Blame view
DECODA_binary_BOW_AEINIT_TANH_MODELS.py
10.2 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 |
# coding: utf-8
# In[2]:
# Import
import pandas
# Alignement
from alignment.sequence import Sequence
from alignment.vocabulary import Vocabulary
from alignment.sequencealigner import *
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
# In[4]:
db=shelve.open("{}.shelve".format(sys.argv[1]),writeback=True)
sparse_corp=shelve.open("DECODA_sparse.shelve")
# In[6]:
ASR_sparse=sparse_corp["ASR_SPARSE"]
TRS_sparse=sparse_corp["TRS_SPARSE"]
# In[11]:
hidden_size=3096
hidden_size2=2048
input_activation="tanh"
out_activation="tanh"
loss="mse"
epochs=500
batch=128
patience=40
print "gogo autoencoder ASR"
sgd = SGD(lr=0.0001)#, momentum=0.9, nesterov=True)
autoencode=Sequential()
autoencode.add(Dense(ASR_sparse["TRAIN"].shape[1],hidden_size,init='glorot_uniform',activation=input_activation))
autoencode.add(Dense(hidden_size,hidden_size2,init='glorot_uniform',activation=input_activation))
autoencode.add(Dense(hidden_size2,hidden_size,init="glorot_uniform",activation=out_activation))
autoencode.add(Dense(hidden_size,ASR_sparse["DEV"].todense().shape[1],init="glorot_uniform",activation=out_activation))
#autoencode.compile(optimizer=sgd,loss=loss)
autoencode.compile(optimizer=sgd,loss=loss)
# In[ ]:
autoencode.fit(ASR_sparse["TRAIN"].todense(),ASR_sparse["TRAIN"].todense(),nb_epoch=epochs,batch_size=batch,
callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)], validation_data=(ASR_sparse["DEV"].todense(),ASR_sparse["DEV"].todense()),verbose=1)
# In[ ]:
auto_decoder=Sequential()
auto_decoder.add(Dense(ASR_sparse["DEV"].todense().shape[1],hidden_size,init='uniform',activation=input_activation,weights=autoencode.get_weights()[:2]))
auto_decoder.add(Dense(hidden_size,hidden_size2,init='glorot_uniform',activation=input_activation,weights=autoencode.get_weights()[2:4]))
auto_decoder.add(Dense(hidden_size2,hidden_size,init='glorot_uniform',activation=input_activation,weights=autoencode.get_weights()[4:6]))
auto_decoder.compile(optimizer=sgd,loss=loss)
# In[77]:
#autoencode.predict(ASR_sparse["DEV"].todense())
# In[ ]:
print "auto encoder et auto decoder asr okay"
ASR_sparse_AE={}
for i in ASR_sparse.keys():
ASR_sparse_AE[i]=auto_decoder.predict(ASR_sparse[i].todense())
#TRS_sparse[i]=dico.transform(TRS[i][2])
db["ASR_AE_H2_RELU"]=ASR_sparse_AE
auto_decoder=Sequential()
auto_decoder.add(Dense(ASR_sparse["DEV"].todense().shape[1],hidden_size,init='uniform',activation=input_activation,weights=autoencode.get_weights()[:2]))
auto_decoder.add(Dense(hidden_size,hidden_size2,init='glorot_uniform',activation=input_activation,weights=autoencode.get_weights()[2:4]))
auto_decoder.compile(optimizer=sgd,loss=loss)
ASR_sparse_AE_H1={}
for i in ASR_sparse.keys():
ASR_sparse_AE_H1[i]=auto_decoder.predict(ASR_sparse[i].todense())
#TRS_sparse[i]=dico.transform(TRS[i][2])
db["ASR_AE_H1_RELU"]=ASR_sparse_AE_H1
print "auto encoder trs learning"
# In[68]:/
sgd_trs = SGD(lr=0.1,momentum=0.9)
autoencode_trs=Sequential()
autoencode_trs.add(Dense(TRS_sparse["DEV"].todense().shape[1],hidden_size,init='glorot_uniform',activation=input_activation))
