Blame view
STACKEDAE_MODELS.py
7.39 KB
b6d0165d1
Initial commit
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 |
# coding: utf-8
# In[2]:
# Import
import pandas
# Alignement
import nltk
import codecs
import gensim
from scipy import sparse
import itertools
from sklearn.feature_extraction.text import CountVectorizer
import scipy.sparse
import scipy.io
from sklearn import preprocessing
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation,AutoEncoder
from keras.optimizers import SGD,Adam
from keras.layers import containers
from mlp import *
import mlp
import sklearn.metrics
import shelve
import pickle
from utils import *
import sys
import json
import itertools
# In[4]:
db=shelve.open("{}.shelve".format(sys.argv[2]),writeback=True)
sparse_model=shelve.open("{}.shelve".format(sys.argv[1]))
#['ASR', 'TRS', 'LABEL']
# In[6]:
ASR=sparse_model["ASR"]
TRS=sparse_model["TRS"]
LABEL=sparse_model["LABEL"]
db["ASR_SPARSE"]=ASR
db["TRS_SPARSE"]=TRS
db["LABEL"]=LABEL
print "todo label"
def select(elm):
return int(elm.split("_")[-1])
#z.apply(select)
label_bin={}
lb = preprocessing.LabelBinarizer(neg_label=0)
lb.fit(LABEL["TRAIN"].apply(select))
for i in ASR.keys():
label_bin=lb.transform(LABEL[i].apply(select))
hidden_size=50
input_activation="tanh"
out_activation="tanh"
loss="mse"
epochs=500
batch=1
patience=60
w1_size=300
w2_size=500
do_do=True
sgd = Adam(lr=0.0001)#SGD(lr=0.00001,nesterov=False) #'rmsprop'# Adam(lr=0.00001)#SGD(lr=0.001, momentum=0.9, nesterov=True)
try :
sgd_repr=sgd.get_config()
except AttributeError :
sgd_repr=sgd
json.dump({ "h1" : hidden_size,
"inside_activation" : input_activation,
"out_activation" : out_activation,
"do_dropout": do_do,
"loss" : loss,
"epochs" : epochs ,
"batch_size" : batch,
"patience" : patience,
"sgd" : sgd_repr},
open("{}.json".format(sys.argv[2]),"w"),
indent=4)
def fintuned(sized,x_data,y_data,input_activation,do_do):
ae = Sequential()
previous = x_data["TRAIN"].shape[1]
for hidden_size,w in sized:
ae.add(Dense(hidden_size,input_dim=previous,activation=input_activation,weights=w[:2]))
if do_do :
ae.add(Dropout(0.5))
previous = hidden_size
ae.compile(sgd,loss)
hist = ae.fit(x_data["TRAIN"].todense(),y_data["TRAIN"].todense(),nb_epoch=epochs,batch_size=batch,
callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)],validation_data=(x_data["DEV"].todense(),y_data["DEV"].todense()),verbose=1)
return (ae.get_weights(),hist)
def get_projection(weights,x_data):
proj_list=[x_data]
lappend=proj_list.append
for w in zip(weights[0::2],weights[1::2]) :
decoder =Sequential()
decoder.add(Dense(w[1].shape[0],input_dim=proj_list[-1]["TRAIN"].shape[1],weights=w))
decoder.compile(sgd,loss)
lappend({})
for key in proj_list[-2].keys():
try :
proj_list[-1][key]=decoder.predict(proj_list[-2][key])
except :
proj_list[-1][key]=decoder.predict(proj_list[-2][key].todense())
return proj_list
def learn_ae(hidden_size,data,do_do,input_activation,output_activation):
autoencode=Sequential()
autoencode.add(Dense(hidden_size,input_dim=data["TRAIN"].shape[1],init='glorot_uniform',activation=input_activation))
if do_do :
autoencode.add(Dropout(0.5))
try :
autoencode.add(Dense(data["DEV"].todense().shape[1],input_dim=hidden_size,init="glorot_uniform",activation=out_activation))
autoencode.compile(optimizer=sgd,loss=loss)
hist =autoencode.fit(data["TRAIN"].todense(),data["TRAIN"].todense(),nb_epoch=epochs,batch_size=batch,
callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)],validation_data=(data["DEV"].todense(),data["DEV"].todense()),verbose=1)
