def train_model_pe( use_char = True):
print('define model')
print('load data')
import data_utils, training_utils
tn_conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load('data.dict')
y = to_categorical(data_dict['y'])
if use_char:
vocab_dict = data_utils.pickle_load(tn_conf.char_dict)
embed_matrix = data_utils.load_embedding(vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
x = data_dict['x']
MAX_LEN = 600
name = 'conditionmodel_PE.h5'
else:
x = data_utils.pickle_load(tn_conf.term_file)
vocab_dict = data_utils.pickle_load(tn_conf.term_dict)
embed_matrix = data_utils.load_embedding(vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
MAX_LEN = 300
name = 'conditionmodel_term_PE.h5'
xe = [[i for i in range(MAX_LEN)] for _ in range(y.shape[0])]
xe = np.array(xe)
x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xe], y, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
model = ConditionModel(embed_matrix=embed_matrix, MAX_LEN=MAX_LEN, name=name, PE=True)
x_tn, y_tn = model.gen_train(x_tn, y_tn)
x_val, y_val = model.gen_train(x_val, y_val)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model_pe():
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.char_dict)
embed_matrix = data_utils.load_embedding(
vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
print('load embed done.')
y = to_categorical(data_dict['y'])
xe = [[i for i in range(600)] for _ in range(y.shape[0])]
xe = np.array(xe)
x_tn, y_tn, x_ts, y_ts = training_utils.split([data_dict['x'], xe],
y,
shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
model = SSCharModel(embed_matrix=embed_matrix,
name='sscharmodel_PE.h5',
PE=True)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model = SSCharModel(embed_matrix=embed_matrix,
name='sscharmodel_PE.h5',
PE=True,
train_embed=True)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model(use_char = False ):
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
if use_char:
vocab_dict = data_utils.pickle_load(conf.char_dict)
embed_matrix = data_utils.load_embedding(vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
MAX_LEN = 600
name = 'deepcnn_model.h5'
x = data_dict['x']
else:
x = data_utils.pickle_load(conf.term_file)
vocab_dict = data_utils.pickle_load(conf.term_dict)
embed_matrix = data_utils.load_embedding(vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
MAX_LEN = 300
name = 'deepcnn_model_term.h5'
print('load embed done.')
y = to_categorical(data_dict['y'])
x_tn, y_tn, x_ts, y_ts = training_utils.split(x, y, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
model = DeepCNNModel(embed_matrix=embed_matrix, MAX_LEN=MAX_LEN, name=name)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model(use_char=True):
print('train condition conv model.\nload data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.char_dict)
char_embed_matrix = data_utils.load_embedding(
vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
MAX_LEN = 600
x = data_dict['x']
xterm = data_utils.pickle_load(conf.term_file)
term_vocab_dict = data_utils.pickle_load(conf.term_dict)
term_embed_matrix = data_utils.load_embedding(
term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
MAX_LEN_TERM = 300
name = 'conditionconvmodel.h5'
print('load embed done.')
y = to_categorical(data_dict['y'])
xt = data_utils.pickle_load('data/lda_vec.pkl')
xfeat = data_utils.pickle_load(conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, conf.feat_norm)
xfeat = scaler.transform(xfeat)
x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xterm, xfeat, xt],
y,
shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
model = ConditionConvModel(char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
MAX_LEN=MAX_LEN,
NUM_FEAT=8,
name=name)
x_tn, y_tn = model.gen_train(x_tn, y_tn)
x_val, y_val = model.gen_train(x_val, y_val)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model = ConditionConvModel(char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
MAX_LEN=MAX_LEN,
NUM_FEAT=8,
name=name,
train_embed=True,
train_top=False,
lr=0.001)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model_n200(model_conf, name, ModelClass):
print(name)
import data_utils200 as data_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.char_dict)
char_embed_matrix = data_utils.load_embedding(
vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
MAX_LEN = conf.MAX_LEN
x = data_dict['x']
xterm = data_utils.pickle_load(conf.term_file)
term_vocab_dict = data_utils.pickle_load(conf.term_dict)
term_embed_matrix = data_utils.load_embedding(
term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
MAX_LEN_TERM = conf.MAX_LEN
print('load embed done.')