autoencode_trs.add(Dense(hidden_size,hidden_size2,init='glorot_uniform',activation=input_activation))
autoencode_trs.add(Dense(hidden_size2,hidden_size,init="glorot_uniform",activation=out_activation))
autoencode_trs.add(Dense(hidden_size,TRS_sparse["DEV"].todense().shape[1],init="glorot_uniform",activation=out_activation))
#autoencode_trs.compile(optimizer=sgd_trs,loss=loss)
autoencode_trs.compile(optimizer=sgd,loss=loss)
# In[69]:
autoencode_trs.fit(TRS_sparse["TRAIN"].todense(),TRS_sparse["TRAIN"].todense(),nb_epoch=epochs,batch_size=batch,
callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)],
validation_data=(TRS_sparse["DEV"].todense(),TRS_sparse["DEV"].todense()),verbose=1)
# In[87]:
auto_decoder_trs=Sequential()
auto_decoder_trs.add(Dense(ASR_sparse["DEV"].todense().shape[1],hidden_size,activation=input_activation,weights=autoencode_trs.get_weights()[:2]))
auto_decoder_trs.add(Dense(hidden_size,hidden_size2,activation=input_activation,weights=autoencode_trs.get_weights()[2:4]))
auto_decoder_trs.add(Dense(hidden_size2,hidden_size,activation=input_activation,weights=autoencode_trs.get_weights()[4:6]))
auto_decoder_trs.compile(optimizer=sgd,loss=loss)
# In[88]:
print "auto encoder trs okay"
TRS_sparse_AE={}
for i in TRS_sparse.keys():
TRS_sparse_AE[i]=auto_decoder_trs.predict(TRS_sparse[i].todense())
#TRS_sparse[i]=dico.transform(TRS[i][2])
db["TRS_AE_H2_RELU"]=TRS_sparse_AE
auto_decoder_trs=Sequential()
auto_decoder_trs.add(Dense(ASR_sparse["DEV"].todense().shape[1],hidden_size,activation=input_activation,weights=autoencode_trs.get_weights()[:2]))
auto_decoder_trs.add(Dense(hidden_size,hidden_size2,activation=input_activation,weights=autoencode_trs.get_weights()[2:4]))
auto_decoder_trs.compile(optimizer=sgd,loss=loss)
# In[88]:
print "auto encoder trs okay"
TRS_sparse_AE_H1={}
for i in TRS_sparse.keys():
TRS_sparse_AE_H1[i]=auto_decoder_trs.predict(TRS_sparse[i].todense())
#TRS_sparse[i]=dico.transform(TRS[i][2])
db["TRS_AE_H1_RELU"]=TRS_sparse_AE
db.sync()
# In[261]:
#pred_dev= model_TRS_AE.predict(TRS_sparse_AE["DEV"],batch_size=1)
TRS_AE={}
ASR_AE={}
for i in TRS_sparse.keys():
TRS_AE[i]=autoencode_trs.predict(TRS_sparse[i].todense())
ASR_AE[i]=autoencode.predict(ASR_sparse[i].todense())
db["TRS_AE_OUT_RELU"]=TRS_AE
db["ASR_AE_OUT_RELU"]=ASR_AE
db.sync()
# # Transfert de couche
# ICI
spe_finetune=Sequential()
spe_finetune.add(Dense(TRS_sparse["DEV"].todense().shape[1],hidden_size,init='glorot_uniform',activation=input_activation,weights=autoencode.get_weights()[:2]))
spe_finetune.add(Dense(hidden_size,hidden_size2,init='glorot_uniform',activation=input_activation,weights=autoencode.get_weights()[2:4]))
spe_finetune.add(Dense(hidden_size2,hidden_size,init="glorot_uniform",activation=out_activation,weights=autoencode_trs.get_weights()[4:6]))
spe_finetune.add(Dense(hidden_size,TRS_sparse["DEV"].todense().shape[1],init="glorot_uniform",activation=out_activation,weights=autoencode_trs.get_weights()[6:8]))
#spe_finetune.compile(optimizer=sgd_trs,loss=loss)
spe_finetune.compile(optimizer=sgd,loss=loss)
auto_decoder_spe_finetune=Sequential()
auto_decoder_spe_finetune.add(Dense(ASR_sparse["DEV"].todense().shape[1],hidden_size,activation=input_activation,weights=autoencode.get_weights()[:2]))