auto_decoder=Sequential()
auto_decoder.add(Dense(hidden_size,input_dim=data["DEV"].todense().shape[1],init='uniform',activation=input_activation,weights=autoencode.get_weights()[:2]))
auto_decoder.compile(optimizer=sgd,loss=loss)
data_proj={}
for i in data.keys():
data_proj[i]=auto_decoder.predict(data[i].todense())
return (data_proj,(hist.epoch,hist.history),autoencode.get_weights())
except AttributeError :
autoencode.add(Dense(data["DEV"].shape[1],input_dim=hidden_size,init="glorot_uniform",activation=out_activation))
autoencode.compile(optimizer=sgd,loss=loss)
hist =autoencode.fit(data["TRAIN"],data["TRAIN"],nb_epoch=epochs,batch_size=batch,
callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)],validation_data=(data["DEV"],data["DEV"]),verbose=1)
auto_decoder=Sequential()
auto_decoder.add(Dense(hidden_size,input_dim=data["DEV"].shape[1],init='uniform',activation=input_activation,weights=autoencode.get_weights()[:2]))
auto_decoder.compile(optimizer=sgd,loss=loss)
data_proj={}
for i in data.keys():
data_proj[i]=auto_decoder.predict(data[i])
return (data_proj,(hist.epoch,hist.history),autoencode.get_weights())
print "gogo autoencoder ASR"
hist = {}
ASR_AE_H1,hist["ASR_AE_H1"],wa1=learn_ae(hidden_size,ASR,do_do,input_activation,out_activation)
db["ASR_AE_H1"]=ASR_AE_H1
ASR_AE_H2,hist["ASR_AE_H2"],wa2=learn_ae(w1_size,ASR_AE_H1,do_do,input_activation,out_activation)
db["ASR_AE_H2"]=ASR_AE_H2
ASR_AE_H3,hist["ASR_AE_H3"],wa3=learn_ae(hidden_size,ASR_AE_H2,do_do,input_activation,out_activation)
db["ASR_AE_H3"]=ASR_AE_H3
ASR_AE_H4,hist["ASR_AE_H4"],wa4=learn_ae(ASR["TRAIN"].shape[1],ASR_AE_H3,do_do,input_activation,out_activation)
db["ASR_AE_OUT"]=ASR_AE_H4
db.sync()
print "fine_tuning"
w,h=fintuned([(hidden_size,wa1),(w1_size,wa2),(hidden_size,wa3),(ASR["TRAIN"].shape[1],wa4)],ASR,ASR,input_activation,do_do)
hist["finetuning_a2a"] = (h.epoch,h.history)
for k,proj in enumerate(get_projection(w,ASR)):
db["ASR_AE_FTA2A_H"+str(k)]=proj
w,h=fintuned([(hidden_size,wa1),(w1_size,wa2),(hidden_size,wa3),(ASR["TRAIN"].shape[1],wa4)],ASR,TRS,input_activation,do_do)
hist["finetuning_a2t"] = (h.epoch,h.history)
for k,proj in enumerate(get_projection(w,ASR)):
db["ASR_AE_FTA2T_H"+str(k)]=proj
print "auto encoder trs learning"
TRS_AE_H1,hist["TRS_AE_H1"],wt1=learn_ae(hidden_size,TRS,do_do,input_activation,out_activation)
db["TRS_AE_H1"]=TRS_AE_H1
TRS_AE_H2,hist["TRS_AE_H2"],wt2=learn_ae(w1_size,TRS_AE_H1,do_do,input_activation,out_activation)
db["TRS_AE_H2"]=TRS_AE_H2
TRS_AE_H3,hist["TRS_AE_H3"],wt3,=learn_ae(hidden_size,TRS_AE_H2,do_do,input_activation,out_activation)
db["TRS_AE_H3"]=TRS_AE_H3
TRS_AE_H4,hist["TRS_AE_H4"],wt4=learn_ae(TRS["TRAIN"].shape[1],TRS_AE_H3,do_do,input_activation,out_activation)
db["TRS_AE_OUT"]=TRS_AE_H4
db.sync()
w,h=fintuned([(hidden_size,wt1),(w1_size,wt2),(hidden_size,wt3),(ASR["TRAIN"].shape[1],wt4)],TRS,TRS,input_activation,do_do)
hist["finetuning_t2t"] = (h.epoch,h.history)
for k,proj in enumerate(get_projection(w,ASR)):
db["ASR_AE_FTT2T_H"+str(k)]=proj
# In[261]:
#pred_dev= model_TRS_AE.predict(TRS_AE["DEV"],batch_size=1)
db.sync()
db.close()
json.dump({ "h1" : hidden_size,
"inside_activation" : input_activation,
"out_activation" : out_activation,
"do_dropout": do_do,
"loss" : loss,
"epochs" : epochs ,
"batch_size" : batch,
"patience" : patience,
"sgd" : sgd_repr,
"hist" : hist},
open("{}.json".format(sys.argv[2]),"w"),
indent=4)
|