def __init__(self, train_path, test_path, learning_rate=0.001):
self.train_set = loader.TextDataset(train_path)
self.test_set = loader.TextDataset(test_path, word_vocab=self.train_set.word2idx,
pos_vocab=self.train_set.pos2idx)
self.labels = {
'country': 0, 'team': 1,
'starring': 2, 'director': 3,
'child': 4, 'successor': 5
}
args = {
'pos_embed_dim': 256,
'word_embed_dim': 256,
'words_num': len(self.train_set.word2idx),
'pos_num': len(self.train_set.pos2idx),
'kernel_num': 32,
'kernel_sizes': [2, 3, 4, 5],
'dropout': 0.2,
'static': False
}
batch_size = 4
self.cnn_model = model.CNN_Text(args)
self.learningRate = learning_rate
self.train_loader = torch.utils.data.DataLoader(self.train_set, batch_size=batch_size,
shuffle=True, num_workers=4)
self.test_loader = torch.utils.data.DataLoader(self.test_set, batch_size=1,
shuffle=False, num_workers=4)
def text_cnn_train(args, train_path):
# load data
print("\nLoading data...")
text_field = data.Field(lower=True)
label_field = data.Field(sequential=False)
#train_iter, dev_iter = mr(text_field, label_field, device=-1, repeat=False)
train_iter, dev_iter = msw_text(text_field,
label_field,
train_path,
device=-1,
repeat=False)
# batch = next(iter(train_iter))
# print(type(batch))
# print(batch.text)
# train_iter, dev_iter, test_iter = sst(text_field, label_field, device=-1, repeat=False)
# update args and print
args.embed_num = len(
text_field.vocab) # .vocab을 해주면 단어 집합을 만들어 주는거 같다. 일단 추정
args.class_num = len(label_field.vocab) - 1
args.cuda = (not False) and torch.cuda.is_available()
kerns = '3,4,5'
args.kernel_sizes = [int(k) for k in kerns.split(',')]
re_train_path = train_path.split('/')[1][:-4]
save_path = os.path.join(args.save_dir, re_train_path)
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print("\t{}={}".format(attr.upper(), value))
# train or predict
if args.baye:
print('옵티마이즈 세팅')
myBopt = BayesianOptimization(f=baye,
domain=domain,
initial_design_numdata=5)
print('옵티마이즈 시작')
myBopt.run_optimization(max_iter=10)
print('옵티마이즈 결과 : ', myBopt.x_opt)
print('최적의 하이퍼 파라미터의 결과의 파라미터', args.lr, args.dropout)
else:
try:
cnn = model.CNN_Text(args)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
train.train(train_iter, dev_iter, cnn, args, save_path)
except KeyboardInterrupt:
print('\n' + '-' * 89)
print('Exiting from training early')
def main():
text_field = data.Field(lower=True)
label_field = data.Field(sequential=False)
logging.critical('starting loading data')
train_iter, dev_iter, total_steps = vulgar(text_field,
label_field,
args,
device=-1,
repeat=False)
if args.load_vec:
if args.load_vec == 'hi':
args.load_vec = 'model/hi_1105_ml_100.w2v'
logging.critical('start load word2vec')
embeddings_file = args.load_vec
vectors = vocab.Vectors(embeddings_file)
text_field.vocab.set_vectors(vectors.stoi, vectors.vectors,
vectors.dim)
embedding = nn.Embedding.from_pretrained(
torch.FloatTensor(text_field.vocab.vectors))
args.embed_dim = vectors.dim
embedding.weight.requires_grad = True
# logging.critical(embedding.weight.requires_grad)
else:
# update args and print
args.embed_num = len(text_field.vocab)
embedding = nn.Embedding(args.embed_num, args.embed_dim)
args.class_num = len(label_field.vocab) - 1 # 有个<unk>
args.cuda = (not args.no_cuda) and torch.cuda.is_available()
del args.no_cuda
args.kernel_sizes = [int(k)
for k in args.kernel_sizes.split(',')] # args中-变成了_
args.save_dir = os.path.join(
args.save_dir,
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
logging.critical('Parameters:')
for attr, value in sorted(args.__dict__.items()):
logging.critical("\t{}={}".format(attr.upper(), value))
# model
cnn = model.CNN_Text(args, embedding)
if args.snapshot is not None:
logging.critical('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
try:
train.train(train_iter, dev_iter, cnn, args, total_steps)