auto_decoder_spe_finetune.add(Dense(hidden_size,hidden_size2,activation=input_activation,weights=autoencode.get_weights()[2:4]))
auto_decoder_spe_finetune.add(Dense(hidden_size2,hidden_size,init="glorot_uniform",activation=out_activation,weights=autoencode_trs.get_weights()[4:6]))
auto_decoder_spe_finetune.compile(optimizer=sgd,loss=loss)
spe_finetune_fixed=Sequential()
spe_finetune_fixed.add(Dense(hidden_size2,hidden_size,init="glorot_uniform",activation=out_activation,weights=autoencode_trs.get_weights()[4:6]))
spe_finetune_fixed.add(Dense(hidden_size,TRS_sparse["DEV"].todense().shape[1],init="glorot_uniform",activation=out_activation,weights=autoencode_trs.get_weights()[6:8]))
spe_finetune_fixed.compile(optimizer=sgd,loss=loss)
# In[88]:
print "auto encoder SPE RAW OKAY"
ASR_AE_SPE_FINETUNE_RAW_OUT={}
ASR_AE_SPE_FINETUNE_RAW_H2={}
ASR_AE_SPE_FINETUNE_RAW_FIXED_OUT={}
for i in TRS_sparse.keys():
ASR_AE_SPE_FINETUNE_RAW_OUT[i]=spe_finetune.predict(ASR_sparse[i].todense())
#TRS_sparse[i]=dico.transform(TRS[i][2])
ASR_AE_SPE_FINETUNE_RAW_H2[i]=auto_decoder_spe_finetune.predict(ASR_sparse[i].todense())
ASR_AE_SPE_FINETUNE_RAW_FIXED_OUT[i]=spe_finetune_fixed.predict(ASR_sparse_AE_H1[i])
db["ASR_AE_SPE_FINETUNE_RAW_OUT"]=ASR_AE_SPE_FINETUNE_RAW_OUT
db["ASR_AE_SPE_FINETUNE_RAW_H2"]=ASR_AE_SPE_FINETUNE_RAW_H2
db["ASR_AE_SPE_FINETUNE_RAW_FIXED_OUT"]=ASR_AE_SPE_FINETUNE_RAW_FIXED_OUT
db.sync()
spe_finetune.fit(ASR_sparse["TRAIN"].todense(),TRS_sparse["TRAIN"].todense(),nb_epoch=epochs,batch_size=batch,
callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)], validation_data=(ASR_sparse["DEV"].todense(),TRS_sparse["DEV"].todense()),verbose=1)
auto_decoder_spe_finetune=Sequential()
auto_decoder_spe_finetune.add(Dense(ASR_sparse["DEV"].todense().shape[1],hidden_size,activation=input_activation,weights=spe_finetune.get_weights()[:2]))
auto_decoder_spe_finetune.add(Dense(hidden_size,hidden_size2,activation=input_activation,weights=spe_finetune.get_weights()[2:4]))
auto_decoder_spe_finetune.add(Dense(hidden_size2,hidden_size,init="glorot_uniform",activation=out_activation,weights=spe_finetune.get_weights()[4:6]))
auto_decoder_spe_finetune.compile(optimizer=sgd,loss=loss)
ASR_sparse_AE_H1[i]
spe_finetune_fixed.fit(ASR_sparse_AE_H1["TRAIN"],TRS_sparse["TRAIN"].todense(),nb_epoch=epochs,batch_size=batch,callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss',
patience=patience, verbose=0)],
validation_data=(ASR_sparse_AE_H1["DEV"],TRS_sparse["DEV"].todense()),verbose=1)
ASR_AE_SPE_FINETUNE_TUNED_OUT={}
ASR_AE_SPE_FINETUNE_TUNED_H2={}
ASR_AE_SPE_FINETUNE_TUNED_FIXED_OUT={}
for i in TRS_sparse.keys():
ASR_AE_SPE_FINETUNE_TUNED_OUT[i]=spe_finetune.predict(ASR_sparse[i].todense())
#TRS_sparse[i]=dico.transform(TRS[i][2])
ASR_AE_SPE_FINETUNE_TUNED_H2[i]=auto_decoder_spe_finetune.predict(ASR_sparse[i].todense())
ASR_AE_SPE_FINETUNE_TUNED_FIXED_OUT[i]=spe_finetune_fixed.predict(ASR_sparse_AE_H1[i])
db["ASR_AE_SPE_FINETUNE_TUNED_OUT"]=ASR_AE_SPE_FINETUNE_RAW_OUT
db["ASR_AE_SPE_FINETUNE_TUNED_H2"]=ASR_AE_SPE_FINETUNE_RAW_H2
db["ASR_AE_SPE_FINETUNE_TUNED_FIXED_OUT"]=ASR_AE_SPE_FINETUNE_RAW_FIXED_OUT
db.sync()
db.close()
|