177,
    0.16031958162784576,
    -0.0778871700167656,
    0.15741103887557983])
    bias_f = np.array([ 0.31866440176963806,
    0.4428155720233917,
    0.5675423741340637,
    0.8959832191467285,
    0.9183741807937622,
    0.499959260225296,
    0.6026825904846191,
    0.5548291206359863,
    0.5360163450241089,
    0.5315966010093689])
    bias_c = np.array([-0.1425662487745285,
    0.05167638882994652,
    0.1590413600206375,
    -0.1351909041404724,
    -0.006993392016738653,
    0.26650679111480713,
    -0.0693528801202774,
    0.00884726271033287,
    0.17367465794086456,
    -0.0636242926120758])
    bias_o = np.array([-0.47558993101119995,
    0.41542863845825195,
    -0.28937411308288574,
    -0.12115256488323212,
    -0.10851451754570007,
    -0.03444031625986099,
    0.7349303960800171,
    0.2082076072692871,
    -0.13617178797721863,
    0.07746882736682892])

    # step 1 计算W * x
    x_i = inputs[0] * kernel_i
    x_f = inputs[0] * kernel_f
    x_c = inputs[0] * kernel_c
    x_o = inputs[0] * kernel_o

    x_i2 = inputs[1] * kernel_i2
    x_f2 = inputs[1] * kernel_f2
    x_c2 = inputs[1] * kernel_c2
    x_o2= inputs[1] * kernel_o2

    # step 2 加上bias
    x_i += bias_i
    x_f += bias_f
    x_c += bias_c
    x_o += bias_o

    x_i2 += bias_i
    x_f2 += bias_f
    x_c2 += bias_c
    x_o2 += bias_o

    # step 3 计算
    if not isinstance(h_tm_i, np.ndarray):
        h_tm_i = np.zeros((1, 10))
        h_tm_o = np.zeros((1, 10))
        h_tm_f = np.zeros((1, 10))
        h_tm_c = np.zeros((1, 10))
        c_tm = np.zeros((1, 10))

        h_tm_i2 = np.zeros((1, 10))
        h_tm_o2 = np.zeros((1, 10))
        h_tm_f2 = np.zeros((1, 10))
        h_tm_c2 = np.zeros((1, 10))
        c_tm2 = np.zeros((1, 10))
    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))


    i2 = hard_sigmoid(x_i2 + np.dot(h_tm_i2, recurrent_kernel_i))
    f2 = hard_sigmoid(x_f2 + np.dot(h_tm_f2, recurrent_kernel_f))
    c2 = f2 * c_tm2 + i2 * np.tanh(x_c2 + np.dot(h_tm_c2, recurrent_kernel_c))
    o2 = hard_sigmoid(x_o2 + np.dot(h_tm_o2, recurrent_kernel_o))

    h = o * np.tanh(c)
    h2 = o2 * np.tanh(c2)

    h_tm_c = h_tm_f = h_tm_o = h_tm_i = h
    c_tm = c

    h_tm_c2 = h_tm_f2 = h_tm_o2 = h_tm_i2 = h2
    c_tm2 = c2

    dense_weights = np.array([[0.25596046447753906], [-0.6001706123352051], [  -0.7026602625846863], [-0.6046009659767151], [ 0.7513906359672546], [-0.6343705654144287], [1.1254067420959473], [0.6913528442382812], [0.3435429632663727], [ 0.7804898023605347]])
    dense_bias = np.array([0.11426012963056564])

    dense_weights2 = np.array([[0.476622074842453], [-0.05293620750308037], [-0.31866875290870667], [0.7019630670547485], [0.3929026424884796], [-0.11945560574531555], [0.2096286118030548], [-0.14515486359596252], [-0.7117556929588318], [ -0.36956626176834106]])
    dense_bias2 = np.array([0.22602976858615875])

    y = np.dot(h, dense_weights) + dense_bias
    y2 = np.dot(h2, dense_weights2) + dense_bias2

    y = y.flatten()
    y2 = y2.flatten()
    y = y.squeeze()
    y2 = y2.squeeze()

    y = np.array([y,y2])
    #y = scaler.inverse_transform(np.array(y,y2))
    #print(y, y2)
    return y

def create_dataset(dataset, look_back=1):
    dataX, dataY = [], []
    predict_len = 5
    for i in range(len(dataset)-look_back-1 -predict_len):
        a = dataset[i:(i+look_back), 0:2]
        dataX.append(a)
        dataY.append(dataset[i + look_back + predict_len, 0:2])
    return np.array(dataX), np.array(dataY)

if __name__ == "__main__":
    dataframe = read_excel(r'C:\Users\gj7520\Desktop\pythob_files\file_select\data4\data_select.xls', usecols=[0, 1], skipfooter=3)
    dataset = dataframe.values
    # 将整型变为float
    dataset = dataset.astype('float32')
    # fix random seed for reproducibility
    np.random.seed(7)

    scaler = MinMaxScaler(feature_range=(0, 1))
    dataset = scaler.fit_transform(dataset)

    # split into train and test sets
    #train_size = int(len(dataset) * 0.8)
    #test_size = len(dataset) - train_size
    train_size = 0
    test_size = len(dataset)
    train, test = dataset[0:train_size, :2], dataset[train_size:len(dataset), :2]

    # use this function to prepare the train and test datasets for modeling
    look_back = 5
    trainX, trainY = create_dataset(train, look_back)
    testX, testY = create_dataset(test, look_back)
    arr = []
    arr_train = []
    arr2 = []
    arr3 = []
    for j in range(len(testX) - 10):
        for i in range(0, len(testX[j])):
            y = lstm_keras_verify(testX[j][i], i,j)
            if i == len(testX[j]) - 1:
                arr.append(y)
                if(  testX[j+5][0][1] > 0.02):
                    #arr2.append((y[0] - testX[j + 5][0][0])/testX[j+5][0][0])
                    arr3.append((y[1] - testX[j + 5][0][1])/testX[j+5][0][1])
                    if ((y[1] - testX[j + 5][0][1])/testX[j+5][0][1] >  30):
                        a = 1;
    arr = np.array(arr)
    #arr = scaler.inverse_transform(arr)
    #arr2 = np.array(arr2)
    arr3 = np.array(arr3)
    #dataset_inve = scaler.inverse_transform(dataset)
    #testY = scaler.inverse_transform(testY)

    # shift train predictions for plotting
    # #
    # for j in range(len(trainX)):
    #     for i in range(0, len(trainX[j])):
    #         y2 = lstm_keras_verify(trainX[j][i], i,j)
    #         if i == len(trainX[j]) - 1:
    #             arr_train.append(y2)
    # arr_train = np.array(arr_train)
    #
    # arr_train = scaler.inverse_transform(arr_train)
    # trainY = scaler.inverse_transform(trainY)
    # predict_len = 5
    #
    # # shift test predictions for plotting
    # plt.plot(arr_train)
    # plt.plot(trainY)
    # plt.plot(testY)
    # plt.plot(arr)
    plt.plot(arr3)
    # plt.plot(dataset)
    # for i in range(0, len(testX[1])):
    #     y3 = lstm_keras_verify(testX[1][i], i,0)
    #     if(i == len(testX[1]) - 1):
    #         print(y3)

    # print(trainY[1])
    # print(testX[1])
    plt.show()