y = to_categorical(data_dict['y'])
xt = data_utils.pickle_load('data/lda_vec.pkl')
xfeat = data_utils.pickle_load(conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, conf.feat_norm)
xfeat = scaler.transform(xfeat)
xe = [[i for i in range(MAX_LEN)] for _ in range(y.shape[0])]
xe = np.array(xe)
xe_term = [[i for i in range(MAX_LEN_TERM)] for _ in range(y.shape[0])]
xe_term = np.array(xe_term)
x_tn, y_tn, x_ts, y_ts = training_utils.split(
[x, xe, xterm, xe_term, xfeat, xt], y, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
model = ModelClass(model_conf,
char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
name=name)
x_tn, y_tn = model.gen_train(x_tn, y_tn)
x_val, y_val = model.gen_train(x_val, y_val)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model = ModelClass(model_conf,
char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
name=name,
train_embed=True,
train_top=False)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_main_pe():
print('load data')
import data_utils, training_utils
tn_conf = TrainConfigure()
data_dict = data_utils.pickle_load(tn_conf.char_file)
y = to_categorical(data_dict['y'])
x = data_dict['x']
xterm = data_utils.pickle_load(tn_conf.term_file)
xfeat = data_utils.pickle_load(tn_conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, tn_conf.feat_norm)
xfeat = scaler.transform(xfeat)
print('loading embed ...')
term_vocab_dict = data_utils.pickle_load(tn_conf.term_dict)
term_embed_matrix = data_utils.load_embedding(
term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
# term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
# 'data/sgns.target.word-word.dynwin5.thr10.neg5.dim300.iter5',
# dump_path='data/term_embed_ww.pkl')
char_vocab_dict = data_utils.pickle_load(tn_conf.char_dict)
char_embed_matrix = data_utils.load_embedding(
char_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
print('load embed done.')
xe_char = [[i for i in range(600)] for _ in range(y.shape[0])]
xe_char = np.array(xe_char)
xe_term = [[i for i in range(300)] for _ in range(y.shape[0])]
xe_term = np.array(xe_term)
x_tn, y_tn, x_ts, y_ts = training_utils.split(
[x, xterm, xfeat, xe_char, xe_term], y, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('define model')
model = HybridModel(char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
NUM_FEAT=8,
name='hybridmodel_PE.h5',
PE=True) # +37
print('feat shape', xfeat.shape)
import sys
print('train')
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model_pe(use_char=False):
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
if use_char:
vocab_dict = data_utils.pickle_load(conf.char_dict)
embed_matrix = data_utils.load_embedding(
vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
MAX_LEN = 600
name = 'gateddeepcnnmodel_PE.h5'
x = data_dict['x']
else:
x = data_utils.pickle_load(conf.term_file)
vocab_dict = data_utils.pickle_load(conf.term_dict)
embed_matrix = data_utils.load_embedding(
vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
MAX_LEN = 300
name = 'gateddeepcnnmodel_term_PE.h5'
print('load embed done.')
y = to_categorical(data_dict['y'])
xe = [[i for i in range(MAX_LEN)] for _ in range(y.shape[0])]
xe = np.array(xe)
x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xe], y, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
model = GatedDeepCNNModel(embed_matrix=embed_matrix,
MAX_LEN=MAX_LEN,
name=name,
PE=True)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model = GatedDeepCNNModel(embed_matrix=embed_matrix,
MAX_LEN=MAX_LEN,
name=name,
PE=True,
train_embed=True)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model():
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
y = to_categorical(data_dict['y'])
x = data_dict['x']
xterm = data_utils.pickle_load(conf.term_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.term_dict)
embed_matrix = data_utils.load_embedding(
vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
print('load embed done.')
x_tn, y_tn, x_ts, y_ts = training_utils.split(xterm, y, shuffle=True)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=True)
print('train')
print('define model')
model = TermModel(embed_matrix=embed_matrix)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_model_cv( cv_index, cv_num ):
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.char_dict)
char_embed_matrix = data_utils.load_embedding(vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
MAX_LEN = 600
x = data_dict['x']
xterm = data_utils.pickle_load(conf.term_file)
term_vocab_dict = data_utils.pickle_load(conf.term_dict)
term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
MAX_LEN_TERM = 300
name = 'hybriddensemodel_cv{}.h5'.format(cv_index)
print('load embed done.')