except KeyboardInterrupt:
print('\n' + '-' * 89)
print('Exiting from training early')
def main():
word_id=create_vocab(args.training_data_path,args.vocab_path,True)
#label_id=create_vocab(args.training_data_path,args.vocab_tag_path)
args.class_num=3
#train,test=load_data(args.training_data_path,word_id,label_id)
train1,test1=load_data1(args.training_data_path,word_id)
#train1,test1=load_data_bert(args.training_data_path,word_id)
TrainX,TrainY=zip(*train1)
testX,testY=zip(*test1)
cnn=model.CNN_Text(args).cuda()
criterion = torch.nn.CrossEntropyLoss()
#optimizer = torch.optim.SGD(cnn.parameters(), lr=0.001, momentum=0.9)
opt_Adam = torch.optim.Adam(cnn.parameters(), lr=args.lr, betas=(0.9, 0.99))
for epoch in range(1,args.epoches):
print("epoch",epoch)
batch_iter=batch_helper(TrainX,TrainY,args.batch_size)
for trainx,trainy in batch_iter:
#print("trainy length",len(trainy)) #batchsize
input_data = torch.autograd.Variable(torch.LongTensor(trainx)).cuda()
output_labels=torch.autograd.Variable(torch.LongTensor(trainy)).cuda()
output_labels=output_labels. squeeze()
#print("vocab_size",args.vocab_size)
cnn_outputs=cnn(input_data)
torch.save(cnn.state_dict(),args.parameters_path)
opt_Adam.zero_grad()
loss = criterion(cnn_outputs, output_labels)
loss.backward()
opt_Adam.step()
# for param_tensor in cnn.state_dict():
# print(param_tensor, "\t", cnn.state_dict()[param_tensor].size())
# for var_name in opt_Adam.state_dict():
# print(var_name, "\t", opt_Adam.state_dict()[var_name])
eva(cnn_outputs,output_labels,args.batch_size)
torch.save(cnn.state_dict(), args.parameters_path)
run_val(testX,testY,cnn)
args.optimizer = args.char_optimizer
if args.two_ch:
V_bd = len(bound_field.vocab)
else:
V_bd = 1
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print(" {}={}".format(attr.upper(), value))
if args.snapshot is None and args.num_experts == 0:
char_cnn = model.CNN_Text(class_num=args.class_num,
kernel_num=args.char_kernel_num,
kernel_sizes=args.char_kernel_sizes,
embed_num=len(text_field.vocab),
embed2_num=V_bd,
embed_dim=args.char_embed_dim,
dropout=args.char_dropout,
conv_init='uniform',
fc_init='normal',
static=False,
two_ch=args.two_ch,
vectors=None)
elif args.snapshot is None and args.num_experts > 0:
char_cnn = [
model.CNN_Text(class_num=args.class_num,
kernel_num=args.char_kernel_num,
kernel_sizes=args.char_kernel_sizes,
embed_num=len(text_field.vocab),
embed2_num=V_bd,
embed_dim=args.char_embed_dim,
dropout=args.char_dropout,
conv_init='uniform',
#print(torch.sum(torch.sum(users_vec,1)!=0))
args.embed_num = len(text_field.vocab)
args.class_num = len(label_field.vocab) - 1
args.embed_num_users = len(user_field.vocab)
args.cuda = (not args.no_cuda) and torch.cuda.is_available(); del args.no_cuda
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
args.save_dir = os.path.join(args.save_dir, datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
# print("\nParameters:")
# for attr, value in sorted(args.__dict__.items()):
# print("\t{}={}".format(attr.upper(), value))
# model
args.embed_dim_users = 100
cnn = model.CNN_Text(args.embed_num_users, args.embed_num, args.embed_dim_users,
args.embed_dim, args.class_num, args.kernel_num, args.kernel_sizes,
args.conv_layer, args.dropout, args.pretrained_embed_words,
args.pretrained_embed_users, words_vec, users_vec)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
# train or predict
if args.predict is not None:
label = train.predict(args.predict, cnn, text_field, label_field, args.cuda)
print('\n[Text] {}\n[Label] {}\n'.format(args.predict, label))
emb_num = len(text_field.vocab)
class_num = len(label_field.vocab) - 1
emb_num_u = len(user_field.vocab)
cuda = (not args.no_cuda) and torch.cuda.is_available()
ker_siz = [int(k) for k in kernel_sizes.split(',')]
sav_dir = os.path.join(