y = to_categorical(data_dict['y'])
xt = data_utils.pickle_load('data/lda_vec.pkl')
xfeat = data_utils.pickle_load(conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, conf.feat_norm)
xfeat = scaler.transform(xfeat)
xe = [[i for i in range(MAX_LEN)] for _ in range(y.shape[0])]
xe = np.array(xe)
xe_term = [[i for i in range(MAX_LEN_TERM)] for _ in range(y.shape[0])]
xe_term = np.array(xe_term)
x_tn, y_tn, x_ts, y_ts = training_utils.split_cv([x, xe, xterm, xe_term, xfeat, xt], y, cv_index=cv_index,cv_num=cv_num)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
model = HybridDenseModel(char_embed_matrix=char_embed_matrix, term_embed_matrix=term_embed_matrix,
MAX_LEN=MAX_LEN, NUM_FEAT=8, PE=True, name=name)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model = HybridDenseModel(char_embed_matrix=char_embed_matrix, term_embed_matrix=term_embed_matrix,
MAX_LEN=MAX_LEN, NUM_FEAT=8, PE=True,
name=name, train_embed=True, train_top=False, lr=0.001)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
def fit(self, X_train, y_train):
self.X_train = X_train
self.y_train = y_train
skf = StratifiedKFold(n_splits=self.n_fold, shuffle=True, random_state=0)
# i = 0
y_test_list = []
y_pred_list = []
# for train_idx, test_idx in skf.split(self.X_train, self.y_train):
# X_train_s, y_train_s = self.X_train[train_idx], self.y_train[train_idx]
# X_test_s, y_test_s = self.X_train[test_idx], self.y_train[test_idx]
for i in range(self.n_fold):
print('cross ', i+1, '/', self.n_fold)
X_train_s, y_train_s, X_test_s, y_test_s = training_utils.split_cv(X_train, y_train, cv_num=self.n_fold, cv_index=i)
# X_train_s, X_val_s, y_train_s, y_val_s = train_test_split(X_train_s, y_train_s, test_size=0.1,
# stratify=y_train_s)
X_train_s, y_train_s, X_val_s, y_val_s = training_utils.split(X_train_s, y_train_s, split_ratio=0.95)
y_pred_s = None
for models in self.base_models:
model = models[i]
print(model.name)
if self.is_condition:
model.train_exp(X_train_s, y_train_s, X_val_s, y_val_s, X_test_s, y_test_s)
else:
model.train(X_train_s, y_train_s, X_val_s, y_val_s, X_test_s, y_test_s)
#作为特征的时候使用one-hot表示
if y_pred_s is None:
y_pred_s = model.predict(X_test_s)
else:
# import ipdb
# ipdb.set_trace( )
y_pred_s = np.hstack( (y_pred_s, model.predict(X_test_s) ) )
i += 1
y_test_list.append(y_test_s)
y_pred_list.append(y_pred_s)
# 使用y_pred_list做特征，y_test_list做目标，再次训练模型
X_top = np.vstack(y_pred_list)
y_top = np.vstack(y_test_list)
y_top = training_utils.convert_y( y_top )
if self.is_condition:
X_top = np.squeeze(X_top, axis=-1)
print(X_top.shape, y_top.shape)
self.top_model = LogisticRegression()
self.top_model.fit(X_top, y_top)
print(X_top)
def fit_tmp(self, X_train, y_train):
self.X_train = X_train
self.y_train = y_train
y_test_list = []
y_pred_list = []
for i in range(self.n_fold):
print('cross ', i+1, '/', self.n_fold)
X_train_s, y_train_s, X_test_s, y_test_s = training_utils.split_cv(X_train, y_train, cv_num=self.n_fold, cv_index=i)
X_train_s, y_train_s, X_val_s, y_val_s = training_utils.split(X_train_s, y_train_s, split_ratio=0.9)
y_pred_s = None
for models in self.base_models:
model = models[i]
print(model.name)
# if self.is_condition:
# model.train_exp(X_train_s, y_train_s, X_val_s, y_val_s, X_test_s, y_test_s)
# else:
# model.train(X_train_s, y_train_s, X_val_s, y_val_s, X_test_s, y_test_s)
#作为特征的时候使用one-hot表示
if y_pred_s is None:
y_pred_s = model.predict(X_test_s)
else:
# import ipdb
# ipdb.set_trace( )
y_pred_s = np.hstack( (y_pred_s, model.predict(X_test_s) ) )
i += 1
y_test_list.append(y_test_s)
y_pred_list.append(y_pred_s)
# 使用y_pred_list做特征，y_test_list做目标，再次训练模型
X_top = np.vstack(y_pred_list)
if self.is_condition:
X_top = np.squeeze(X_top, axis=-1)
y_top = np.vstack(y_test_list)
y_top = training_utils.convert_y( y_top )
print(X_top.shape, y_top.shape)
self.top_model = LogisticRegression()
self.top_model.fit(X_top, y_top)
print(X_top)
joblib.dump(self.top_model, self.name)
if len(sys.argv) > 2 and sys.argv[2] == 'pe':
print('define char model with position embedding')
print('load data')
import data_utils, training_utils
data_dict = data_utils.pickle_load(tn_conf.char_file)
y = to_categorical(data_dict['y'])
char_vocab_dict = data_utils.pickle_load(tn_conf.char_dict)
char_embed_matrix = data_utils.load_embedding(
char_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
xe = [[i for i in range(600)] for _ in range(y.shape[0])]
xe = np.array(xe)
x_tn, y_tn, x_ts, y_ts = training_utils.split([data_dict['x'], xe],
y,
shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn,
y_tn,
shuffle=False)
print('train')
model = RCNNModel(MAX_LEN=600,
name='RCNNmodel_char_PE.h5',
embed_matrix=char_embed_matrix,
PE=True)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
else:
print('define char model')
print('load data')
import data_utils, training_utils
data_dict = data_utils.pickle_load(tn_conf.char_file)
def train_model_peoe( ):
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.char_dict)
char_embed_matrix = data_utils.load_embedding(vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
MAX_LEN = 600
x = data_dict['x']
xterm = data_utils.pickle_load(conf.term_file)
term_vocab_dict = data_utils.pickle_load(conf.term_dict)
term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
char_embeds = []
char_embed_matrix_oe = data_utils.load_our_embedding(vocab_dict)
char_embeds.append( char_embed_matrix_oe )
for windows in [3, 5, 8]:
sg = 1
# for sg in [0,1]:
embed_file = 'data/char_embed_{}_{}.model'.format(windows, sg)
char_embed_tmp = data_utils.load_our_embedding(vocab_dict, model_file = embed_file,
dump_path = 'data/our_char_embed_{}_{}.pkl'.format(windows, sg))
char_embeds.append(char_embed_tmp)
term_embeds = []
term_embed_matrix_oe = data_utils.load_our_embedding(term_vocab_dict, model_file='data/term_embed.model',
dump_path='data/our_term_embed.pkl')
term_embeds.append( term_embed_matrix_oe )
for windows in [3, 5, 8]:
sg = 1
# for sg in [0,1]:
embed_file = 'data/term_embed_{}_{}.model'.format(windows, sg)
term_embed_tmp = data_utils.load_our_embedding(term_vocab_dict, model_file=embed_file,
dump_path='data/our_term_embed_{}_{}.pkl'.format(windows,
sg))
term_embeds.append(term_embed_tmp)
MAX_LEN_TERM = 300
name = 'hybriddpcnnmodel_PEOE.h5'
print('load embed done.')
y = to_categorical(data_dict['y'])
xt = data_utils.pickle_load('data/lda_vec.pkl')
xfeat = data_utils.pickle_load(conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, conf.feat_norm)
xfeat = scaler.transform(xfeat)
xe = [[i for i in range(MAX_LEN)] for _ in range(y.shape[0])]
xe = np.array(xe)
xe_term = [[i for i in range(MAX_LEN_TERM)] for _ in range(y.shape[0])]
xe_term = np.array(xe_term)
x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xe, xterm, xe_term, xfeat, xt], y, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('train')
print('define model')
#加入更多embedding模型以后，学习率要降低才能正常学习
model = HybridDPCNNModel(char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix, MAX_LEN=MAX_LEN, NUM_FEAT=8, PE=True,
name=name, char_embeds=char_embeds, term_embeds = term_embeds, lr=0.0004)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model = HybridDPCNNModel(char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix, MAX_LEN=MAX_LEN, NUM_FEAT=8, PE=True,
name=name, train_embed=True, train_top=False, lr=0.001, char_embeds=char_embeds, term_embeds = term_embeds)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
def train_main():
print('load data')
import data_utils, training_utils
tn_conf = TrainConfigure()
data_dict = data_utils.pickle_load(tn_conf.char_file)
y = to_categorical(data_dict['y'])
x = data_dict['x']
xterm = data_utils.pickle_load(tn_conf.term_file)
xfeat = data_utils.pickle_load(tn_conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, tn_conf.feat_norm)
xfeat = scaler.transform(xfeat)
print('loading embed ...')