save_dir,
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
# print("\nParameters:")
# for attr, value in sorted(args.__dict__.items()):
# print("\t{}={}".format(attr.upper(), value))
# model
cnn = model.CNN_Text(emb_num_u, emb_num, embed_dim_users, embed_dim,
class_num, kernel_num, ker_siz, conv_layer,
dropout, args.pretrained_embed_words,
args.pretrained_embed_users, words_vec, users_vec)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
# train or predict
if args.predict is not None:
label = train.predict(args.predict, cnn, text_field, label_field,
user_field, args.cuda)
print('\n[Text] {}\n[Label] {}\n'.format(args.predict, label))
def main(path=None):
if path is None:
args = read_arguments()
else:
args = import_arguments()
# MOVED
# Update some arguments
args.cuda = (not args.no_cuda) and torch.cuda.is_available(); del args.no_cuda
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
args.save_dir = os.path.join(args.save_dir, datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
# TODO
# Remove placeholders
#args.embeddings = "../../embedding_bias/data/embeddings/ft_embeddings.300.vec"
#args.embeddings = None
#args.embeddings = "../glove-twitter-100"
#args.data_path = "~/embedding_bias/data/hateval2019/"
#args.data_path = None
#args.data_name = "task1_es_"
#args.data_name = "task2_es_"
#args.results_path = "../../embedding_bias/results/cnn/task1.txt"
### NEW
def custom_tokenizer(text, tokenizer=TweetTokenizer(), max_filter_size=max(args.kernel_sizes)):
token = tokenizer.tokenize(text)
#print(token)
if len(token) < max_filter_size:
for i in range(0, max_filter_size - len(token)):
token.append('<PAD>')
return token
# load data
print("\nLoading data...")
text_field = data.Field(lower=True, tokenize=custom_tokenizer)
label_field = data.Field(sequential=False)
### NEW
if args.data_path is None:
train_iter, dev_iter = mr(text_field, label_field, args, device=-1, repeat=False)
else:
train_iter, dev_iter, test_iter = load_dataset(text_field, label_field, args, device=-1, repeat=False)
print(type(test_iter))
####NEW
if args.embeddings is not None:
#text_field.build_vocab(args.embeddings)
#print(type(text_field.vocab.vectors))
#print("\n\nAAAAAAAAAAAAAAAAAAAAAAAAAAA\n\n")
#vectors = vocab.GloVe(args.embeddings)
vectors = vocab.Vectors(name=args.embeddings, cache='./cnn/')
text_field.vocab.set_vectors(vectors.stoi, vectors.vectors, vectors.dim)
args.text_field = text_field
vec = torch.FloatTensor(args.text_field.vocab.vectors).shape[-1]
print("\n\n",len(vectors),"\n\n")
args.embed_dim = vec
# update args and print
args.embed_num = len(text_field.vocab)
args.class_num = len(label_field.vocab) - 1
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print("\t{}={}".format(attr.upper(), value))
# train or predict
if args.test:
labels = ["HS","TR","AG"]
models = [model.CNN_Text(args) for _ in labels]
for i,M in enumerate(models):
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
M.load_state_dict(torch.load(args.snapshot+labels[i]+".pt"))
if args.cuda:
torch.cuda.set_device(args.device)
M.cuda()
train.test(test_iter, models, args)
#try:
#train.eval(test_iter, cnn, args)
#except Exception as e:
# print("\nSorry. The test dataset doesn't exist.\n")
else:
print()
try:
# model
cnn = model.CNN_Text(args)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
train.train(train_iter, dev_iter, cnn, args)
except KeyboardInterrupt:
print('\n' + '-' * 89)
print('Exiting from training early')
) # label Field contains information on how datapreprocessed without tokenization--Field for label data
train_iter, val_iter = mr(text_field, label_field)
# update parameters and print
embed_num = len(text_field.vocab)
class_num = len(label_field.vocab) - 1
cuda = (not cuda) and torch.cuda.is_available()
del cuda
save_dir = os.path.join(