term_vocab_dict = data_utils.pickle_load(tn_conf.term_dict)
term_embed_matrix = data_utils.load_embedding(
term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
# term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
# 'data/sgns.target.word-word.dynwin5.thr10.neg5.dim300.iter5',
# dump_path='data/term_embed_ww.pkl')
char_vocab_dict = data_utils.pickle_load(tn_conf.char_dict)
char_embed_matrix = data_utils.load_embedding(
char_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
print('load embed done.')
x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xterm, xfeat],
y,
shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn, y_tn, shuffle=False)
print('define model')
model = HybridModel(char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
NUM_FEAT=8) # +37
print('feat shape', xfeat.shape)
import sys
if len(sys.argv) <= 1 or sys.argv[1] == 'train':
print('train')
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
if len(sys.argv) > 1 and sys.argv[1] == 'val':
val_conf = ValidConfigure()
data_dict = data_utils.pickle_load(val_conf.char_file)
y = to_categorical(data_dict['y'])
x = data_dict['x']
ids = data_dict['id']
xterm = data_utils.pickle_load(val_conf.term_file)
xfeat = data_utils.pickle_load(val_conf.feat_file)
xfeat = scaler.transform(xfeat)
model.load_weights()
model.test([x, xterm, xfeat], ids, val_conf.out_file)
if len(sys.argv) > 1 and sys.argv[1] == 'error':
start_index = y_tn.shape[0] + y_val.shape[0]
texts = data_utils.load_all_text(tn_conf)
model.load_weights()
model.error_analysis(x_ts, y_ts, texts, start_index)
def stacking_main_condition():
print('load data')
tn_conf = TrainConfigure()
data_dict = data_utils.pickle_load(tn_conf.char_file)
y = to_categorical(data_dict['y'])
x = data_dict['x']
xterm = data_utils.pickle_load(tn_conf.term_file)
xfeat = data_utils.pickle_load(tn_conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, tn_conf.feat_norm)
xfeat = scaler.transform(xfeat)
xe = [[i for i in range(600)] for _ in range(y.shape[0])]
xe = np.array(xe)
xe_term = [[i for i in range(300)] for _ in range(y.shape[0])]
xe_term = np.array(xe_term)
xtopic = data_utils.pickle_load('data/lda_vec.pkl')
print('loading embed ...')
term_vocab_dict = data_utils.pickle_load(tn_conf.term_dict)
term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/term_embed.pkl')
# term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
# 'data/sgns.target.word-word.dynwin5.thr10.neg5.dim300.iter5',
# dump_path='data/term_embed_ww.pkl')
char_vocab_dict = data_utils.pickle_load(tn_conf.char_dict)
char_embed_matrix = data_utils.load_embedding(char_vocab_dict,
'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
dump_path='data/char_embed.pkl')
print('load embed done.')