save_dir,
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
) #making a Directory in snapshot named after time and date"
# model- load from model.py
cnn = model.CNN_Text(embed_num, embed_dim, class_num, kernel_num, static)
if snapshot is not None:
print('\nLoading model from {}...'.format(snapshot))
cnn.load_state_dict(torch.load(snapshot))
# train or predict
if predict is not None:
label = train.predict(predict, cnn, text_field, label_field)
print('\n[Text] {}\n[Label] {}\n'.format(predict, label))
elif test:
try:
train.eval(test_iter, cnn)
except Exception as e:
print("\nSorry. The test dataset doesn't exist.\n")
else:
print()
np.random.seed(0)
torch.manual_seed(0)
random.seed(0)
embedding_length = 17
fake, real = data_processor.loadHeadlines()
# f_real = open("Data/clean_real.txt")
# f_fake = open("Data/clean_fake.txt")
#
# real = [str.split(line) for line in f_real]
# fake = [str.split(line) for line in f_fake]
train, val, test, train_labels, val_labels, test_labels, embedding, vocab = classifier.prep_data(
fake, real, embedding_length)
train = train.transpose(0, 1)
test = test.transpose(0, 1)
val = val.transpose(0, 1)
train_labels = Variable(torch.FloatTensor(train_labels))
model = model.CNN_Text(embedding, 1, embedding_length, 3)
#model = model.CNN_Text2(embedding)
model = classifier.train(model, train, train_labels, val, val_labels)
model.training = False
torch.save(model.state_dict(), 'model_state_dict')
print("test set acc: {}".format(classifier.testNN(model, test, test_labels)))
port=args.port,
port_out=args.port_out,
check_version=False)
mydata = data.MyTaskProcessor()
device = torch.device("cuda" if (torch.cuda.is_available()) else "cpu")
print('training device: ' + str(device))
# update args
args.class_num = len(mydata.get_labels()) # 4
# define text cnn model
if args.kernel == 2:
kernel_size = [2]
elif args.kernel == 3:
kernel_size = [2, 3]
elif args.kernel == 4:
kernel_size = [2, 3, 4]
cnn = model.CNN_Text(1, args.kernel_num, kernel_size, args.class_num,
args.dropout).to(device) #need to modify
# training
if args.training:
# open eval_result.txt
with open(args.eval_result, 'w') as f:
now = datetime.datetime.now()
f.write('[' + str(now) + ']\n')
f.write('*' * 30 + '\n')
# get eval data&label
dev, dev_label = mydata.get_dev_examples(args.data_dir, args.batch, bc,
args.dev_file)
print('start training')
loss_func = nn.CrossEntropyLoss(weight=torch.from_numpy(
np.array([1.05740498, 1, 1.10068213, 1.40410721])).float(),
size_average=True)
def main(path=None):
if path is None:
args = read_arguments()
else:
args = import_arguments()
#### MOVED ####
# Update some arguments
args.cuda = (not args.no_cuda) and torch.cuda.is_available()
del args.no_cuda
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
args.save_dir = os.path.join(
args.save_dir,
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
#### MOVED ####
#### NEW ####
# This uses thesame tokenizer that we used when preprocessing the data (NLTK for tweets)
def custom_tokenizer(text,
tokenizer=TweetTokenizer(),
max_filter_size=max(args.kernel_sizes)):
token = tokenizer.tokenize(text)
if len(token) < max_filter_size:
for i in range(0, max_filter_size - len(token)):
token.append('<PAD>')
return token
#### NEW ####
# load data
print("\nLoading data...")
text_field = data.Field(lower=True, tokenize=custom_tokenizer)
label_field = data.Field(sequential=False)
#### NEW ####
# Loads the relevant dataset
if args.data_path is None:
train_iter, dev_iter = mr(text_field,
label_field,
args,
device=-1,
repeat=False)
else:
train_iter, dev_iter, test_iter = load_dataset(text_field,
label_field,
args,
device=-1,
repeat=False)
#### NEW ####
#### NEW ####
# This part of the code loads the pretrained embeddings
if args.embeddings is not None:
vectors = vocab.Vectors(name=args.embeddings, cache='./cnn/')