name = 'model/stack_condition_model.pkl'
model_dir = 'model/stack/'
n_fold = 3
name = 'model/stack_condition_model5.pkl'
model_dir = 'model/stack5/'
n_fold = 5
stk_model = stacking(n_fold, name=name, is_condition=True)
conf = conditionmodelbase.ModelConfigure()
conf.PE = True
stk_model.add_model(ConditionConvModel, {"conf":conf,"char_embed_matrix":char_embed_matrix,
"term_embed_matrix":term_embed_matrix,
"name":model_dir+'conditionconvmodel_PE.h5'})
stk_model.add_model(ConditionGatedConvModel, {"conf":conf,"char_embed_matrix": char_embed_matrix,
"term_embed_matrix": term_embed_matrix,
"name": model_dir+'conditiongatedconvmodel_PE.h5'})
stk_model.add_model(ConditionGatedDeepCNNModel, {"conf":conf,"char_embed_matrix": char_embed_matrix,
"term_embed_matrix": term_embed_matrix,
"name": model_dir+'conditiongateddeepcnnmodel_PE.h5'})
conf.lr = 0.0005
stk_model.add_model(ConditionDPCNNModel, {"conf": conf, "char_embed_matrix": char_embed_matrix,
"term_embed_matrix": term_embed_matrix,
"name": model_dir + 'conditiondpcnnmodel_PE.h5'})
#采样0.1用于测试
# x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xe, xterm, xe_term, xfeat, xtopic], y, split_ratio=0.005, shuffle=False)
# x_tn, y_tn, x_ts, y_ts = training_utils.split(x_tn, y_tn, shuffle=False)
x_tn, y_tn, x_ts, y_ts = training_utils.split([x, xe, xterm, xe_term, xfeat, xtopic], y, split_ratio=0.95)
stk_model.fit(x_tn, y_tn)
# joblib.dump(stk_model, 'model/stack_model_3.pkl')
y_pred = stk_model.predict(x_ts)
acc = accuracy_score(training_utils.convert_y(y_pred), training_utils.convert_y(y_ts) )
print(acc)
cnf_matrix = confusion_matrix(training_utils.convert_y(y_pred), training_utils.convert_y(y_ts) )
print(cnf_matrix)
stk_model.save( )
def train_model_ftoe(model_conf,
model_name='hybridconvmodel_FTOE.h5',
ModelClass=HybridModelBase,
char_embed_file=None,
term_embed_file=None):
print(model_name)
print('load data')
import data_utils, training_utils
conf = data_utils.TrainConfigure()
data_dict = data_utils.pickle_load(conf.char_file)
print('loading embed ...')
vocab_dict = data_utils.pickle_load(conf.char_dict)
# char_embed_matrix = data_utils.load_embedding(vocab_dict,
# 'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
# dump_path='data/char_embed.pkl')
MAX_LEN = 600
x = data_dict['x']
xterm = data_utils.pickle_load(conf.term_file)
term_vocab_dict = data_utils.pickle_load(conf.term_dict)
# term_embed_matrix = data_utils.load_embedding(term_vocab_dict,
# 'data/sgns.target.word-character.char1-2.dynwin5.thr10.neg5.dim300.iter5',
# dump_path='data/term_embed.pkl')
char_embed_matrix = data_utils.load_embedding(
vocab_dict,
char_embed_file,
dump_path='data/{}.pkl'.format(char_embed_file[5:]))
term_embed_matrix = data_utils.load_embedding(
term_vocab_dict,
term_embed_file,
dump_path='data/{}.pkl'.format(term_embed_file[5:]))
MAX_LEN_TERM = 300
print('load embed done.')
y = to_categorical(data_dict['y'])
xt = data_utils.pickle_load('data/lda_vec.pkl')
xfeat = data_utils.pickle_load(conf.feat_file)
# normalization
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit(xfeat)
data_utils.pickle_dump(scaler, conf.feat_norm)
xfeat = scaler.transform(xfeat)
xe = [[i for i in range(MAX_LEN)] for _ in range(y.shape[0])]
xe = np.array(xe)
xe_term = [[i for i in range(MAX_LEN_TERM)] for _ in range(y.shape[0])]
xe_term = np.array(xe_term)
x_tn, y_tn, x_ts, y_ts = training_utils.split(
[x, xe, xterm, xe_term, xfeat, xt], y, split_ratio=0.95, shuffle=False)
x_tn, y_tn, x_val, y_val = training_utils.split(x_tn,
y_tn,
split_ratio=0.95,
shuffle=False)
print('train')
print('define model')
model = ModelClass(model_conf,
char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
name=model_name,
train_embed=False,
train_top=True)
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
model_conf.lr *= 0.5
model = ModelClass(model_conf,
char_embed_matrix=char_embed_matrix,
term_embed_matrix=term_embed_matrix,
name=model_name,
train_embed=True,
train_top=False)
model.load_weights()
model.train(x_tn, y_tn, x_val, y_val, x_ts, y_ts)
del model