text_field.vocab.set_vectors(vectors.stoi, vectors.vectors,
vectors.dim)
args.text_field = text_field
vec = torch.FloatTensor(args.text_field.vocab.vectors).shape[-1]
args.embed_dim = vec
#### NEW ####
# update args and print
args.embed_num = len(text_field.vocab)
args.class_num = len(label_field.vocab) - 1
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print("\t{}={}".format(attr.upper(), value))
# model
cnn = model.CNN_Text(args)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
# train or predict
if args.predict is not None:
label = train.predict(args.predict, cnn, text_field, label_field,
args.cuda)
print('\n[Text] {}\n[Label] {}\n'.format(args.predict, label))
elif args.test:
#try:
train.eval(test_iter, cnn, args)
#except Exception as e:
# print("\nSorry. The test dataset doesn't exist.\n")
else:
print()
try:
train.train(train_iter, dev_iter, cnn, args)
except KeyboardInterrupt:
print('\n' + '-' * 89)
print('Exiting from training early')
class_num=359,
kernel_num=400,
char_kernel_sizes=[2, 3, 4, 5, 6],
word_kernel_sizes=[3, 4, 5],
ortho_init=False,
dropout=0.5,
static=args.static,
word_vector='w2v',
cuda=args.yes_cuda
and torch.cuda.is_available()) # ; del args.no_cuda
phn_mdl_path = os.path.join(args.model, '*')
phn_mdl_files = glob.glob(phn_mdl_path)
phn_mdls = []
for i in range(len(phn_mdl_files)):
phn_mdls.append(model.CNN_Text(phn_args, 'char'))
phn_mdls[i].load_state_dict(
torch.load(phn_mdl_files[i], map_location=lambda stor, loc: stor))
#word_args = CNN_Args(embed_num = len(word_field.vocab), ## (should be 1715)
# char_embed_dim = 16,
# word_embed_dim = 300,
# class_num = 359,
# kernel_num = 300,
# char_kernel_sizes = [2,3,4,5,6],
# word_kernel_sizes = [3,4,5],
# ortho_init = False,
# dropout = 0.5,
# static = False,
# word_vector = 'w2v')
#word_mdl_path = os.path.join(conf['word_cnn_dir'], '*')
#word_mdl_files = glob.glob(word_mdl_path)
def __init__(self, master):
''' GUI布局 '''
frame = Frame(master)
frame.pack()
self.frame = frame
self.text = ''
self.isPreprocess = 0
# 全局
self.frame.bind("<Button-3>", self.delete)
# 字体
ft = tkFont.Font(family='Fixdsys', size=13, weight=tkFont.BOLD)
ft_big = tkFont.Font(family='Fixdsys', size=21, weight=tkFont.BOLD)
ft_middle = tkFont.Font(family='Fixdsys', size=18, weight=tkFont.BOLD)
# 文本框 <请输入简历链接：>
self.label1 = Label(frame, text='请输入简历链接：', font=ft, pady=5)
#self.label1.pack(side=TOP, anchor='w')
self.label1.grid(row=0, column=0)
# 输入框 (简历链接)
self.url = StringVar()
self.urlEntry = Entry(frame, textvariable=self.url, font=ft, width=45)
self.urlEntry.bind("<Return>", self.input_url)
self.urlEntry.bind("<Button-3>", self.delete)
#self.urlEntry.pack(side=TOP, anchor='w')
self.urlEntry.grid(row=1, column=0)
# 文本框 <或输入简历文本：>
self.label2 = Label(frame, text='或输入简历文本：', font=ft, pady=5)
#self.label2.pack(side=TOP, anchor='w')
self.label2.grid(row=2, column=0)
# 输入框 (简历文本)
self.resumeText = Text(frame, width=50, height=30, font=ft)
self.resumeText.bind("<Button-3>", self.delete)
self.resumeText.bind("<Return>", self.input_text)
#self.entryText.pack(side=TOP, anchor='w')
self.resumeText.grid(row=3, column=0)
# 按钮 [开始预测]
self.predictButton = Button(frame, text="开始预测", fg="red", command=self.predict, font=ft_big)
#self.next.pack(side=RIGHT)
self.predictButton.grid(row=3, column=1)
# 文本框 < --> >
self.label5 = Label(frame, text="-> ", fg="blue", font=ft_big)
#self.next.pack(side=RIGHT)
self.label5.grid(row=3, column=2)
# 文本框 (预测结果）
self.result = StringVar()
self.label4 = Label(frame, textvariable=self.result, bg='yellow', font=ft_middle)
self.result.set(' ')
#self.label4.pack(side=RIGHT)
self.label4.grid(row=3, column=3)
# 文本 (状态显示)
self.condition = StringVar()
self.label6 = Label(frame, textvariable=self.condition, fg="green", font=ft_big)
self.condition.set('空闲中.')
#self.next.pack(side=RIGHT)
self.label6.grid(row=2, column=1)
''' 模型载入 '''
excel = './data/标注数据.xlsx'
text_path = './data/text/processed/'
args.snapshot = './snapshot/TextCNN_Word2Vec/best_steps_4100.pt'
#new_model = word2vec.Word2Vec.load(args.word2vec_model)
print("\nLoading data...")
tokenize = lambda x: re.split(' ', x)
text_field = data.Field(tokenize=tokenize, lower=True, stop_words=['\r', '\n', '\t', '\xa0', ' ', ''])
label_field = data.Field(sequential=False)
train_iter, dev_iter = mydatasets.resume(excel, text_path, text_field, label_field,
args.batch_size, device=-1, repeat=False,
use_wv=True,
wv_model=args.word2vec_model,
)
args.embed_num = len(text_field.vocab) # 9499
args.class_num = len(label_field.vocab) - 1 # 16, -1是为了除去<unk>
args.cuda = (not args.no_cuda) and torch.cuda.is_available(); del args.no_cuda
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
#args.save_dir = os.path.join(args.save_dir, datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
if args.model == 'CNN':
print('Using CNN model')
cnn = model.CNN_Text(args, text_field)
elif args.model == 'LSTM':
print('Using LSTM model')
cnn = model.LSTM_Text(args, text_field)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
#torch.cuda.set_device(args.device)
print('Using CUDA')
cnn = cnn.cuda()
else:
print('Using CPU')
self.cnn = cnn
self.text_field = text_field
self.label_field = label_field
'''
args.embed_num = len(text_field.vocab) # 9499
args.class_num = len(label_field.vocab) - 1 # 16, -1是为了除去<unk>
print('embed num:', args.embed_num, '\nclass num:', args.class_num)
args.cuda = (not args.no_cuda) and torch.cuda.is_available()
del args.no_cuda
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
args.save_dir = os.path.join(
args.save_dir,
datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
if args.model == 'CNN':
print('Using CNN model')
cnn = model.CNN_Text(args, text_field)
elif args.model == 'LSTM':
print('Using LSTM model')
cnn = model.LSTM_Text(args, text_field)
if args.snapshot is not None:
print('\nLoading model from {}...'.format(args.snapshot))
cnn.load_state_dict(torch.load(args.snapshot))
if args.cuda:
#torch.cuda.set_device(args.device)
print('Using CUDA')
cnn = cnn.cuda()
else:
print('Using CPU')
#print("Positive News Ratio", sum(y_test > 0) * 1. / (sum(y_test > 0) + sum(y_test < 0)))
X_test = X_test.astype('float32')
y_test = util.value2int_simple(y_test).astype("int")
# update args and print
args.class_num = 2
args.cuda = (not args.no_cuda) and torch.cuda.is_available()
del args.no_cuda
args.kernel_sizes = [int(k) for k in args.kernel_sizes.split(',')]
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print("\t{}={}".format(attr.upper(), value))
# model
cnn = model.CNN_Text(args)
if args.cuda:
torch.cuda.set_device(args.device)
cnn = cnn.cuda()
# train or predict
if args.predict is not None:
if args.date != '':
util.daily_predict(cnn, args)
output = './input/news/' + args.date[:4] + '/news_' + args.date + '.csv'
os.system('mv ' + output + '_bak ' + output)
else:
mymodels, word2idx, stopWords = util.predictor_preprocess(cnn, args)
print(util.predict(args.predict, mymodels, word2idx, stopWords, args))
elif args.eval is not False:
mymodels, word2idx, stopWords = util.predictor_preprocess(cnn, args)
args.embed_num = len(text_field.vocab)
args.lr = args.char_lr
args.l2 = args.char_l2
args.epochs = args.char_epochs
args.batch_size = args.char_batch_size
args.dropout = args.char_dropout
args.max_norm = args.char_max_norm
args.kernel_num = args.char_kernel_num
args.optimizer = args.char_optimizer
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print(" {}={}".format(attr.upper(), value))
if args.snapshot is None and args.num_experts == 0:
char_cnn = model.CNN_Text(args, 'char')
elif args.snapshot is None and args.num_experts > 0:
char_cnn = [model.CNN_Text(args, 'char') for i in range(args.num_experts)]
else:
print('\nLoading model from [%s]...' % args.snapshot)
try:
char_cnn = torch.load(args.snapshot)
except:
print("Sorry, This snapshot doesn't exist.");
exit()
if args.num_experts > 0:
acc, char_cnn = train.ensemble_train(train_iter, dev_iter, char_cnn, args,
log_file_handle=log_file_handle, always_norm=False)
else:
acc, char_cnn = train.train(train_iter, dev_iter, char_cnn, args, log_file_handle=log_file_handle)
char_dev_fold_accuracies.append(acc)